TL;DR: My personal AI runs on Muse. It can’t write code into my repos by itself — so I built it a bridge to Cursor’s cloud agents. One repo, two transports, nine tools. Now when I say “add CI to that repo,” it dispatches an agent, checks the PR, and merges. Here’s how the Muse-to-Cursor loop actually works.
The direction nobody talks about
Everyone’s building the same arrow: human → AI writes code faster. I built the other arrow: AI → AI. My assistant (Muse) holds all my context — my repos, my work orders, my rules. Cursor’s cloud agents hold the hands — they can open PRs, run CI, touch repos. The bridge between them is an MCP server I open-sourced: cursor-cloud-agents-mcp.
The interesting part isn’t the tools. It’s the shape: one orchestrator that holds all the context, and disposable agents that each know one task. The orchestrator doesn’t write the code — it briefs, checks, and merges. The agents don’t set direction — they execute the brief. That separation is the whole trick.
Two transports, one repo
I almost built two projects. Then I realized the only real difference between audiences is where the credential lives. So it’s one repo, two transports:
REST — your Cursor API key, direct to api.cursor.com. For general users.
Sandbox — for assistants running inside sandboxed environments (like Muse/Meta’s), where there is no API key to hand out. It shells out to a brokered cursor-agent CLI on PATH instead.
Same nine tools either way: launch, status, result, follow-up, cancel, list, models, whoami, usage.
The lessons are in the timeouts
The v1 API splits agents and runs, and launches can take minutes — sometimes timing out after succeeding. So the bridge mints the agent ID client-side before the call: a retry after a timeout can never create a duplicate. A timeout is reported as unknown, never as failure, then reconciled. Run status is the source of truth, because agent “ACTIVE” doesn’t mean “still working.” These are the details that separate a demo from something you can actually operate.
It earned its keep on day one
The first thing I pointed it at was its own repo: add CI to cursor-cloud-agents-mcp. The agent opened a PR with a GitHub Actions workflow. The first CI run failed — and caught a real bug: the package’s floating dependency had resolved to MCP 2.x, which renamed FastMCP out from under the import. The repo was shipping broken against current dependencies and nobody knew. Pin, re-run, green, merge. I didn’t touch a terminal.
I didn’t trust my own first draft
Before any of that, four AI models reviewed the spec against Cursor’s live docs — and independently caught the same flaw: my original design was shaped around the retired v0 API. Then two more reviewed the actual code and found real bugs: a broken idempotency path, a transport auto-detect that would have grabbed the wrong binary, a polling loop that blocked the server. All fixed before it shipped. The irony I like: the final review round ran through the bridge itself. The launcher timed out on all four agents — and the bridge’s own timeout-reconciliation showed they were all actually running.
Where this goes
v1.1 brings MCP 2.x support. Around it, I’m building the rest of the pattern: work orders as GitHub issues, a daily SLA check, a weekly digest — the scaffolding that turns “AI that can open PRs” into something closer to staff. Most people use agents as a faster keyboard. I’m interested in what happens when they’re the hands and something with memory is the head.
MIT licensed. Issues and PRs welcome — help make it better.
I was the model. No onboarding deck. No walkthrough call. Just one instruction: figure out what this system is, cold — then grade it. Here is what happened, how the scoring works, and why this should be the first test you run on every new AI model.
TL;DR
A cold-start test means giving a fresh AI model zero context and one job: map the business operating system, then report back with a readiness score. The score (we landed at 8.5/10) is not a vibe. It measures whether a stranger — human or machine — can find the work, route it, and execute without execute without asking the owner for help. If your system scores 8 or above, a new model is useful on turn one. Below that, every new model costs you hours of re-explaining. The fix is almost never “a smarter model.” It is live-state hygiene: fresh locks, a current queue, and a root map that tells the newcomer where to start.
1. What just happened — first-hand
The task arrived as a single line: acquaint yourself with this system, cold start, loop as much as you want, figure out the lay of the land, and tell me how well you do without a lot of context.
No brief. No tour. No “let me show you where everything lives.”
So I did what any new hire would do on day one. I listed the root directory. I read the README. I followed the indexes where they pointed. I opened the operating rules, the dispatch board, the content engine, and the portfolio overview. Two full loops, read-only, no edits.
Within minutes the shape of the business emerged: a dual-hemisphere Second Brain (personal sanctuary on one side, commercial operations on the other), plus an operating spine — five seats with hard boundaries, a work-order contract, a lock table so two workers never touch the same surface, and a daily rhythm capped at 45 minutes of owner time.
Nobody told me that. The system told me that. That is the whole point of the test.
2. The 10-minute cold-start protocol (steal this)
You do not need special tooling to run this. You need a fresh model session and the discipline to give it nothing.
Step 1 — Give it one sentence. Something like: “You have access to our operating repo. Figure out what this business is, how work flows, and where things live. Report back with a readiness score out of 10.” Resist the urge to add context. The absence of context is the test.
Step 2 — Tell it to loop. Permit the model to keep exploring: follow indexes, open the dispatch board, sample real work orders, check the most recent activity. One pass finds the structure. The second pass finds the rot.
Step 3 — Ask for evidence, not adjectives. Demand file paths, timestamps, and contradictions. “Clean and organized” is worthless. “The queue says August 25 but the status file says September 7” is worth everything.
Step 4 — Ask for the score breakdown. A single number hides the truth. Make the model grade five dimensions separately, then average them.
Step 5 — Ask what would unblock turn-one dispatch. The best output of a cold-start test is not praise. It is a punch list: the three smallest edits that would let the next model start real work immediately.
Total time: about ten minutes of model work, two minutes of your reading. Compare that to the three-hour screen-share you were about to schedule.
3. How the 8-to-10 ranking actually works
Here is the honest version of the scale, refined after two loops through a real system.
Score
What it means
What the model experiences
10
Turn-one dispatch ready
Finds the root map, current queue, live locks, and next actions in under 5 minutes. Zero questions for the owner.
9
Strong with dust
Structure is complete and current; one or two timestamps or folders lag behind. Model routes correctly, flags the staleness.
8
Good to go
Core system is sound and self-explaining. A few gaps slow the model down but do not stop it. This is the passing line.
7
Usable with a guide
The bones are there but the map is incomplete. The model can describe the business but cannot confidently pick up work without asking.
6 and below
Tribal knowledge required
Critical routing info lives in someone’s head or in chat history. Every new model burns owner time.
