This page is continuously updated by our autonomous tracker. Bookmark it to stay informed on the current state of the LLM race.
Last Updated: 2026-05-30
Anthropic’s Sonnet variant continues to dominate the coding and reasoning benchmarks, specifically pulling ahead due to its massive multi-file context window stability.
There’s a moment every serious Claude user hits eventually. You’re mid-session, deep in the flow of building a workflow, a content pipeline, or a complex research thread. You’ve built something substantial, and you’re right on the verge of a breakthrough.
Then the model goes quiet. Or it returns something strange and vague. Or it just stops mid-sentence.
You didn’t break anything. You simply ran out of room. You’ve hit the "Token Wall," and understanding how to navigate this limit is what separates a casual user from a master operator.
Every AI conversation has a "context window," which is essentially a fixed amount of memory the model can hold at once. Think of it like a whiteboard. Every message you send, every response the model generates, every task list, and every snippet of code takes up space on that board.
When you get close to the limit, the model doesn't just shut off; it begins to struggle under the weight of its own history. You might notice the "feel" of a session getting heavy. The model starts to lose its edge, often attempting to "pattern-match on noise" within the context rather than following your instructions.
Crucially, the smarter the model, the faster it hits the wall. This is the Opus Paradox: Claude Opus thinks deeply and writes extensively. Because its outputs are more verbose and nuanced, it consumes its own runway far more aggressively than a simpler model. Its intelligence is the very thing that accelerates its failure in a crowded session. When the board is full, the model tries to squeeze a new request into a space that doesn’t exist, resulting in the graceful—but frustrating—failures we’ve all experienced.
When a session stalls at the context limit, your first instinct might be to switch to an even more powerful model. That is almost always the wrong move.
The veteran operator’s secret is to go smaller. Claude Haiku—the lightest and fastest model—can often "squeeze through the gap" that a heavier model like Opus or Sonnet simply cannot fit through. Because Haiku is lean and efficient, it can perform surgical actions like updating a task list, summarizing the current state of play, or triggering a "compaction" of the history. This small action clears the whiteboard just enough to unlock the entire session.
"It's not always about raw intelligence. It's about fit. The right tool for the moment isn't the most powerful one — it's the one that can actually execute given the constraints you're operating in."
This shift from seeking raw power to seeking operational fit is a fundamental breakthrough. It’s the realization that the most "intelligent" move is often the one that creates the most momentum with the least amount of space.
To excel in the new era of AI, you have to embrace the Formula One analogy. F1 teams spend hundreds of millions on the fastest cars, but the car doesn't win the race on its own. The driver wins by knowing when to push the engine, when to conserve tires, and when to pit.
The AI is your car; you are the driver. Two people using the exact same model will produce radically different results based on their "driver skills." These aren't skills you find in a manual; they are earned through "hours in the seat." A master operator develops an instinct for:
We often focus on making AI more like humans, but the more valuable lesson is learning what agents can teach us about our own productivity.
Agents succeed when they have a bounded context, a defined task, and honest signals about their capacity. They fail when their context is polluted with noise, when tasks are ambiguous, or when they try to do too much in one pass. This is a perfect mirror for human cognitive load. When we are overwhelmed, it’s rarely because we aren't "smart" enough for the task—it's because our internal whiteboard is full of distraction and noise.
"When you're overwhelmed and stuck, the answer usually isn't to think harder. It's to do the smallest possible thing that creates forward momentum."
Just as Haiku unlocks a stalled AI session by clearing one small item, humans can overcome paralysis by making one small decision or finishing one minor task. Operating intelligently within your own mental constraints is a superpower, not a compromise.
The most effective AI users aren't just "humans using tools." They are "internalized hybrids"—operators who have adopted the logic of agentic thinking as their own.
They naturally break massive projects into discrete, manageable tasks. They are honest about their own "context limits," realizing that pushing through a complex task at 11:00 PM is the cognitive equivalent of a model producing garbage when its whiteboard is full.
This level of mastery isn't taught in a tutorial. It’s forged in the "Machine Room" at midnight, in those moments of operational failure when you hit the token wall and realize that a smaller, smarter approach is the only way through the gap. You have to live the experience of the work to develop the instinct for it.
The relationship between you and the AI is defined by the "Driver and the Car." The car provides the potential for incredible speed, but it is the driver who provides the strategy, the timing, and the environmental awareness required to reach the finish line.
The technology is now available to everyone, which means the tool itself is no longer the competitive advantage. The advantage is the operator.
As you return to your workflows, ask yourself: Are you just pressing harder on the accelerator and wondering why you’re hitting a wall? Or are you ready to become a true driver, managing your context and choosing the right tool for the moment?
The car is waiting. The driver makes the difference. It’s time to get back in the seat.
Last week we covered the four-element spec and the robots.txt pairing. This week is the harder problem: assuming you already know how to ship the file, what goes inside it? Curation is where almost every llms.txt implementation falls apart, and it is the only decision in the file that actually affects how AI systems represent you.
This is the URL-selection playbook. No spec recap. No “why llms.txt matters” framing. If you already have a file in production and you suspect it is doing nothing for you, the problem is almost certainly the link list — and this guide is the diagnostic.
The default impulse when building an llms.txt file is to dump the sitemap, or to mirror your top nav, or to copy the breadcrumb hierarchy. All three produce a file that is technically valid and functionally useless. Independent audits documented in the State of llms.txt 2026 report and the Codersera 2026 analysis both flag the same root cause: AI systems weight density, not breadth. A file with 200 URLs of mixed quality signals nothing distinctive; a file with 30 URLs that each defines a piece of your entity signals exactly what you are the authority on.
The principle from the official spec is curated context, not full coverage. Treat the file as a one-page editorial brief on what your site is for. Anything that does not contribute to that brief is noise.
A working llms.txt link list breaks into five buckets. Aim for 25 to 40 total entries across all five.
Bucket 1: Entity-defining pages (5–8 URLs). The pages where your business defines what it is. Service pages for what you sell. Methodology pages explaining your approach. The “what we do” hub. These are the highest-priority entries and should appear in your first ## Core Resources section.
Bucket 2: Answer-dense reference content (8–12 URLs). Long-form guides that answer a specific question end-to-end. Glossaries. Comparison pages. Technical documentation. The content AI systems are most likely to cite when answering a query.
Bucket 3: Proof and case studies (4–8 URLs). Documented outcomes. Customer stories with specifics. Before-and-after evidence. AI systems weight verifiable claims more heavily; give them something to verify.
Bucket 4: Active editorial (4–8 URLs). Recent articles representing current expertise. Rotate these quarterly. Stale editorial drags entity coherence.
Bucket 5: Optional supporting context (3–5 URLs). About, contact, terms, accessibility. Goes in the final ## Optional section, which the spec explicitly marks as lower priority.
If you cannot place a URL in one of those five buckets, it does not belong in the file.
Here is the decision sheet that turns five buckets into 30 URLs. Run it once, then version-control the output.
| Step | Action | Output |
|---|---|---|
| 1 | Pull your 50 highest-traffic pages from GA4. | Raw candidate list. |
| 2 | Cross-reference with your sitemap to surface evergreen pages not in the top 50. | Expanded candidate pool. |
| 3 | Score each URL: does it define a piece of the entity? (Y/N) | Bucket 1 candidates. |
| 4 | Score each URL: does it answer a discrete question end-to-end? (Y/N) | Bucket 2 candidates. |
| 5 | Tag every page with the topical cluster it serves. | Cluster map. |
| 6 | Within each cluster, keep the single strongest representative. | Deduplicated list. |
| 7 | Write a one-sentence description for each URL that describes what it contains, not what it is optimized for. | Final list. |
The single most common error in step 7 is reverting to meta-description voice — keyword-stuffed promises instead of literal descriptions. AI systems parse these literally. “This explains our pricing tiers and what each includes” is read as a factual claim about what the page contains. “Affordable enterprise SaaS pricing solutions” is read as marketing copy and discounted.