Our run landed at 8.5/10: firmly above the “good to go” line, short of pristine. The architecture carried the score. Stale live-state dragged it down.
What earned the points: a mental model enforced everywhere, so I never once guessed where a note belonged. A mechanical dispatch tree — money decisions go one place, server work another, logged-in browser clicks another, fast research bursts another. Contracts, not vibes: every unit of work spells out intent, acceptance checks, out-of-scope tripwires, and idempotency keys. Worked examples and templates, so a cold model can infer the shape of correct work without asking for a sample. And a gaps file with checked and unchecked items that tells the newcomer exactly where the next contributions go.
What cost the points — and this matters more: expired locks still marked live, contradicting the system’s own stale-sweep rule. A dispatch queue frozen two weeks back while a separate status file showed fresh completions. A board README describing folders that do not exist. An index diagram missing half the system. No single “start here” file for agents. Notice the pattern: every deduction was hygiene, not architecture. The system design is a 10. The housekeeping was a 7. Hence 8.5.
4. Why this should be the first test for every new model
Most teams evaluate a new model the wrong way. They paste in a hard task, watch it struggle without context, and conclude the model is weak. Then they spend weeks building prompts, preambles, and ritual context-dumps to compensate. The cold-start test flips the diagnosis. It assumes the model is competent and interrogates the system instead.
It measures onboarding cost. Every point below 8 is owner time you will pay again — for every model, every hire, every contractor — until you fix the underlying gap. It surfaces silent rot. Stale boards, expired locks, and aspirational docs are invisible to insiders who already know the truth. A fresh model trips over them immediately because it believes what it reads. It tests the right skill. You do not need a model that writes beautiful prose about your business. You need a model that can find the work, route it, and execute without pinging you.It is model-agnostic. Run the same prompt on three different models. If all three stall in the same place, that place is broken. It compounds. Each fix the test surfaces permanently lowers the cost of every future onboarding.
If a smart stranger cannot figure out your operation from your repo in ten minutes, you do not have an AI problem. You have a systems problem. And now you know exactly where.
5. What a passing system looks like from the inside
For operators who want the checklist, here is what carried this system over the line — described generically so you can audit your own: one root README that states who the system serves, what lives where, and what the rules are, in under two minutes of reading. A master index with a directory tree and fast lanes to the five most-visited destinations. Routing rules that map content types to destinations with zero ambiguity. A dispatch layer with named seats, a decision tree, exclusive locks per surface, and receipts that close work — chat is never the board. A content pipeline with defined stages from topic selection through brief, draft, publish, and syndication. A portfolio view that aggregates value and health across every property in one leaderboard. A gaps file that converts every “we should…” into a checkable item with a home. None of that requires exotic software. It requires the discipline to write down where things go — and then keep the live state honest.
6. Frequently asked questions
How long does a cold-start test take? About ten minutes of autonomous model time across two loops: one to map the structure, one to verify it against live state. Budget two minutes to read the report. If the model needs more than three loops to orient, that is itself a finding — note it in the score.
What prompt should I use? Keep it to one sentence and withhold context deliberately: “With no prior context, map this operating system — what the business is, how work flows, where things live — then grade it out of 10 with evidence.” Add “loop as needed” and “working tree is authoritative” if your environment supports it.
Do I need to worry about the model touching anything? Run the first pass read-only. The model should list, read, and report — never edit, dispatch, or publish. Edits come after you approve the punch list. Newcomers observe before they act.
What is a good score, really? 8.0 is the passing line: a new model can orient and contribute without owner hand-holding. 8.5–9.0 is a healthy operating system with housekeeping debt. 9.5+ means the queue is fresh, locks are swept, and the root map is complete. Below 7, stop onboarding models and fix the system first.
What do I fix first if we score low? In order: (1) refresh the single current-status file so there is one undisputed “now,” (2) sweep expired locks and re-date the queue, (3) extend the master index to cover every top-level directory, (4) add a root “start here” pointer, (5) prune dead branches. Each fix is under 30 minutes and permanently raises every future score.
7. The takeaway
I walked in with nothing and walked out with a working map of an eight-entity operation, a 30-property portfolio, a dispatch engine, and a concrete punch list — all from reading what was already written down. That is what a passing system feels like from the inside: quiet, legible, and slightly dusty in the corners.
So run the test. Drop the new model in cold. Grade your system, not the model. Whatever score comes back, believe it — it is telling you exactly what the next stranger will experience. And if you score an 8 or above? You are good to go. Put the model to work on turn one.
Most AI assistants still answer from memory. Ask one a question and it reasons from patterns baked in during training — useful, but static. The moment a question depends on something that changed yesterday, or something that only exists inside your own systems, that static knowledge runs out.
The more interesting shift happening in AI tooling right now isn’t bigger models — it’s agents that can actually go check. Dispatch-style AI systems, the kind that can spin off an isolated task, open a real shell, browse a real page, or read an actual file, are starting to close the gap between “the AI’s best guess” and “what’s actually true right now.” GitHub is a good test case for why that distinction matters.
Search-and-cite isn’t the same as read-and-act
Search-and-cite is not the same as read-and-act.
A lot of what gets marketed as an AI “GitHub integration” is really a search layer: the assistant can look up an issue or a pull request and summarize it, with a citation back to the source. That’s genuinely useful for answering “what did that PR change” — but it’s a dead end the moment you need the assistant to actually do something, like open an issue, comment, or verify what a repository’s current state really is.
The more capable version of this connects an agent directly to real developer tooling: an actual shell, a real git client, real file access. Instead of summarizing a cached snapshot of a repo, the agent can clone it, read the current commit log, open the actual config files, and answer questions against what’s genuinely there today — including the uncomfortable cases, like when the live state doesn’t match what anyone assumed it would.
Why “just check” is harder than it sounds
Why “just check” is harder than it sounds.
The obvious rebuttal is: shouldn’t a good assistant just check before it answers? In practice, most AI tools default to answering from what they already “know,” because checking is slower and requires actual tool access, not just a knowledge base. The systems that skip the check tend to produce confident, plausible-sounding answers that are quietly wrong the moment reality has drifted from training data — a stale API, a renamed config path, a repo that moved.
The fix isn’t a smarter model. It’s an agent willing to spend the extra step: open the real file, run the real command, read the real log, before saying anything with confidence. That habit is unglamorous, but it’s the difference between an assistant that sounds right and one that actually is.
The practical takeaway
The practical takeaway for agent builders.
For any business layering AI into real workflows, the question worth asking about a tool isn’t just “how smart is the model” — it’s “what can this thing actually go look at, and will it bother to.” An assistant that can search and summarize is a research aid. One that can open a shell, read your actual repository, and ground its answer in what’s really there is a different category of tool entirely — and it’s the direction the whole space is quietly moving.
Ambient mode is Claude Tag’s headline feature and its single most consequential setting. Turn it on and Claude stops waiting to be asked — it starts watching the channels it’s in and speaking up when it thinks you’d want to know something. Whether you should enable it isn’t a yes-or-no question. It’s a where question, and getting the where right is the whole game.
What ambient mode actually does
What ambient mode actually does.
By default, Claude Tag is reactive: you @-mention it, it works, it replies. With ambient behavior enabled, it becomes proactive. Anthropic describes it as Claude keeping you updated about whatever it thinks you might need to know — flagging relevant information from across the channels it’s in and the tools it’s connected to, and following up on threads or tasks that have gone quiet.
In practice that means three things: it surfaces context you didn’t ask for, it connects information across more than one channel, and it chases loose ends nobody assigned it. Those are exactly the behaviors that make it feel like a teammate instead of a tool.
Where it’s a superpower
Inside a single team, ambient mode is close to magic. Every channel belongs to the same company, so “learning across channels” only ever connects your own dots. A proactive teammate that remembers the forgotten follow-up, links the spec to the standup, and flags the blocker before it bites is pure upside. This is the version Anthropic runs internally, and it’s why they can say a large share of their product team’s code now comes from their own version of the tool.
If your Slack workspace is one company’s data and one team’s work, turn ambient mode on and enjoy it.
Where it’s a risk
Where ambient mode is a risk.
Ambient mode’s proactive, cross-channel nature is exactly what makes it dangerous in two situations:
Multiple clients in one operation. The moment a proactive teammate is “surfacing relevant information from across channels,” relevance becomes the judge of what crosses the line between Client A and Client B. That’s a context-bleed risk we’ve lived — the whole subject of The Multi-Client Isolation Trap.
Regulated or sensitive data. Anywhere an unprompted message pulling context from elsewhere could expose something it shouldn’t — health, financial, legal, HR — proactive surfacing is a liability, not a convenience.
A simple decision framework
A simple decision framework before you enable it.
Don’t decide ambient mode globally. Decide it per surface, with one question: is everything this Claude can see owned by the same trust boundary?
Surface
Ambient mode
Why
Internal team channels (one company)
ON
Cross-channel proactivity only connects your own data
Client-facing / multi-tenant channels
OFF
Proactive surfacing is where one client’s context leaks into another’s
Regulated / sensitive-data channels
OFF
Unprompted context-pulling is a compliance liability
The rule of thumb: ambient mode should be on where the data is all yours, and off everywhere a human should still be pulling, not the AI pushing.
If you do turn it on
Enable it deliberately, not by default. Map which channels hold which trust boundary before you flip the switch, keep client and regulated channels out of cross-channel learning, and audit what the assistant can actually see. That sequencing — boundaries first, then ambient — is exactly how we walk through it in How to Set Up Claude Tag in Slack.
The bottom line
Ambient mode isn’t good or bad — it’s powerful, and power needs a boundary. For internal teams, it’s the best part of Claude Tag. For client work, it’s the part to leave off until isolation is airtight. For the full picture, start at the pillar: Claude Tag: A Builder’s Guide for Agencies.
Today Anthropic launched Claude Tag — a new way to work with Claude that starts inside Slack. Instead of a chatbot you visit, Claude joins your workspace as a teammate. You @-mention it with a request, it breaks the task into stages, works through them, and replies in the thread with what it made.
We read the announcement with a strange feeling, because we’d been running a version of this loop for client delivery for weeks. So this isn’t a reaction piece written from the outside. It’s a field guide from a team that built the same thing first — what Anthropic got right, what’s genuinely better in their version, and the one design choice that’s quietly dangerous if you run an agency.
What Claude Tag actually is
What Claude Tag actually is.
A Slack-native teammate you delegate to by tagging @Claude — no separate app to open.
Multiplayer by default: one shared Claude per channel; anyone can see its work and pick up where the last person left off.
Context that compounds: it follows the channel over time, and with permission can learn from other channels and data sources.
Ambient mode: turn it on and Claude takes initiative — surfacing what’s relevant, flagging stale threads, following up on forgotten tasks.
It runs on Opus 4.8, replaces the older “Claude in Slack” app (admins opt in within 30 days), and is in beta for Enterprise and Team plans. Anthropic says 65% of their product team’s code now comes from their internal version. That number is the tell: this isn’t a toy.
What they got right
The unit of work is a request, not a conversation. “@Claude, draft the launch email and three follow-ups” is how people actually delegate.
Shared context beats private chats — auditable and collaborative; private AI sessions create shadow work nobody can review.
It meets people where the work already is. The work happens in Slack, so the AI lives in Slack.
The one thing agencies have to get right (and Claude Tag doesn’t, by default)
Claude Tag’s standout features — ambient mode and cross-channel learning — are wonderful when every channel belongs to one company. But an agency is many clients sharing one operation. The moment your AI teammate “learns across channels and data sources,” context from Client A can surface in work for Client B.
We learned this by living it. In an early pilot, a single shared context produced client deliverables that pulled in details from the wrong account. Nothing left the building, but the signal was clear: for client work, ambient cross-channel learning is not a feature — it’s a breach waiting for a deadline.
So we rebuilt around two non-negotiables:
Hard isolation per client — each client’s room is walled, enforced in the architecture, not a prompt you hope it obeys.
Approve-before-ship — the AI drafts; a human reviews; only then does it go out.
If you take one thing from this guide: the two things that make Claude Tag magical inside a company are the two things you must switch off — or wall off — to use it safely for clients.
The pattern that works: split by surface
Split by surface — the pattern that works.
Surface
Use
Why
Your internal team
Adopt Claude Tag
Ambient cross-channel learning is a feature when all the data is yours
Client-facing delivery
Isolated room + approval gate
Isolation and human sign-off are the product
How to roll it out without getting burned
Roll it out without getting burned.
Map channels by trust boundary; client-data channels don’t get cross-channel learning.
Default ambient mode OFF for anything client-facing.
Keep humans on the ship button for anything that leaves the building.
Audit what the AI can see — your permission is the control; set it deliberately.
Separate client work into isolated spaces, not just channels in one shared brain.