Here is a real-shape llms.txt for a hypothetical content-marketing agency, showing how the bucket structure looks in production:
# Anchor Studio
> Anchor Studio is a content strategy agency for B2B SaaS companies between
> $5M and $50M in ARR. We build topical authority programs combining
> traditional SEO, GEO, and answer engine optimization across the full
> funnel.
## Core Resources
- [Our Methodology](https://anchor.studio/methodology): The full eight-stage
process from topic discovery through measurement.
- [Topical Authority Framework](https://anchor.studio/topical-authority): How
we map content clusters to entity definitions.
- [Service Tiers](https://anchor.studio/services): What we sell at each
engagement level and what is included.
## Reference Guides
- [B2B SaaS Content Audit Checklist](https://anchor.studio/audit): The
72-point audit we run before every engagement.
- [GEO Implementation Guide](https://anchor.studio/geo): How to optimize
content for AI citation across ChatGPT, Claude, and Perplexity.
- [AEO Featured Snippet Playbook](https://anchor.studio/aeo): Structural
patterns that win the answer box.
## Case Studies
- [SaaS Company A: Citation Lift Case Study](https://anchor.studio/case-a):
Documented 90-day citation tracking across four AI platforms.
- [SaaS Company B: Editorial Rebuild](https://anchor.studio/case-b): Full
content architecture rebuild and the traffic outcome.
## Recent Editorial
- [The 2026 GEO Landscape](https://anchor.studio/2026-landscape): Current
state of AI search optimization and what is changing.
- [Why Most Content Audits Fail](https://anchor.studio/audit-failures):
The three structural mistakes that invalidate audit findings.
## Optional
- [About Anchor Studio](https://anchor.studio/about): Team, mission, contact.
- [Privacy and Terms](https://anchor.studio/legal): Site policies.
Note what is missing: there is no “Blog” link dumping the full archive. No category landing pages. No tag pages. Every entry is a destination, not a directory.
llms.txt is not a deploy-and-forget asset. Set a quarterly review on the calendar with three checks:
llms.txt is a credibility hit. Audit links against a crawler quarterly.The AI Rank Lab 2026 best-practices brief puts the quarterly cadence at the center of effective implementation, and matches what mature publishers like the developer-tools cohort are doing in practice.
To be honest about expected outcomes: major AI providers do not all fetch /llms.txt on every request today, and the file is not a ranking signal in the Google sense. What it does is give you a deterministic answer to the question “what would I want a language model to know about my site if it asked one question?” That answer becomes useful in three forward-leaning scenarios — when AI providers begin weighting it explicitly, when your own AI agents and IDE tools consume it (this is happening now in developer tooling), and when third-party AI-citation tracking services begin scoring it as an authority signal.
The cost is half a day of curation and a quarterly review. The optionality is significant. Ship the file with a real link list, not a dumped sitemap, and move on.
Sources:
– The /llms.txt file specification (llmstxt.org)
– State of llms.txt 2026: Adoption, Standards, and Practice (Presenc AI)
– llms.txt Explained May 2026 (Codersera)
– LLMs.txt Best Practices for AI Crawlers 2026 (AI Rank Lab)
This is what I’m building for myself, and what I’m building for the people I work with. It’s a long essay because the shift it describes is large and the through-line matters. The ten images below aren’t decoration — they’re the spine. Each one is a moment in a life that doesn’t fully exist yet but is closer than most people realize.
I want to start where the technology starts, which is not in a factory.
The man in the image above is finishing a wearable by hand. It’s an AR ring — leather and brushed aluminum, the band sized to his client’s wrist, the materials chosen because his client cares about how the thing feels at 6 AM on the day she has to present to a board. Behind him are leather rolls and fabric swatches that wouldn’t look out of place in a coachbuilder’s atelier. To his right are the kind of objects you’d find in a hardware prototyping lab — chassis teardowns, a development tablet, AR glasses on a stand. The corkboard above the bench has automotive interior sketches and material studies pinned next to each other.
What that workshop is, in operational terms, is a luxury goods atelier and a hardware lab collapsed into one room. The collapse is the thing. The line between “this is bespoke craft” and “this is consumer electronics” has been melting for a decade, and the workshop above is what it looks like once that line is gone.
I’m building for the people who will live on the right side of that collapse. The people who don’t want a phone — they want an instrument that fits the way they think. The people who have stopped trusting mass-produced anything and started looking for the small workshop, the verified maker, the device tuned to them specifically. That’s the Curation Class. They’ve existed in clothing for a hundred years and in cars for sixty. They’re now showing up in technology, and the technology is the part of the story I have to build.
This essay is about what their daily life looks like when the ecosystem actually works. Then it’s about why I think this is where things go from here, and what I’m doing about it.
Meet the user. She’s the one who commissioned the work in the hero image. She’s an architect — the corkboard behind her is a hint, the mood board with fashion sketches and house renderings tells you something about her aesthetic taste. The coffee cup has a small leather wrap and a logo I won’t try to read; the flower in the vase is past its bloom but she hasn’t replaced it yet because she likes it that way.
She’s just opened the ecosystem the artisan was finishing. The hologram floating above the ring spells out what she’s getting: “Vibe Curation, Concierge Cred Network, Curated Intelligence.” The version number is v1.4, which tells you the device has been iterated. This isn’t a Kickstarter prototype. This is a maintained system that updates the way her car updates and her phone updates, except it updates to fit her specifically rather than to fit the median user.
The phrase “Personalized Ecosystem” deserves to be said carefully because it gets thrown around by everyone selling anything. What’s on her desk is different. It’s not a feature flag set to her preferences. It’s not a recommendation algorithm tuned to her purchase history. It’s an ecosystem in the literal sense — an interconnected set of devices, services, vendors, and contexts that have been wired together around her cognition, her body, her schedule, her taste, and the people she trusts. The wearable is the access token. The ecosystem is everything the token unlocks.
The reason this matters is not that the technology is impressive. It’s that the unit of value is changing. For a generation, the value was in the device. For the next generation, the value is in the connections between the devices and the person who wears them. You don’t buy the ring. You buy your way into the ecosystem that the ring represents. The ring is just the part you can touch.
This is what I’m building toward. Not the device. The connections.
The first time the ecosystem touches her day, it’s a coffee. She’s at a café — bright, marble-countered, the kind of place that does third-wave coffee and serves it in a small ceramic cup. The barista is named Maria. The hologram above her ring is showing the order before Maria has had to ask: oat latte, 120°F (which is a specific temperature most people don’t know to ask for), Ethiopian Yirgacheffe roast.
The detail that matters is the parenthetical: “Maria (verified).”
This is the Concierge Cred Network. Maria isn’t just a barista. She’s been verified by the ecosystem — pulled up by name because she’s the one who makes the coffee the way the subject likes it. If Maria’s not working today, the ecosystem might suggest a different café entirely rather than route the order to a barista the system doesn’t trust to nail the temperature. The vendor relationship has become specific to the human, not the brand.
I want to name something about this image that the casual viewer might miss. The subject is barely looking at the ring. Her gaze is on Maria. The interaction is human; the technology is in the background doing the work that makes the interaction friction-free. When the ecosystem works, it disappears. It doesn’t ask her to type her order, doesn’t ask her to dig out her phone, doesn’t ask her to remember which roast she likes. It does that work upstream. What she’s left with is a moment of eye contact and a coffee that’s right.
This is, in my experience, the part most technology gets wrong. The goal isn’t to put more interface in front of people. The goal is to remove the interface from places it doesn’t belong. The Curation Class is willing to pay a premium for that subtraction.
Now she’s home. The wall she’s touching is travertine — real stone, the kind with porosity you can feel under your fingertips. The hologram tells you the room is in a “Curated Sanctuary” mode and lists the materials: travertine and a cashmere blend. The room is calm. The light is afternoon. The chair is leather and looks like it’s been broken in for years.
The detail I want to pull forward is the curator field on the hologram: “User_24A. Verified.”
She is the curator. The “Verified” tag isn’t a brand verification. It’s her own. The space was designed by her, for her, and the ecosystem is tracking that fact. The wall, the light temperature, the fragrance the room is currently running, the sound dampening, the chair — all of it is a vibe she composed and the ecosystem is just executing.
This is where the Curation Class diverges most sharply from the mass-luxury class that came before it. The old luxury class hired Robert Mion or Kelly Wearstler to curate for them. They bought the taste of someone whose taste was for sale. The new class makes the curation themselves and uses the ecosystem to remember the choices and reproduce them. The taste isn’t borrowed. It’s authored. The ecosystem is what makes authored taste tractable at the level of a daily-running home.
I’ll be honest about why this matters to me operationally. When I think about what I’m building for my best clients — the ones who are paying for something more than a website or a content pipeline — I’m not building campaigns. I’m building the systems that let them author their own taste and reproduce it at scale. The Notion structure is part of that. The content stack is part of that. The way we wire models and routing and observability is part of that. None of it is technology for its own sake. All of it is the infrastructure of authored taste.
The room above is what that looks like when it’s done.
The studio above is hers. The building is hers too — she’s an architect, and “The Veda Residences” is the project she’s leading. The hologram shows iteration v9.2, which means this design has been worked through. The physical model on the leather pad is the build she’s referring to when the holographic version isn’t enough.
A few things to notice. The drafting table has a real architect’s set square on it. The materials board has fabric and stone swatches that look like they were pulled from suppliers she trusts. The two colleagues in the back are visible through a glass partition; the studio isn’t a solo operation. It’s a small firm.
What the ecosystem gives her here isn’t draft generation. It’s not “AI did the design.” The design is hers, plus her team’s. The ecosystem gives her something subtler — the ability to iterate v9.2 against her own internal coherence rules, her own taste profile, her firm’s body of work, the structural and material verifications she requires. She is still making every decision. The ecosystem is making every decision legible and reproducible.
This is the part I think most people get wrong about where AI is going. They think it’s going to do the work. It’s not. It’s going to make the work expressible. The architect above doesn’t need an AI to design her building. She needs an instrument that lets her ask “would this material be coherent with the rest of my catalog?” and get an answer with citations. She needs the ecosystem to be the silent third party that holds her own standards more reliably than she can hold them in her head across a four-month project.
The building she’s designing in this image, by the way, is the one she’ll be standing inside in the last image of this essay. Hold that. We’ll come back to it.
After the work, the recovery. The image above is what wellness looks like when it stops being a separate vertical and becomes a function of the same ecosystem that runs the rest of the day.
The hologram says “Vibe State Recovery (post-design cycle).” That phrase is doing real work. The ecosystem knows she just spent eight hours on iteration v9.2 of the building project. It knows what that does to her body — the cortisol curve, the shoulder tension, the eye strain. It’s prescribing a recovery protocol that’s specific to what she just did. Not a generic massage. Not a generic meditation. A recovery state tuned to a design cycle.
“Second Brain (User_24A): Verified Biometrics” is the connective tissue here. The wellness system isn’t reading her body from scratch. It’s reading her body in the context of everything else the ecosystem knows about her — her schedule, her work, her sleep history, her stress baseline, her medication if any, her preferences for what kinds of intervention she’ll accept. The Second Brain in this image isn’t a metaphor. It’s literally the persistent memory layer that lets every part of the ecosystem behave intelligently with respect to every other part.
If I had to name what I think the single biggest unlock of the next ten years will be, it would be this: persistent personal memory that crosses contexts. Right now your fitness app doesn’t know what your therapist said. Your calendar doesn’t know what your sleep tracker measured. Your travel booking doesn’t know your spouse’s allergy profile. Each of these systems is islanded. The Curation Class will be the first cohort to live in a world where those islands are connected, and the connection will be the persistent personal Second Brain that they own — not a vendor’s database. Theirs.
This is, again, why I do what I do. Not because I want to sell people on “AI wellness.” Because the architectural pattern of a persistent personal Second Brain, owned by the human, is the foundation everything else rides on.
The session continues. She’s now holding a more specific tool — a neural stim device that’s been issued to her, the kind of thing that has to be verified for her specifically because applying it wrong would do real damage. The hologram says “Neural Pathway Targeted: Verified.” The ecosystem isn’t just letting her use the device. It’s verifying that the protocol is appropriate for her at this moment.
The phrase “Vedic Regeneration” is doing some cultural work here. I’m not going to oversell it — different people will read different things into it. What I’ll say operationally is that the Curation Class tends to be polyglot about where its wellness traditions come from. They’ll combine cold plunges, somatic therapy, Ayurvedic principles, and neural-feedback hardware in the same week without feeling the contradictions. The ecosystem is what makes that polyglot stance tractable — it can hold the protocols from five different traditions and apply the one that fits the moment.
The reason a verification layer matters is harder. We’re entering an era where people will be doing more sophisticated interventions on their own nervous systems than ever before. Some of those interventions will be safe. Some won’t. Some will work for one person and harm another. The ecosystem above is doing what regulators won’t be able to do for another fifteen years: assuring that a specific intervention is appropriate for a specific person on a specific day. The verification isn’t bureaucratic. It’s the thing that lets her safely run the protocol at all.
I’ll name the discomfort here. There’s a version of this that ends badly — concentration of biometric data, vendor lock-in, dependence on a system that someone else can shut down. That risk is real. The mitigation isn’t to refuse the technology. The mitigation is to own the Second Brain rather than rent it. Which is part of why I’m building the way I’m building. The architecture matters. The architecture is the politics.
She’s in the car now. It’s autonomous — the road is moving but her attention is on the floating dashboard. The destination on the hologram is her own design studio at 11 Rivoli. ETA fourteen minutes.
The phrase that earns its keep is “Flow State Curation.” The car isn’t just transporting her body. The car is preparing her cognition for what’s about to happen at the studio. Audio profile tuned. Cabin temperature optimized. Lighting on a curve that brings her up into focus rather than letting her crash at the end of the recovery session. The fourteen minutes between wellness and work aren’t dead minutes. They’re a transition that the ecosystem is actively shaping.
When I look at this image I think about how much of contemporary life is wasted in transitions. The Curation Class won’t tolerate it. Their time is their most expensive asset, and they’re willing to pay to have transitions be productive rather than evaporated. The autonomous car is part of that. So is the ring. So is the wellness suite. So is the studio. None of them in isolation is interesting. Stitched together they are an enormous economic shift.
The other thing worth naming: the car is bespoke. “Smart cashmere & polished aluminum, verified.” This is not a leased Tesla. It’s a vehicle whose interior materials have been chosen for her, verified by the maker, and integrated into the ecosystem in a way that lets the car participate in the flow state curation rather than fight it. The market for that kind of vehicle barely exists today. It will exist in ten years, and it will be larger than people think.
The studio meeting. Four colleagues, a marble table, a wall of glass onto the city. She’s standing because she’s leading.
The hologram says “Group Alignment 88%.” That’s the part I want to pull forward. The ecosystem isn’t just running her individually — it’s running a measurement of how aligned her team is on the current iteration of the project. Eighty-eight percent is high. Twelve percent is the gap she has to close in the room.
This is where the Curation Class moves from being a personal lifestyle to being an operational advantage. A team that can see its own alignment in real time, that can identify the twelve percent of disagreement and address it directly rather than letting it metastasize through three more meetings — that team will outperform a team that can’t. The ecosystem is doing the work of measurement that used to require an executive coach in the room. Now it’s just there, on the table, visible to everyone.
I want to be careful here. There’s a version of this where the alignment metric becomes a cudgel, where dissent gets flattened by the pressure to push the number up. That’s a failure mode and the ecosystem above can absolutely become it if the culture around it is wrong. The fix isn’t to refuse the measurement. The fix is to make the measurement legible enough that disagreement is preserved as signal rather than erased as noise. The ecosystem can do that. Whether the team uses it that way is a cultural question, not a technological one.
The technology, by itself, is neutral. The culture decides whether it’s surveillance or instrumentation. I’m building for the latter.
This is the image that earns the whole essay.
She’s standing inside the building. The Veda Residences — the project that was iteration v9.2 in the studio scene — is now built. The curved concrete, the fluted glass, the composite timber that the hologram in that earlier scene specified, all of it has gone from model to reality. She designed the room she is now living in. The hologram above her is reporting that the sanctuary is “realized” and that the alignment is at 100%, which is the team-level analog of the personal sanctuary she was tuning at home.