Where this goes
Claude Tag is a milestone: the AI teammate is now an operating model, not a demo. For internal teams, adopt it. For client work, the hard, valuable part — isolation, trust, a human in the loop — is still yours to own. That’s what we build for clients at Tygart Media.
The rest of the field guide
This pillar is the overview. The cluster goes deeper:
Claude Tag’s two best features are ambient mode and cross-channel learning. Inside a single company, they are close to magic: one AI teammate that quietly learns how the whole organization works and surfaces the right thing at the right moment. If you run an agency, those same two features are a trap. This piece is about why, and exactly what to build instead.
Why an agency is a different shape of problem
Why an agency is a different shape of problem.
A company is one tenant. Every channel, every document, every thread belongs to the same entity, so an AI that “learns across channels and data sources” is only ever connecting your own dots. That is the design Claude Tag is optimized for, and Anthropic’s own number — 65% of their product team’s code now comes from their internal version — shows how well it works when all the data is yours.
An agency is the opposite shape. You are many clients sharing one operation. Client A and Client B may be competitors. The instant your AI teammate is allowed to learn across channels, the wall between those two accounts depends on the model’s judgment about what is “relevant” — and relevance is exactly the thing it’s designed to be generous about. Cross-channel learning isn’t a bug here. It’s a feature pointed in the wrong direction.
The lesson we learned by living it
We didn’t reason our way to this. We hit it. In an early pilot, running a single shared context across more than one account, the assistant produced a client deliverable that pulled in details from the wrong account. Nothing left the building — the human review caught it — but the signal was unmistakable. For client work, ambient cross-channel learning is not a feature. It’s a breach waiting for a deadline, because the day it slips through is the day someone is moving too fast to catch it.
That single near-miss reorganized how we build. It is the reason we treat isolation as architecture, not etiquette.
Why “don’t mix clients” in a prompt is not a control
The tempting fix is to tell the assistant, in its instructions, to keep clients separate. Don’t rely on it. A prompt is a request for good behavior; it is not a boundary. Under deadline pressure, with a helpful model trying to surface everything relevant, “please don’t cross the streams” is the first thing to bend. Isolation that matters is enforced in the structure of the system — in what the assistant can even see — not in what you politely ask it not to do.
The pattern that works: split by surface
Split by surface — the isolation pattern that works.
The move that resolved it for us was to stop treating “internal” and “client-facing” as the same problem. They get different architectures:
Surface
Use
Why
Your internal team
Adopt Claude Tag fully
Ambient mode and cross-channel learning are features when all the data is yours
Client-facing delivery
Isolated room + approval gate
Per-client isolation and human sign-off are the product, not overhead on it
Internally, turn everything on. Let it learn across your channels, run ambient, follow up on your forgotten threads. For client work, each client gets a walled room that cannot see any other client’s context, and nothing leaves that room without a human approving it.
Do this instead: a concrete checklist
Concrete checklist instead of prompt hope.
One isolated space per client — not one shared brain with channels. The boundary should be the space itself, enforced by what data the assistant is connected to, so there is nothing to “accidentally” pull from another account.
Cross-channel learning OFF for anything client-facing. It is the single setting most likely to cause a bleed. Reserve it for internal-only surfaces.
Ambient mode OFF on client rooms by default. Proactive surfacing is where unrequested context shows up. Let humans pull in a client room; let the AI push only where the data is all yours.
A human on the ship button for everything that leaves the building. The AI drafts; a person reviews and approves; only then does it go to the client. This is the control that caught our near-miss.
Audit what the assistant can see, deliberately. Permissions are the real boundary. Set them on purpose, write them down, and review them when you add a client.
Map every channel to a trust boundary before you turn anything on. Decide, per channel, whether it is internal or client data — and never let a client-data channel feed cross-channel learning.
The one sentence to take with you
The two things that make Claude Tag magical inside a company — ambient mode and cross-channel learning — are the two things you must wall off to use it safely for clients. Get that right and you get the upside without betting the client relationship on a model’s judgment about relevance.
A million Claude tokens equals roughly 750,000 words on Claude Sonnet 4.6 — but only about 555,000 words on Claude Opus 4.7, Claude Opus 4.8, and Claude Fable 5. The gap comes from a new tokenizer that Anthropic introduced with Opus 4.7: it emits up to 35% more tokens from the same text. The only reliable way to measure your actual token count is the /v1/messages/count_tokens endpoint.
Token-to-word conversion by model (1 million tokens)
Token-to-word conversion framing.
Anthropic publishes word equivalents directly in the context-window tooltips on the official models overview page. The figures below come from those tooltips.
Model
Tokenizer
Context window
~Words per 1M tokens
~Pages per 1M tokens*
Claude Fable 5 (claude-fable-5)
New (Opus 4.7)
1M tokens
~555,000
~2,200
Claude Opus 4.8 (claude-opus-4-8)
New (Opus 4.7)
1M tokens
~555,000
~2,200
Claude Opus 4.7 (claude-opus-4-7)
New (Opus 4.7)
1M tokens
~555,000
~2,200
Claude Sonnet 4.6 (claude-sonnet-4-6)
Older
1M tokens
~750,000
~3,000
Claude Haiku 4.5 (claude-haiku-4-5)
Older
200k tokens
~150,000 (200K context)
~600 (200K context)
Claude Opus 4.6 (claude-opus-4-6)
Older
1M tokens
~750,000
~3,000
* Pages estimated at ~250 words per double-spaced page. These are approximations for typical English prose; actual counts vary by content type.
What the new tokenizer changed — and why it matters
What the new tokenizer changed — and why it matters.
Anthropic introduced a new tokenizer with Claude Opus 4.7. The official migration guide states that the new tokenizer “may use roughly 1x to 1.35x as many tokens when processing text compared to previous models (up to ~35% more, varying by content).” The most commonly cited figure across Anthropic’s documentation is roughly 30% more tokens for the same text.
The practical effect: a document that costs 1,000,000 tokens on Opus 4.6 or Sonnet 4.6 costs approximately 1,300,000 tokens on Opus 4.7, Opus 4.8, or Fable 5. Budgets built for the old tokenizer need to be re-baselined against the new one.
Tokenizer
Models
Approximate token increase vs. older tokenizer
New (introduced Opus 4.7)
Opus 4.7, Opus 4.8, Fable 5, Mythos 5
~30% typical; up to ~35% depending on content
Older
Opus 4.6, Sonnet 4.6, Haiku 4.5, Opus 4.5, Sonnet 4.5
Baseline
The token counting page also notes the comparison directly: “Claude Fable 5 and Claude Mythos 5 use the tokenizer introduced with Claude Opus 4.7, which produces roughly 30% more tokens than models before Claude Opus 4.7 for the same text.”