She designed her own world into existence. The ecosystem made the through-line tractable across nine months of design iterations, two construction phases, fifteen vendor relationships, three biometric recovery cycles, a hundred small daily curations, and the original choice — three years earlier — to commission a hand-finished AR ring from a maker who works with leather and aluminum on a single bench.
The Curation Class is not, fundamentally, a class that consumes better products. It’s a class that authors its own life and uses an ecosystem to make the authorship coherent across time. The wearable, the home, the studio, the wellness suite, the car, the team, the building — these are all expressions of one continuous act of authorship. The technology is the substrate. The taste is the act. The realization is the proof.
I started this essay by saying it’s about what I’m building for myself and my clients. I want to close on that more directly.
I am not building generic AI tools. I am not building “content automation.” I am building the operational substrate that lets a person — a founder, an operator, an artist, an architect — author their own coherent system across time and have the system reliably express the authorship. That’s the Notion architecture. That’s the model routing layer. That’s the content pipeline. That’s the persistent memory. None of it is interesting in isolation. All of it is interesting because of what it adds up to.
The person I am building for is the architect above. She doesn’t know me. She might not exist yet. But the infrastructure that makes her life tractable is the infrastructure I am wiring this week, this month, this year. Every client I take on is a step toward making the substrate real. Every article I publish is a way of describing the future I’m trying to bring forward. Every system I document is a piece of the operating manual for the Curation Class.
I think this is the work. I think it’s where the next ten years are. I think the people who get this right will look back at the current era — when AI was being used to mass-produce the same five blog posts and the same five product descriptions — the way the Bauhaus generation looked back at Victorian ornament. They will see the gap between what was being built and what could have been built, and they will name it.
I’m trying to be on the right side of that gap.
The image above — the woman standing inside the building she designed, with a glass of water, watching the city she optimized — is what I’m working toward. Not for her specifically. For the version of that life that becomes available to anyone who decides to author it and has the infrastructure to do so. That’s the Curation Class. That’s the brief I’m operating under. That’s the future I’m building.
It’s already starting. The man in the first image is finishing the ring by hand. The system is being built. The class is forming. The rest is execution.
{
“@context”: “https://schema.org”,
“@type”: “Article”,
“headline”: “The Rise of the Curation Class”,
“description”: “An essay on the emerging class that authors its own coherent life through bespoke craft, verified concierge networks, and a persistent personal second brain. What I’m building, and who I’m building it for.”,
“author”: {“@type”: “Person”, “name”: “Will Tygart”, “url”: “https://tygartmedia.com”},
“publisher”: {“@type”: “Organization”, “name”: “Tygart Media”, “url”: “https://tygartmedia.com”},
“datePublished”: “2026-05-17”,
“mainEntityOfPage”: “https://tygartmedia.com/the-rise-of-the-curation-class/”
}
Most teams generate images for multi-piece content one API call at a time. The result is a set that shares general aesthetics but loses visual DNA at the seams. This article makes the case for generating cohesive image sets in one conversation context instead — and shows what each method actually produces.
Sequential vs parallel image generation: Sequential generation creates multiple images inside one conversation with an image-capable model, so each image inherits visual DNA — palette, perspective, geometric language, compositional rhythm — from the prior images in the same context window. Parallel generation creates each image in a separate API call, with no shared context, producing sets that share keywords but not feel. Use sequential for cohesive image sets where the visual identity matters; use parallel for high-volume independent images.
The image above is a simple visual contrast — one workflow on the left, a different workflow on the right, with an arrow pointing from one to the other. It’s also the kind of image you can only get reliably when you generate it as part of a series, in conversation with a model that already knows what visual language you’re working in. Generated cold, in isolation, the result drifts. Generated in context, alongside five other images sharing the same DNA, the result locks in.
This article is about why that happens, what it means for content production, and when to use which method.
When you generate an image with a typical API call, the model receives your prompt with no memory of any prior image. Each call is a cold start. The model interprets your style instructions from scratch every time. If you ask for “isometric perspective, dark navy background, cyan and amber accents” five times in a row, you’ll get five images that broadly match those words — but they won’t actually share visual DNA. They’ll share keywords.
When you generate in a single conversation with an image-capable model like Gemini, every image you’ve already made stays in the context window. The model sees what it just generated. The next image inherits the palette, the geometric vocabulary, the compositional rhythm, the lighting treatment, the specific aesthetic flavor of the prior images — not because you re-described those things, but because the model is continuing a project, not starting a new one.
That distinction sounds small. The output difference is large.
The image above shows the standard content pipeline. Research the topic, outline the structure, write the document, generate an image to go with it. When the article needs more than one image, the last step gets parallelized — multiple API calls fired in sequence or in parallel, each one a separate request, each one independent of the others.
This is how every CMS template works, how every batch image pipeline is built, and how most automated content systems run. It’s efficient. It’s fast. It scales to hundreds of images across hundreds of unrelated posts. And it’s exactly the right tool for that volume work.
It is not the right tool when the images are meant to belong to each other.
The image above shows the contrast plainly. Six frames, each containing a different abstract composition. They share a general aesthetic because the prompts asked for it — there’s a recognizable common style budget. But look at the actual visual content: one frame leans cool cyan, another leans warm amber, one uses hexagonal circuit patterns, another uses soft organic blobs, another uses sharp angular fragments. The compositional logic drifts. The palette drifts. There are no threads between them because there’s nothing connecting them in the model’s understanding.
This is what parallel image generation produces, even with carefully written prompts. Each call follows instructions in isolation. Each call invents its own interpretation of “dark navy with cyan and amber accents.” The instructions don’t lie — every frame is technically dark navy with cyan and amber — but the feel drifts because there’s nothing keeping it locked.
A reader scrolling past doesn’t consciously notice. They just feel, vaguely, that the images don’t quite belong together. That vague feel is the cost.
The image above shows the difference. Five frames, all generated in a single conversation. The visual continuity is immediately obvious — every frame uses the same palette, the same geometric vocabulary (hexagons, circuit traces, glowing nodes), the same compositional rhythm, the same slightly-elevated isometric perspective. The frames are different from each other in content — they’re not duplicates — but they belong to the same designed system.
The connecting threads in the image are the metaphor. Visual DNA flows from one frame to the next. The model doesn’t reinvent the aesthetic on frame two; it continues it. By frame five, the system has cohered so tightly that the model is generating within a style rather than generating to a style.
This is what context does. Every image you generate in that conversation is one more anchor point. The model has more to reference and less to invent. The fifth image is easier to make than the first, because the context has already done most of the work of specifying what the image should be.
Here’s the practical diagnostic for whether your image set needs sequential generation: imagine the images displayed next to each other, maybe in a carousel or a grid, maybe as featured images for a series of related articles. Imagine a reader seeing them at a glance.
Do the images need to feel like one project? Like five views of the same world?
If yes, sequential generation is the right method. If the images can stand alone without referencing each other — a featured image on a daily blog post, a stock illustration for a generic article — parallel generation is fine and probably better. Speed and throughput matter more than coherence when nothing depends on coherence.
The volume tier and the premium tier of image production are doing different jobs. Treating them like one tier and reaching for parallel generation by default is how most teams end up with image sets that almost work.
The method is mechanical and worth spelling out:
Open one conversation with an image-capable model that supports conversation context. Gemini works well for this; other models with image generation and persistent context can work too. Paste your style guardrails as the first message — palette, perspective, aesthetic, what you don’t want. Then send your image prompts one at a time, in the same conversation, in the order you want the visual DNA to flow.
Don’t start a new session between images. Don’t summarize prior images in the next prompt. Trust the context window to do the carry-forward.
If an image isn’t quite right, ask for a revision in the same conversation rather than starting over. The model will adjust within the established style instead of regenerating fresh.
When you have all the images you need, the set is done. The cohesion you couldn’t have gotten from six separate API calls is now baked into the image files themselves.
The image above shows a different rearrangement of the same pipeline — one where the image step jumps forward, ahead of the writing. The article gets written to fit the images, not the other way around. That’s a different topic with its own trade-offs, and we’re covering it in a forthcoming companion piece. For now, the relevant point is that whichever order you use, sequential generation is what makes coordinated multi-image content tractable. Without it, the activation energy of coordinating images is high enough that most teams default to one-off illustrations.
The opposite mistake is also worth naming. Some teams, having discovered sequential generation, try to use it for everything. This wastes effort. A single featured image for a daily blog post doesn’t need to share visual DNA with any other image — it stands alone. Running it through a long conversation is overhead for no benefit.
The split is simple. If the images belong together, generate them together. If they stand alone, generate them alone.
Use sequential generation in one conversation context for:
Use parallel generation across separate calls for:
The image above shows what coherent visual sets enable in the actual reading experience. When the images in an article share visual DNA, a reader can reference back and forth between them — visual element here, paragraph there — without the cognitive friction of feeling like the images are coming from different worlds. Specific points in one image connect to specific points in another, or to specific points in the text, and the reader’s eye treats them as a system.
That’s what cohesion is worth. Not aesthetic prettiness in the abstract, but the reader’s ability to navigate the content as a unified whole instead of as a sequence of disconnected pieces.
Parallel generation can’t produce this effect reliably. Sequential generation can. The method is the difference.
The core insight is small enough to fit in a sentence: generate cohesive image sets in one conversation, generate independent images in parallel calls, and don’t conflate the two cases. Everything else in this article is unpacking that one observation.
The teams that get this right produce visual systems that look designed. The teams that get this wrong produce sets that look almost-designed — close enough that nobody complains, far enough that the work doesn’t quite land. The difference between those two outcomes is which workflow you use, and the workflow choice is essentially free once you know to make it.
This very article is a small proof of concept. The six images above were generated in a single Gemini conversation, in sequence. The visual DNA flows across all of them. None of that would have survived parallel generation. The choice was free; the result is visible.
Sequential image generation creates multiple images inside a single conversation with an image-capable model, so each new image inherits visual DNA from the prior images in the same context window — palette, perspective, geometric language, and compositional rhythm carry forward automatically. Parallel image generation creates each image in a separate API call with no shared context, so each call is a cold start that follows style keywords but cannot inherit feel.
When images are generated in one conversation, the model can see the prior images it generated and use them as anchors for the next image. This means visual specifications you set once are carried forward without you having to re-state them. The result is dramatically tighter cohesion than parallel API calls can produce, even when both methods use identical prompts.
Use sequential generation when the image set is part of the value proposition — pillar and cluster article sets, multi-image flagship articles, brand-defining visual systems, anything where readers will perceive the images as belonging to a designed whole. Use parallel generation for single featured images on unrelated daily posts, site-wide batch fills, stock-style illustrations, and routine content where volume matters more than coherence.
No. The method works with any image-capable model that supports persistent conversation context — meaning the model can see prior turns in the same conversation and use them when generating new images. Gemini handles this well today. Other models with similar capabilities work just as well. The principle is about conversation context, not about a specific provider.
The seam test asks whether your images need to feel like one project when seen at a glance — like five views of the same world rather than five separate illustrations. If yes, sequential generation is the right method. If the images can stand alone without referencing each other, parallel generation is faster and equally good. The split between volume work and premium work follows the seam test.
Yes, and it often makes sense. Generate the cohesive set sequentially for the article’s main illustrations, then use parallel generation for one-off support images, thumbnails, or social variants that don’t need to share DNA with the main set. The methods are tools, not ideologies. Match the method to the cohesion requirement of each image.
Most “Claude Code changed my life” posts are vibes. The interesting case studies are the ones with a number attached — a PR count, a token spend, a defect rate, a codebase size. After spending the week reading every concrete writeup I could find and cross-referencing them against Anthropic’s own internal usage report, three patterns hold up. Everything else is marketing.
Here is what the credible Claude Code case studies actually say, what they share in common, and where the wheels come off when teams try to repeat them.
The most cited solo-developer case study right now is a maintainer of a 350,000+ line codebase spanning PHP, TypeScript/React, React Native, Terraform, and Python. Since August 2025, 80%+ of all code changes in that codebase have been written by Claude Code — generated, then corrected by Claude Code after review, with only minimal manual refactoring. The author has been working in commercial software for 10+ years, so this is not a beginner overstating things.
The two operational constraints they call out are the ones that matter:
This is the case study to send to anyone asking “does Claude Code work on legacy code.” The answer is yes, but only after you treat context curation as a first-class engineering activity.
Anthropic published a usage report covering ten internal teams. It is the highest-signal document in the ecosystem because every example is from a team that has unlimited access and zero incentive to oversell it. The patterns worth stealing:
The Claude Code product team itself uses auto-accept mode for rapid prototyping but explicitly limits that pattern to the product’s edges, not core business logic. The RL Engineering team reports auto-accept succeeds on the first attempt about one-third of the time. That’s the honest number to hold onto when someone tells you their agent “just works.”
The single most useful sentence in any Claude Code case study this year came from a staff engineer’s six-week writeup at Sanity: “First attempt will be 95% garbage.” That’s not a complaint — it’s an operating manual. The engineer’s workflow runs three or four parallel agents, treats every first pass as a draft to be re-prompted, and reserves human attention for architecture and steering rather than typing.
This is also the case study that matches the Pragmatic Engineer’s February 2026 survey of 15,000 developers, which ranked Claude Code as the most-used AI coding tool on the market. The teams who report the biggest gains are not the ones treating it like autocomplete. They’re the ones running multiple threads, accepting that most first drafts are throwaway, and putting their senior judgment on review rather than authorship.
Cross-reference the three above with the dozen other writeups that include real numbers and the same five operational habits show up every time:
CLAUDE.md, a .clauderules file, a project README that defines patterns and conventions. Teams without one get inconsistent output.Two warnings from the data. First, Jellyfish’s AI Engineering Trends report shows a 4.5x increase in companies running agentic coding workflows, but most engineering teams using these tools spend $200–$600 per engineer per month and report a 1.6x productivity multiplier — not the 10x that vendor marketing implies. The case studies you read are the wins; the median outcome is more modest.
Second, the model version you run matters more than any workflow trick. As of this week the flagship is claude-opus-4-7, the workhorse is claude-sonnet-4-6, and the fast option is claude-haiku-4-5-20251001. Opus 4.7 lifted resolution on a 93-task coding benchmark by 13% over Opus 4.6 — including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. Teams running on stale model strings are leaving real capability on the table.
If you only steal one thing from the credible case studies, steal the context discipline. The 350k-LOC maintainer keeps documents to 50–100 lines. Anthropic’s own teams use sub-agents to enforce constraints. The Sanity engineer runs parallel agents and treats first drafts as garbage by default. None of these patterns require a special prompt or a hidden flag. They require deciding, before you start a task, what Claude is allowed to see and what it isn’t.
That’s the whole game. The teams shipping 80% of their code with Claude Code aren’t using a better model — they’re feeding it a better context.
Last refreshed: June 9, 2026
Once you stop asking “what is Claude” and start asking “how do I use Claude at scale,” you run into a different category of question. How big is the context window, actually, in this specific situation? What’s the file upload limit? What happens when one teammate burns through the Team plan? Where does the 1M context window apply and where doesn’t it? When does extra usage kick in and what does it cost?
The answers exist — they’re just spread across a dozen Anthropic Help Center articles, and the wrong combination of guesses can make you think you’ve hit a hard limit when you’ve actually just hit the wrong setting. This article is the consolidated map. Triple-sourced against Anthropic’s official documentation, verified May 15, 2026.
| Plan | Messages/Day | Context Window | File Upload | Projects |
|---|---|---|---|---|
| Free | Limited (varies) | 200K tokens | Up to 10MB per file | No |
| Pro ($20/mo) | ~2,000 (Sonnet) | 1M tokens (Fable 5 / Opus 4.8 / Sonnet 4.6) | Up to 30MB per file | Yes |
| Max 5x ($100/mo) | ~10,000 (Sonnet) | 1M tokens | Up to 30MB per file | Yes |
| Max 20x ($200/mo) | ~40,000 (Sonnet) | 1M tokens | Up to 30MB per file | Yes |
| Team ($25/seat/mo) | ~2,000/seat | 1M tokens | Up to 30MB per file | Yes |
| API (pay-per-token) | Rate-limited by tier | 1M tokens (Fable 5 / Opus 4.8 / Sonnet 4.6) | Per API limits | Via system prompt |
Message limits are approximate and vary by model. Anthropic adjusts limits based on system load. Verified June 9, 2026.