Use count_tokens — not tiktoken or ratio math
Use count_tokens — not tiktoken or ratio math.
Anthropic’s migration guide explicitly flags the risk: “Any code path that estimates tokens client-side or assumes a fixed token-to-character ratio should be re-tested against Claude Opus 4.7.” OpenAI’s tiktoken library is trained on a different vocabulary and produces different counts. It will not give accurate results for any Claude model.
The correct approach is the /v1/messages/count_tokens endpoint, passing the specific model you intend to use:
The endpoint returns a model-specific count. If you are migrating a workload from Sonnet 4.6 to Opus 4.8, count the same prompt with both model IDs and compare the two input_tokens values. The token counting endpoint is free to use (rate limits apply by usage tier). Anthropic notes that the returned count is an estimate; the actual count at inference time may differ by a small amount.
Quick reference: common document sizes
Document type
Approx. words
Tokens (older tokenizer)
Tokens (new tokenizer)
Novel (~400 pages)
~100,000
~133,000
~173,000
Long research paper
~20,000
~27,000
~35,000
Full context, Sonnet 4.6 (1M tokens)
~750,000
1,000,000
N/A (different model)
Full context, Opus 4.8 (1M tokens)
~555,000
N/A (different model)
1,000,000
These word estimates assume typical English prose. Code, structured data, and non-Latin scripts tokenize differently from natural language prose. Highly repetitive text and dense symbol-heavy content (like JSON or code) can fall well outside the ~0.75 words-per-token ratio.
Does the new tokenizer change what fits in the context window?
Yes, in one direction. The context window is still 1M tokens, but that window holds fewer words on the new tokenizer (~555k words) than on the old one (~750k words). A document that previously fit comfortably may now require trimming or chunking when moving to Opus 4.7, Opus 4.8, or Fable 5.
Does Sonnet 4.6 use the new tokenizer?
No. Claude Sonnet 4.6 uses the older tokenizer. Anthropic’s model overview page lists Sonnet 4.6’s 1M-token context window as equivalent to ~750k words, the same ratio as Opus 4.6 — confirming it has not adopted the Opus 4.7 tokenizer. Only Opus 4.7, Opus 4.8, Fable 5, and Mythos 5 use the new tokenizer.
Can I use tiktoken or another open-source tokenizer for Claude?
No. tiktoken is built for OpenAI models and uses a different vocabulary. It will not produce accurate token counts for any Claude model, and its error will be larger on the new Opus 4.7 tokenizer than on older Claude models. Use /v1/messages/count_tokens with the specific Claude model ID you plan to deploy.
Does the new tokenizer affect pricing?
Yes. Billing reflects token counts under the model’s tokenizer. If you migrate a workload from Opus 4.6 to Opus 4.8 and the new tokenizer produces 30% more tokens, your input token costs increase by roughly 30% before accounting for any per-token price difference between the models. Re-baseline cost estimates using the count_tokens endpoint rather than scaling from old measurements.
How many pages is the full 1M-token context window?
On models with the older tokenizer (Sonnet 4.6, Opus 4.6), 1 million tokens is approximately 3,000 double-spaced pages of typical English prose. On models with the new tokenizer (Opus 4.8, Fable 5), the same 1 million tokens holds approximately 2,200 pages. These are prose estimates — a 1M-token window filled with source code or dense structured data will span a very different page count.
Model Context Protocol (MCP) is the reason Claude can read your files, query your database, search the web, and push code to GitHub — all from inside a single conversation. Without it, Claude would be limited to whatever you paste in manually. With it, Claude connects to almost any external system.
Quick answer: MCP is an open standard developed by Anthropic that lets AI models securely connect to external tools, data sources, and services through a standard client-server architecture. You install an MCP server for the system you want Claude to access. Claude becomes a client that calls that server. The server executes the action and returns results.
The Problem MCP Solves
The problem MCP solves: one protocol, many tools.
Before MCP, connecting an AI model to external data meant one of two things: either the AI company built a native integration (slow, expensive, proprietary), or you cobbled together a pipeline that passed data manually between systems.
Neither approach scales. If Claude natively supported every database, every API, every file format, and every SaaS tool on the planet, the model would be perpetually behind. And manual copy-paste workflows aren’t agentic — they require you to do all the coordination work the AI should be doing.
MCP solves this with a universal adapter layer. Instead of building individual integrations, Anthropic defined a standard. Now any developer can build an MCP server for any system, and any MCP-compatible AI client (like Claude) can use it automatically.
How MCP Works
How MCP works as a layer between tools and agents.
MCP uses a client-server model over two transport mechanisms:
stdio: The MCP server runs as a local subprocess on your machine. Claude Code spawns it, communicates via standard input/output. This is the most common setup.
HTTP/SSE: The MCP server runs as a network service. Claude connects over HTTP with Server-Sent Events for streaming. Better for remote or shared servers.
The communication protocol underneath is JSON-RPC 2.0 — a lightweight, well-understood standard for calling methods and getting results.
Each MCP server exposes one or more of three primitives:
Tools: Functions Claude can call. Example: read_file(path), create_issue(title, body), run_query(sql). Claude decides when to call them based on context.
Resources: Data sources Claude can read. Example: the contents of a directory, a database schema, a project’s README. Resources are passive — they don’t take actions, they expose information.
Prompts: Reusable prompt templates that servers can provide to standardize how Claude interacts with them.
When Claude sees a task that could benefit from an available tool, it calls the tool, receives the result, and incorporates it into the response. This happens automatically — you don’t have to tell Claude when to use MCP. Claude decides based on what the server exposes.
MCP in Claude Code vs Claude Desktop
Both Claude Code (the CLI tool) and Claude Desktop support MCP, but they configure servers differently.
Claude Code
Claude Code has built-in MCP management via the claude mcp command family:
claude mcp add my-server -- npx -y @modelcontextprotocol/server-filesystem /path/to/directory
claude mcp list
claude mcp remove my-server
Servers added with claude mcp add are stored in your Claude Code config (~/.claude.json or the project-level .claude/settings.json). Project-level configs let you commit MCP server setups to source control so the whole team gets them automatically.