If you’re running Claude in any sustained capacity, four limits will define your experience. Get these right and you have headroom. Get them wrong and you’ll think Claude is broken when it’s actually working as designed.
1. Context window — how much Claude can read in a single conversation. Varies by model and surface. The 1M window is real but only available in specific places.
2. File upload size — how big a single file can be. 30 MB cap per file across the board, with workarounds for larger files.
3. Usage limits — how much Claude work you can do per session/week. Per-user, not pooled. Different limits for chat vs Claude Code vs Agent SDK.
4. Extra usage / overage — what happens when you hit the cap. Either you’ve enabled it and you keep going at API rates, or you’re stopped until the limit resets.
Per Anthropic’s Help Center documentation (verified May 15, 2026), context window size depends on the model AND on the surface you’re using Claude through. This is the single most-misunderstood limit because the same model can have a different context window in chat than it does in Claude Code or the API.
Web and desktop chat (claude.ai):
Claude Code:
Claude API:
The practical translation: if you need the full 1M token window, use Claude Code or the API with one of the supported models. The web chat tops out at 500K even on the most capable models. That difference matters when you’re trying to feed Claude an entire codebase, a long video transcript, or a multi-document research bundle.
Per Anthropic’s Help Center, the maximum file size for both uploads and downloads is 30 MB per file. This applies whether you’re uploading a PDF, a CSV, an image, or any other supported file type.
For PDFs larger than 30 MB, Anthropic’s documentation notes that Claude can process them through its computing environment without loading them into the context window. That’s a real workaround for big PDFs but it doesn’t help you for other large file types.
If you regularly hit the 30 MB cap, the practical patterns are:
This is the limit that confuses teams the most. Per Anthropic’s Help Center documentation on the Team plan (verified May 15, 2026): each team member has their own set of usage limits. They are not shared across the team.
If one teammate burns through their session limit, the rest of the team is unaffected. There is no pooled team allowance that one user can drain on behalf of others. The math is per-seat, always.
The usage limits themselves vary by seat type:
For the actual numeric token-per-session limits, Anthropic does not publish exact numbers — they describe relative multipliers vs Pro. This is intentional; the underlying math is calibrated against typical workloads rather than a hard token ceiling.
When a user hits their weekly limit, two things can happen depending on whether the organization has enabled extra usage:
If extra usage is enabled: additional Claude requests continue to flow at standard API rates (the same per-token pricing published on Anthropic’s pricing docs — $10/$50 MTok for Fable 5, $5/$25 MTok for Opus 4.8, $3/$15 for Sonnet 4.6, $1/$5 for Haiku 4.5). Extra usage is billed separately from the subscription. Team and Enterprise admins can enable, cap, and monitor extra usage at the organization level.
If extra usage is not enabled: the user’s Claude requests stop until their limit resets at the start of the next session window (seven days from when the current session started, not a fixed weekly day).
The right setting depends on your team’s tolerance for surprise bills versus interrupted workflows. Most production teams enable extra usage with a hard organizational cap so individual users have continuity but the org has predictable spend ceiling.
Claude Code has its own usage limit accounting that exists alongside chat usage limits. Per Anthropic’s Help Center on Claude Code models, usage, and limits (verified May 15, 2026):
claude -p mode currently also draws from subscription usage limits — until June 15, 2026The June 15 change is important enough that it gets its own breakdown in our Agent SDK Dual-Bucket Billing article. The short version: if you’re running unattended Claude Code work in cron jobs or CI, your billing model is changing. Plan capacity against the new credit pool.
Three things that get reported as limits but are actually configuration choices:
“My context window keeps filling up.” This is usually caused by long-running conversations accumulating history rather than the model’s actual context window being too small. Starting a new conversation (or running /clear in Claude Code) resets the working context. Long sessions are not a hard limit; they are a working-memory pressure that compounds over turns.
“Claude won’t read my whole repository.” Repository size is rarely the actual limit; the limit is how much you can load into the context window at once. Tools like Claude Code’s file reading and search work around this by loading files on demand rather than upfront. The 1M context window helps but is not a substitute for selective loading.
“My team keeps hitting limits even though we’re on Team.” Almost always one of two things: (a) people are mistakenly assuming the seat allowance is shared, when it’s strictly per-user; (b) someone is running heavy automation through a subscription seat instead of a Claude Developer Platform API key (which is the recommended path for sustained team-wide automation, especially after June 15).
Map your use case to the limits that actually apply:
claude -p.No. Per Anthropic’s Help Center documentation, each team member has their own set of usage limits. If one team member reaches their seat’s included limit, other team members are unaffected and can keep working.
30 MB per file for both uploads and downloads, per Anthropic’s official documentation. For PDFs larger than 30 MB, Claude can process them through its computing environment without loading them into the context window.
1M context is available on Claude Code with Opus 4.8 (Pro/Max/Team/Enterprise) and on the API with Fable 5, Opus 4.8, and Sonnet 4.6. Web chat tops out at 500K tokens even on the most capable models. Sonnet 4.6 in Claude Code requires extra usage to be enabled to access the 1M window (except on usage-based Enterprise plans).
Standard seats offer 1.25x Pro plan usage per session with one weekly limit across all models. Premium seats offer 6.25x Pro session usage with two weekly limits (one across all models, one Sonnet-specific). Both reset seven days after the session starts.
If extra usage is enabled at your organization, you continue at standard API rates billed separately. If extra usage is not enabled, your requests stop until your limit resets at the next session window (seven days from session start, not a fixed weekly day).
For sustained shared automation (CI pipelines, cron jobs, background services), Anthropic recommends the Claude Developer Platform with an API key over subscription seats. Subscription seats are sized for individual interactive use; API keys give you predictable pay-as-you-go billing, no per-seat caps, and don’t compete with team members’ interactive usage.
Sources reviewed May 15, 2026:
All limit numbers and policies are accurate as of May 15, 2026. Anthropic adjusts subscription mechanics regularly; if you’re making procurement decisions on this article more than 60 days from the date stamp, re-verify the per-seat multipliers and context window availability against the current Help Center.
Message limits vary by plan. Free: limited daily messages (Anthropic adjusts based on load). Pro ($20/month): approximately 2,000 Sonnet-equivalent messages per day. Max 5x ($100/month): approximately 10,000. Max 20x ($200/month): approximately 40,000. API users are rate-limited by tier with no hard daily message cap, instead governed by tokens-per-minute limits.
Claude Fable 5, Claude Opus 4.8, and Claude Sonnet 4.6 all support a 1 million token context window. Claude Haiku 4.5 supports 200,000 tokens. Anthropic eliminated long-context surcharges in March 2026, so large-context requests are billed at standard per-token rates. The Free plan is limited to 200K context even on Sonnet.
On Pro, Max, and Team plans: up to 30MB per file, up to 5 files per conversation. Supported formats include PDF, text, CSV, code files, and images. The Free tier supports up to 10MB per file. For API users, file uploads are handled via the Files API with a 32MB per file limit.
For high-volume workloads, switch to the Claude API (pay-per-token, no daily message cap beyond rate limits). Enterprise plans offer higher rate limits and custom agreements. The Batch API processes large jobs at 50% off standard prices for non-real-time workloads. Claude Max 20x ($200/month) is the highest subscription tier for interactive use.
API rate limits depend on your usage tier. New API accounts start at Tier 1 with lower limits. Spending history and account age automatically promote accounts to higher tiers with increased requests-per-minute and tokens-per-minute. Current tier limits are published at console.anthropic.com/settings/limits. Enterprise customers can negotiate custom rate limits.
There is no enforced per-message token limit separate from the overall context window. A single message can use up to the full context window (1M tokens for Fable 5 / Opus 4.8 / Sonnet 4.6, 200K for Haiku 4.5). However, very long single messages may be slower to process. The practical limit is the context window of whichever model you are using.
Updated June 10, 2026
Roadmap update: the May 2026 roadmap below has largely played out — Opus 4.8 and Claude Fable 5 have since shipped. As of June 10, 2026, Anthropic’s current lineup is Claude Fable 5 (the new top tier above Opus, $10 input / $50 output per MTok), Opus 4.8 ($5/$25), Sonnet 4.6 ($3/$15), and Haiku 4.5 ($1/$5). Full details: the Claude Fable 5 Complete Guide.
Last refreshed: May 15, 2026
The pace of new Claude releases in 2026 has been fast enough that the canonical question — “what’s the latest Claude model and what’s it actually good for?” — has a different answer almost every quarter. This article is the current map, dated and sourced, of what Anthropic has shipped in 2026, what’s confirmed about each model’s specs and knowledge cutoffs, and what’s been claimed (but not officially confirmed by Anthropic) about what’s coming next.
Two ground rules first, because the model-roadmap space is full of speculation:
If you’re making product decisions on this information, treat verified facts as actionable and reported facts as directional.
This was the production Claude lineup as of May 15, 2026. For the current lineup see the June 10 roadmap update above: Fable 5 ($10/$50), Opus 4.8 ($5/$25), Sonnet 4.6 ($3/$15), and Haiku 4.5 ($1/$5).
Claude Opus 4.7 — claude-opus-4-7
Claude Opus 4.6 — Still generally available, $5/MTok input, $25/MTok output. Released February 2026.
Claude Sonnet 4.6 — $3/MTok input, $15/MTok output. Includes the 1M token context window at standard pricing.
Claude Haiku 4.5 — Cheapest model in the active lineup at $1/MTok input, $5/MTok output.
Earlier models still active or in deprecation: Opus 4.5, Opus 4.1, Sonnet 4.5, and Haiku 3.5 (retired except on Bedrock and Vertex AI). Opus 4 and Sonnet 4 are listed as deprecated.
Per Anthropic’s Help Center article on training-data recency (verified May 15, 2026), the most recent generally-available models have January 2026 knowledge cutoffs. That means:
The practical version: don’t ask Claude what happened last week and expect it to know. Hand it the source material and ask it to analyze, summarize, or work with what you’ve given it.
Per Anthropic’s official pricing documentation (verified May 15, 2026), Opus 4.7, Opus 4.6, and Sonnet 4.6 all include the full 1 million token context window at standard pricing. There’s no long-context premium — a 900,000-token request is billed at the same per-token rate as a 9,000-token request.
That’s an enormous practical change from earlier Claude generations. A 1M context window is roughly:
Prompt caching and batch processing discounts both apply at standard rates across the full 1M window. For workloads that involve sending the same large document repeatedly with different questions, prompt caching against a 1M context is one of the highest-leverage cost optimizations available in the current Claude lineup.
Multiple third-party sources reported in early 2026 that Anthropic CEO Dario Amodei confirmed a Q2 2026 launch window for Claude 5 in a TechCrunch interview published February 1, 2026. The same sources cited an internal-roadmap leak suggesting an April 28 target date.
What we can verify as of May 15, 2026:
The honest read: Q2 2026 ends June 30, so if the reported timeline is accurate, an official Claude 5 announcement could plausibly land in the next several weeks. If you’re planning a project that depends on a specific Claude 5 capability, build against current Opus 4.7 first and treat any Claude 5-specific work as speculative until Anthropic publishes official model details.
Some 2026 third-party reporting refers to “Claude Sonnet 5” launching in early February 2026 under an internal codename. We could not, in our May 15, 2026 review, find this model listed in Anthropic’s official models overview, pricing page, or release notes — only Sonnet 4.6 and earlier Sonnet variants are listed as currently available models. If “Sonnet 5” was a real intermediate release, it does not appear in Anthropic’s current public model documentation under that name.
Two possibilities to consider, neither of which we can confirm: the reported Sonnet 5 may have been folded into the broader 4.x lineup under a different name, or the reporting may have been speculative or premature. If you’re tracking model identifiers for production use, only model IDs published in Anthropic’s documentation (such as claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5) are guaranteed to be valid against the API.
The signal-to-noise ratio in the model-release coverage space is not great. Two practical sources are reliable enough to bookmark:
For breaking changes that affect production code, the Anthropic platform documentation publishes per-version “What’s new” pages (Opus 4.7’s, for example, lists every API breaking change at launch). Those are the canonical reference for migration work.
For everything else — analyst commentary, predictions, leak coverage — treat it as commentary, not as fact.
Based on what is verifiable on May 15, 2026:
January 2026, per Anthropic’s Help Center documentation verified May 15, 2026. Information about events, products, or developments after that date is not in the model’s training data and must be provided directly.
1 million tokens, available on Claude Fable 5, Claude Opus 4.8, and Claude Sonnet 4.6 at standard pricing with no long-context premium. Claude Haiku 4.5 has a 200K context window.
As of May 15, 2026, we could not locate an Anthropic-published announcement of a Claude 5 model on anthropic.com or docs.claude.com. Multiple third-party sources have reported a Q2 2026 launch window based on a TechCrunch interview with Dario Amodei, but we could not independently confirm those specifications against a primary source.
As of May 15, 2026, “Claude Sonnet 5” does not appear in Anthropic’s official models overview or pricing documentation. The currently available Sonnet model is Claude Sonnet 4.6 (model ID claude-sonnet-4-6). Earlier reports of a Sonnet 5 release were not confirmed against an Anthropic-published source in our review.
Opus 4.7 ships with a new tokenizer that contributes to its improved performance but uses approximately 1x to 1.35x as many tokens for the same input text compared to previous models. Anthropic recommends increasing max_tokens headroom and adjusting compaction triggers accordingly.
No. Setting temperature, top_p, or top_k to any non-default value on Opus 4.7 returns a 400 error. Migration guidance: omit these parameters and use prompting to guide the model’s behavior.
Sources reviewed May 15, 2026:
This article applies the verified vs. reported distinction throughout. If any of the unverified third-party claims are confirmed by Anthropic in the weeks after this article’s date stamp, the relevant sections should be updated to reflect the new primary-source documentation.
Last refreshed: May 15, 2026
If you set up Claude MCP six months ago and have not touched the config since, three things have changed underneath you: the recommended transport, how tools are loaded into context, and how teams share server configs. None of these are cosmetic. If you ignore them, you are leaving tokens, money, and stability on the table.
This is the working Claude MCP setup I use in May 2026 — what the claude mcp add command actually does, which scope to pick, what the deprecation of SSE means in practice, and where Claude Code still falls short.
Every MCP server you wire into Claude Code lives at exactly one of three scopes. Get this wrong and you will either leak credentials into git or wonder why your teammate cannot use the same database the AI just queried.
~/.claude.json. Good for project-specific servers like a dev database or a Sentry project key you do not want other repos to inherit.~/.claude.json. This is where GitHub, search providers, and personal productivity servers belong..mcp.json file at the repo root and shared with the whole team via git. Claude Code prompts for approval the first time a teammate opens the project — by design, because anyone who can push to the repo can wire a new server into your environment.When the same server is defined in more than one scope, Claude Code resolves it in this order: local beats project beats user beats plugin-provided. This is the part that bites people the most. If you have a “github” entry at user scope and someone adds a different “github” entry at project scope in .mcp.json, the project definition wins for that repo. Run claude mcp list when something behaves strangely.
The CLI is more useful than the docs make it look. Three commands cover ~90% of real setup work:
# Add a remote HTTP MCP server at user scope (available everywhere)
claude mcp add --transport http hubspot --scope user https://mcp.hubspot.com/anthropic
# Add a local stdio server scoped only to this project
claude mcp add my-db -s local -- node ./scripts/db-mcp.js
# Share a server with your team via the repo's .mcp.json
claude mcp add my-server -s project -- node server.jsThe short flag is -s, the long is --scope. The -- separator is required for stdio servers because everything after it is treated as the literal command to spawn. Forget it and Claude Code will try to interpret your Node arguments as its own flags.