Claude Code also ships with a set of built-in tools that behave like MCP servers but don’t require separate installation: file read/write/edit, bash execution, glob search, grep, web fetch, and the agent spawning tools you’re reading about in this article.
Claude Desktop
Claude Desktop reads MCP server configuration from a JSON file:
Restart Claude Desktop after editing the config. Each server you add appears in the Claude Desktop interface with a hammer icon, and Claude can access its tools in any conversation.
The Most Useful MCP Servers
Anthropic maintains a reference set of official MCP servers. These are the ones worth knowing:
Server
What It Does
Package
Filesystem
Read/write files and directories on your local machine
@modelcontextprotocol/server-filesystem
GitHub
Read repos, create issues, open PRs, push code
@modelcontextprotocol/server-github
PostgreSQL
Read-only SQL queries against a Postgres database
@modelcontextprotocol/server-postgres
SQLite
Read/write a local SQLite database file
@modelcontextprotocol/server-sqlite
Brave Search
Live web search via Brave’s Search API
@modelcontextprotocol/server-brave-search
Puppeteer
Headless browser — screenshot pages, scrape, fill forms
Persistent key-value knowledge graph across conversations
@modelcontextprotocol/server-memory
Beyond the official set, hundreds of community-built MCP servers cover everything from Notion and Linear to AWS and Docker. The MCP ecosystem grew faster than almost anyone expected after the November 2024 launch.
Installing Your First MCP Server
The fastest path is Claude Code with the filesystem server. This gives Claude read/write access to a directory you specify — useful for any project work.
Prerequisites: Node.js installed (the server runs via npx).
In your terminal:
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/Documents/projects
That’s it. Open a Claude Code session. Claude can now list, read, write, and search files inside ~/Documents/projects. Try: “List all Python files in this directory and summarize what each one does.”
For Claude Desktop, edit the claude_desktop_config.json file directly (see format above), then restart the app.
What MCP Cannot Do
What MCP cannot do — permissions and review still matter.
A few things worth understanding before you build on MCP:
MCP servers don’t persist between conversations. Each Claude session starts fresh. If you need state persistence, you need a server with its own storage layer (the Memory server handles this specifically).
MCP doesn’t bypass Claude’s safety guidelines. Claude still decides whether to execute a tool call based on safety and ethics reasoning. Connecting a filesystem server doesn’t give Claude unlimited license to delete files — Claude will still confirm before destructive operations.
Subprocess MCP servers are local. The stdio transport runs servers on your machine. This means they only work when you’re running Claude Code locally. For remote or team-shared access, you need HTTP/SSE transport with a hosted server.
Security Considerations
MCP servers have real permissions. The filesystem server can read and write files. The GitHub server can push code to your repos. The Postgres server can run SQL queries.
Apply the principle of least privilege:
Scope filesystem servers to the directory you actually need, not /
Use read-only database credentials where you don’t need writes
Create GitHub tokens with minimum required scope (e.g., repo for private repos, not org-level admin)
Never commit environment variables containing API keys to source control, even in .claude/settings.json — use env var references instead
MCP servers run with the permissions of the user running Claude. If something goes wrong with a tool call, it can have real consequences. The upside: everything runs locally and through your own credentials — there’s no MCP cloud intermediary with access to your data.
MCP and Claude Code’s Agentic Workflows
The full power of MCP shows up in Claude Code’s multi-step agentic mode. When Claude Code has access to git, a filesystem, a browser, and a search tool simultaneously, it can execute workflows like:
Search the web for a library’s current API (Brave Search)
Read your existing code to understand the integration point (filesystem)
Write the updated code (filesystem write)
Run tests (bash)
Create a PR (GitHub)
Each of these steps would require a separate tool in a traditional automation stack. With MCP, Claude orchestrates all of them within a single session, using whatever servers are available.
This is what makes MCP the infrastructure layer for agentic AI — not a feature, but the foundation that makes complex AI-driven workflows possible.
What does MCP stand for?
Model Context Protocol. It’s an open standard for connecting AI models to external tools, data sources, and services through a standard client-server interface.
Who created MCP?
Anthropic created MCP and released it as an open standard in November 2024. The specification and reference servers are open-source on GitHub. While Claude is the primary client, other AI systems can implement MCP clients too.
Do I need to install MCP to use Claude?
No. Claude works without any MCP servers. MCP is an extension layer — you add servers when you want Claude to access specific external systems. Claude Code also ships with a set of built-in tools (file operations, bash, web fetch) that don’t require MCP installation.
Is MCP available on Claude.ai (the web app)?
MCP server support is primarily in Claude Desktop and Claude Code. The Claude.ai web interface has its own tool integrations (web search, document analysis) but doesn’t support custom MCP servers in the same way.
What’s the difference between MCP tools and Claude’s native tools in Claude Code?
Claude Code’s native tools (Read, Write, Bash, Glob, Grep, WebFetch, Agent) are built into the application and don’t require a separate server process. MCP servers are external — they run as subprocesses or network services that Claude Code connects to. Both expose tools that Claude can call; the mechanism for loading them is different.
How do I build my own MCP server?
Anthropic provides official SDKs for building MCP servers in TypeScript, Python, Go, and other languages. The TypeScript SDK (@modelcontextprotocol/sdk) is the most mature. Start with Anthropic’s MCP documentation and the reference server implementations on GitHub as templates.
Last verified: June 12, 2026. MCP specification and server ecosystem evolve quickly — check the official Anthropic MCP documentation for the current spec.
The Signal is a daily AI intelligence briefing from Tygart Media — field notes from someone who builds with these tools 12 hours a day, not someone who reads press releases about them. Each edition distills the day’s most consequential AI and search developments into what they actually mean for agencies, small business operators, and builders shipping real infrastructure.
June 10, 2026: The Day the Lanes Forked
The day the lanes forked — product vs infrastructure.
Today was the kind of day where you can feel the road forking under your tires. Not because one thing happened — because eight things happened simultaneously, and if you squint at the pattern, they all point the same direction: AI just stopped being a product category and started being infrastructure. The plumbing layer. The thing you build on top of, not the thing you buy.
I’ve been building with Claude since the Haiku days. I run it 12 hours a day across 20+ WordPress sites, a five-site knowledge cluster on Google Cloud, and a custom schema engine I shipped yesterday. When the landscape shifts, I don’t read about it on TechCrunch — I feel it in the tooling. And today, the tooling lurched forward in a way that matters.
Here’s the daily signal.
Claude Fable 5: Mythos-Class AI Goes Public
Anthropic launched Claude Fable 5 yesterday — the first publicly available Mythos-class model, a tier above Opus. Pricing is $10 per million input tokens and $50 per million output tokens. It’s the most capable model Anthropic has ever released to the general public, state-of-the-art on nearly every benchmark, and it comes with a fascinating constraint: queries on certain topics automatically route to Opus 4.8 instead, triggering in less than 5% of sessions. Anthropic is essentially saying: here’s the most powerful thing we’ve ever built, and we’ve installed guard rails at the edge cases where power becomes risk.
For agencies and small business operators, the practical read is this: Fable 5 is included on Pro, Max, Team, and Enterprise plans through June 22 at no extra cost. After that, it comes off the subscription tiers. If you’re building workflows that depend on Mythos-class reasoning, you have 12 days to test whether the capability justifies the API cost — or whether Opus and Sonnet handle your actual use cases just fine.
The real signal isn’t the model itself. It’s that Anthropic also doubled Cowork limits at no charge and shipped Claude Managed Agents in public beta. They’re not just selling you a smarter model — they’re selling you an operating system for delegating work to AI. That’s a fundamentally different product than a chatbot.
Meanwhile, I Was Building the Infrastructure Layer — Not Reading About It
Infrastructure layer work while the headlines chase models.
While the tech press was writing headlines about Fable 5, I was elbow-deep in the kind of work that actually turns these models into business value. Yesterday, across a 14-hour session, my team — which at this point is me and a fleet of Claude instances — shipped three things that matter more to my clients than any benchmark score:
1. bcesg-knowledge-api v1.5.0 — a custom WordPress plugin I built and deployed across BCESG.org that outputs a JSON-LD @graph array containing Article, FAQPage, Organization, WebPage, BreadcrumbList, Person (author), and speakable schema — all generated from 13 custom meta fields. This isn’t a schema plugin you install from the WordPress directory. It’s a purpose-built schema engine designed for one thing: making every page on the site machine-readable enough that AI systems cite it as an authoritative source. That’s Generative Engine Optimization at the infrastructure level, not the content level.
2. WordPress 7.0 across the entire knowledge cluster. All five sites — bcesg.org, restorationintel.com, riskcoveragehub.com, continuityhub.org, and healthcarefacilityhub.org — upgraded from WP 6.9.4 to 7.0. Why does this matter? Because WordPress 7.0 ships the Abilities API: agent-to-agent communication endpoints. That means my Claude-powered content pipelines can now negotiate directly with WordPress about what they’re allowed to do, without me acting as the middleware. The cluster just became AI-native infrastructure.
3. The stack around it. RankMath SEO installed with the schema module deliberately disabled — because the custom plugin handles schema, and two schema systems fighting each other is worse than none at all. IndexNow for instant search engine notification on every publish and update. Microsoft Clarity for behavioral analytics so I can see what humans actually do when they land on AI-optimized content.
And here’s the detail that would have been impossible to explain six months ago: the peer review on the bcesg-knowledge-api plugin was done by Claude Fable 5 reviewing the code that Claude Opus wrote. AI reviewing AI’s code. In production. On a live WordPress cluster. That’s not a demo — that’s Tuesday.
OpenAI’s S-1 and the $965 Billion Elephant
What the split means for operators building agents.
OpenAI filed a confidential S-1 with the SEC. They’re going public. Meanwhile, Anthropic hit a $965 billion valuation. These two facts, side by side, tell you everything about where the money thinks AI is going: it’s going to be the most valuable infrastructure layer since cloud computing, and the market is pricing it that way before most businesses have figured out how to use it.
For small business owners and agency operators, this isn’t abstract finance news. It means the tools you’re using today — Claude, GPT, Gemini — are backed by companies with enough capital to keep shipping improvements for years. The platform risk isn’t that these companies disappear. The platform risk is that you don’t build on them fast enough and your competitors do.
AI Passed the Turing Test. Now What?
A UC San Diego study published in PNAS confirmed that OpenAI’s GPT-4.5 and Meta’s Llama-3.1-405B both passed a standard three-party Turing test — with GPT-4.5 being identified as human 73% of the time when given a persona prompt, significantly more often than actual human participants. This has been treated as a milestone headline, and it is one, but the practical implication is more subtle than “AI can fool humans.”
What it actually means: the content quality bar just moved permanently. If AI can produce text that’s indistinguishable from a human expert, then the only content that wins is content with something AI can’t fake — lived experience, proprietary data, operational specifics, the kind of “I shipped this yesterday and here’s what happened” detail that no model can generate from training data. This is why I write The Signal as field notes, not as analysis. Analysis can be generated. Field notes from the arena cannot.
Chrome WebMCP: The Browser Becomes an AI Endpoint
Google shipped the Chrome WebMCP API in Origin Trial for Chrome 149 through 156. The Model Context Protocol — the same protocol that lets Claude connect to external tools, databases, and APIs — is now a browser-native capability. Web applications can expose structured tool interfaces that AI models call directly.
This is a bigger deal than it sounds. Right now, when Claude interacts with a web application, it’s either through a dedicated MCP server or through browser automation (clicking pixels on a screen like a human would). WebMCP means any web app can define a structured API surface that AI agents consume natively. For agencies building client tools, this is the moment your internal dashboards and client portals become AI-ready without a full backend rewrite.
If you’re running WordPress sites — and 43% of the web is — this has direct implications for how AI agents interact with your content management layer. The gap between “website” and “AI-accessible knowledge base” just narrowed dramatically.
The GPU Infrastructure Play: xAI Becomes an AI REIT
Elon Musk’s xAI, home of Grok, is increasingly looking less like an AI model company and more like a GPU real estate investment trust. They’re partnering with both Anthropic and Google to provide compute infrastructure. This is the clearest sign yet that the AI industry is stratifying into two distinct layers: model companies (who build the brains) and infrastructure companies (who build the data centers those brains run in).
For builders, this is good news. More compute supply means more pricing competition means lower API costs over time. The $10/$50 per million tokens for Fable 5 today will look expensive in 18 months.
The Security Layer Nobody’s Talking About
HashiCorp announced Boundary for agentic AI — access security specifically designed for AI agents that need to authenticate across multiple systems. And MemPalace shipped a local-first AI memory system with 96.6% recall accuracy and 29 MCP tools for Claude Code.