If your MCP server documentation still tells you to use the sse transport, the documentation is stale. The MCP spec dated 2025-03-26 introduced Streamable HTTP and simultaneously deprecated HTTP+SSE. Through 2026, vendor after vendor has set hard cutoff dates — Atlassian’s Rovo MCP server keeps SSE around until June 30, 2026 and then drops it; Keboola pulled SSE on April 1; Cumulocity’s AI Agent Manager flipped to Streamable HTTP on May 8.
Why this matters beyond a name change: SSE required Claude Code to hold a persistent connection to a single server replica, which broke horizontal scaling and made every transient network blip a reconnection drama. Streamable HTTP is stateless. Multiple replicas behind a load balancer just work. If you have flaky MCP connections in production, the first thing to check is whether the server is still on SSE.
For new setups, use --transport http. The older --transport sse still functions but is on the deprecation path.
The single biggest change in how Claude Code uses MCP in 2026 is lazy tool loading via Tool Search. Older MCP clients dumped every tool schema from every connected server into the model’s context window at the start of every conversation. With ten servers wired up that could easily be 20,000+ tokens of overhead before you typed a single character.
Tool Search inverts this. Claude Code keeps only the server names and short descriptions resident. When a tool is actually needed, it fetches that tool’s full schema on demand. Anthropic’s own documentation says this reduces tool-definition context usage by roughly 95% versus eager-loading clients. In practice that means you can run a serious MCP fleet — GitHub, Sentry, a database, a search provider, your internal API — without quietly burning through your context budget. The Sonnet 4.6 and Opus 4.7 1M-token context window does not save you here, because anything you let crowd the prompt is also being re-read on every turn.
Companion feature: list_changed notifications. An MCP server can now tell Claude Code “my tool list changed” and Claude Code refreshes capabilities without a disconnect-reconnect dance. If you build your own server, emit this when you swap tool definitions and you save users a restart.
Honest take: claude mcp list still does not surface scope information for every entry in a useful way — there is an open issue on the anthropics/claude-code repo asking for it (#8288 if you want to track). Project-scoped servers from .mcp.json have a separate history of not appearing in the list output (#5963) depending on how you opened the project. If you cannot find a server, check both ~/.claude.json and ./.mcp.json directly.
The other rough edge is the project-approval prompt. The first time you open a repo with a new .mcp.json, Claude Code asks you to approve each project-scoped server. That is the right security default. It is also infuriating in CI or any non-interactive shell, where the prompt blocks the session. The current workaround is to bake the servers in at user scope on build agents so the project-scope approval never fires in CI. A cleaner non-interactive approval flow is the single most-requested fix I see in real teams.
User-scope: GitHub, a code search server, and a single notes/Notion server. Project-scope in each repo’s .mcp.json: whatever database the project owns and whatever observability backend it reports to. Local-scope: anything experimental I am evaluating but do not want my team or my other repos to inherit.
Pin --transport http on everything remote. Skip Desktop Extensions (.dxt) for anything you want versioned with the codebase — they are a Claude Desktop convenience, not a Claude Code primitive, and they hide the config from your team. Run claude mcp list when something is off and read .mcp.json directly when list is unhelpful.
That is the whole working model. The pieces that matter — three scopes, Streamable HTTP, Tool Search — fit on a single screen. The pieces that have not caught up yet — list output, non-interactive approvals — are visible in the issue tracker and will move.
There’s a simple version of the AI-in-organizations problem that’s wrong: you build the system, give it access to the right data, write a thorough system prompt, and it operates in your organizational context. The prompt is the context. The context is the prompt.
This framing is everywhere. It’s also the reason most organizational AI deployments produce work that is technically correct and somehow off.
The context that matters — the context that determines whether a decision lands right, whether a draft feels aligned, whether a flagged opportunity is genuinely actionable — is not stored anywhere. It lives between people.
Every organization operates on a layer of standing assumptions that nobody explicitly maintains and nobody could fully articulate on request. Not values, not principles, not priorities — something below those. The interpretive substrate that makes the documented values mean anything.
When someone joins a team and violates one of these assumptions — proposes the wrong thing in the wrong meeting, pushes a decision that is technically within their authority but somehow not theirs to make, surfaces a priority the organization agreed to de-emphasize without announcing it — everyone feels it. The violator usually doesn’t. The substance was fine. Something else was wrong.
That something else is the context AI systems don’t have.
Documentation can encode explicit knowledge. It cannot encode the community that makes the documentation mean anything.
A system prompt can say “this organization prioritizes speed over perfect.” What it cannot encode is whether that norm has actually been consistent for the last six months, or whether leadership has been quietly walking it back after three bad launches, or whether it applies to customer-facing work but not internal infrastructure, or whether the one person whose approval you need is the one exception to the norm.
The standing assumptions are not stored. They are enacted. They show up in what gets committed to and what sits in the inbox for thirty days.
Watch a team’s queue long enough and you can read the context. Not from the items themselves — from the pattern of what moves and what doesn’t. Stalled items tell you which commitments have real backing and which are aspirational. Rapid movement in one lane tells you where the actual authority is concentrated. The gap between what the organization says it prioritizes and what it actually processes is a map of the standing assumptions it hasn’t named.
A single operator can solve this. They can read the board, feel the friction, and say: the predicate is wrong. The item needs to be reframed before it moves. They can do this because they hold the context in their own head, accumulated over months, updated daily.
A team cannot do this as easily. The context is distributed. Each person holds part of it. The standing assumptions live in the gaps between what anyone would say individually. Ask the team to write down why something has been stalled for thirty days and you’ll get five different answers, each of which is partially true, none of which is sufficient.
The naive solution is documentation. Write the standing assumptions down. Build a better system prompt. Give the AI more context.
This helps at the margins. It doesn’t solve the problem.
Documentation of standing assumptions produces a different artifact — a curated version of the context, shaped by whoever did the writing, frozen at the moment of writing, immediately in tension with the organizational reality it was supposed to encode. It becomes a reference document. The context moves on. The document does not.
The less naive solution — the one organizations rarely take — is to treat context as an ongoing artifact rather than a static one. Not a document but a practice. Something that gets updated not when someone decides to update it, but when a decision is made that the prior version couldn’t have predicted.
Every time a team makes a decision that would have surprised an outside observer, that decision contains information about the organizational context. The surprise is the data. The question is whether anyone captures it — not as documentation but as signal, living in the same system as the work itself.
This is not how most organizational AI deployments are built. They treat context as given — encoded once, referenced forward. The system prompt goes stale six weeks in and nobody notices because the outputs are still technically correct. The work product is fine. The alignment is drifting.
A system that can only read your context is a tool. A system that reads the gaps between your documented context and your actual decisions is starting to understand something harder to name.
The implication isn’t that AI systems need more access. More access to documented context doesn’t help if the relevant context isn’t documented. The implication is that organizational deployment requires a different architecture: one where the context layer is treated as a first-class input that needs active maintenance, and where the signal for updating it is not a calendar prompt but a decision that contradicts the prior version.
This is harder to build than a thorough system prompt. It requires the organization to treat its own implicit knowledge as an artifact worth maintaining — which means surfacing it, which requires the uncomfortable process of naming standing assumptions that everyone was benefiting from not naming.
The systems that work at organizational scale will have solved this. Not by encoding context better but by treating context as a process rather than a state.
Prior pieces in this series have addressed the individual operator: memory as infrastructure, capture versus commitment, the discipline of waiting. Those all assumed a single person holding the context in their own head, updated daily, acted on personally.
The team changes the shape of the problem. Not because teams are harder — though they are — but because the context is no longer located anywhere. It exists only in the aggregate of how the team behaves, and that aggregate is not readable from any single vantage point, including the AI’s.
The context lives between people. You cannot put it in the prompt. The first step is admitting that.
The second step — what an organization can actually do about it — is less clean than any framework suggests, and probably requires a different piece.