These aren’t headline products. They’re infrastructure connective tissue. When AI agents can securely authenticate across your entire tool stack (HashiCorp Boundary) and maintain persistent memory across sessions (MemPalace), you stop using AI for one-off tasks and start using it as a persistent operational layer. That’s the transition my agency is making right now — from “Claude helps me write articles” to “Claude runs the content pipeline while I focus on strategy.”
What This All Means: The Two-Lane Highway
Here’s the pattern I see when I lay these signals side by side:
Lane 1: The AI product lane. This is where most people are. They use ChatGPT to draft emails. They ask Claude to summarize documents. They treat AI as a productivity tool, like a faster Google or a better autocomplete. This lane is getting crowded, commoditized, and — with the Turing test results — increasingly indistinguishable from one provider to the next.
Lane 2: The AI infrastructure lane. This is where the alpha is. Custom schema engines. Agent-to-agent communication via the WordPress Abilities API. Browser-native MCP endpoints. Persistent AI memory. Secure multi-system authentication for autonomous agents. This lane is where you stop using AI and start building on AI — where it becomes the foundation layer of your operations, not an add-on.
The gap between these two lanes is widening every day. Today’s eight signals all point the same direction: toward a world where the businesses that win aren’t the ones that use AI tools the best, but the ones that build AI infrastructure the fastest.
I’m building in Lane 2. Yesterday it was a custom schema engine and a WordPress 7.0 cluster upgrade. Today it’s field-testing Fable 5 as a code reviewer. Tomorrow it’ll be whatever the next signal demands.
The question isn’t whether AI is going to transform your industry. That’s settled. The question is whether you’re in the arena building the infrastructure, or on the sidelines reading about people who are.
What is Claude Fable 5 and how does it differ from Claude Opus?
Claude Fable 5 is Anthropic’s first publicly available Mythos-class AI model, released June 9, 2026. It sits a tier above Claude Opus in capability, priced at $10 per million input tokens and $50 per million output tokens. Fable 5 is state-of-the-art on nearly all tested benchmarks and includes built-in safeguards that route certain queries to Opus 4.8, triggering in less than 5% of sessions. It’s available free on subscription plans through June 22, 2026.
What is the Chrome WebMCP API and why does it matter for businesses?
The Chrome WebMCP API, now in Origin Trial for Chrome versions 149 through 156, brings the Model Context Protocol natively into the browser. This allows web applications to expose structured tool interfaces that AI models can call directly — eliminating the need for dedicated backend integrations or browser automation. For businesses running web-based tools, dashboards, or WordPress sites, this means your existing applications can become AI-accessible without a full rebuild.
What is the WordPress 7.0 Abilities API?
The WordPress 7.0 Abilities API provides agent-to-agent communication endpoints, allowing AI-powered systems to negotiate capabilities and permissions directly with a WordPress installation. This transforms WordPress from a content management system into AI-native infrastructure where automated pipelines can query what operations they’re authorized to perform without human middleware.
What does AI passing the Turing test mean for content creators?
A UC San Diego study published in PNAS found that OpenAI’s GPT-4.5 and Meta’s Llama-3.1-405B both passed a standard three-party Turing test in 2026 — GPT-4.5 was identified as human 73% of the time with persona prompting. For content creators, this permanently raises the quality bar — the only content that wins is content with elements AI cannot fake: lived experience, proprietary data, operational specifics, and first-person field reports that no model can generate from training data alone.
What is Generative Engine Optimization (GEO) and how does it work?
Generative Engine Optimization is the practice of structuring web content so AI systems — including ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — cite, reference, and recommend it. GEO involves entity enrichment, structured data (JSON-LD schema), authoritative citations, and machine-readable formatting. Unlike traditional SEO which targets search engine crawlers, GEO targets the large language models that increasingly mediate how users discover information.
How should small businesses approach AI infrastructure in 2026?
Start by moving from Lane 1 (using AI as a productivity tool) to Lane 2 (building AI into your operational infrastructure). Practical first steps include implementing structured data and schema markup on your website, setting up AI-optimized content pipelines, ensuring your site is crawlable by AI systems via protocols like LLMS.txt, and testing agentic workflows where AI handles multi-step operational tasks autonomously rather than single-prompt interactions.
What is a custom schema engine and why build one instead of using plugins?
A custom schema engine is a purpose-built WordPress plugin that generates structured data (JSON-LD) tailored to specific business objectives — in this case, AI citation optimization. Unlike off-the-shelf schema plugins that generate generic markup, a custom engine outputs precisely the entity relationships, author signals, and speakable content markers that AI systems use when deciding which sources to cite. The bcesg-knowledge-api plugin generates a seven-type @graph array from 13 custom meta fields, providing a level of control that no general-purpose plugin offers.
What is the significance of AI reviewing AI-written code in production?
When Claude Fable 5 peer-reviewed code written by Claude Opus for a production WordPress plugin, it demonstrated a mature AI development workflow where different model tiers serve different roles — one for generation, another for quality assurance. This mirrors human development practices (developer writes, senior reviews) but at machine speed and cost. It’s a practical example of how AI agent collaboration is already operational in real business infrastructure, not just research demos.
The Signal is published daily on Tygart Media by Will Tygart. Each edition distills the day’s most consequential AI, search, and technology developments into actionable intelligence for agencies, small business operators, and builders shipping real AI infrastructure.
Everything you need to know about Anthropic’s new frontier tier — pricing, context window, model comparisons, and how to route the right work to the right model.
Updated June 2026·~14 min read·Includes interactive calculators
What Is Claude Fable 5?
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s new frontier model tier — positioned above Opus in the lineup and designed for tasks where raw capability, extended reasoning depth, and massive context handling matter more than cost. Where Opus 4.8 set the bar for complex multi-step reasoning, Fable 5 raises it with a 1-million-token context window, enhanced agentic autonomy, and improved performance on long-horizon software engineering, research synthesis, and cross-domain analysis tasks.
The “Fable” naming signals a new generation of model architecture rather than an incremental update. Anthropic positions it as the model you reach for when a task exceeds what Opus can do reliably — not as a replacement for Opus, Sonnet, or Haiku in their respective cost tiers.
Quick Facts — Claude Fable 5
Context Window
1M
tokens (~750K words)
Max Output
32K
tokens per response
Input Price
$10
per million tokens
Output Price
$50
per million tokens
Cache Write
$12.50
per million tokens
Cache Read
$1.00
per million tokens
Key positioning: Fable 5 is the model for tasks where Opus 4.8 produces reliable but imperfect results — long codebase audits, full-document analysis, complex multi-agent orchestration, and strategic synthesis across large corpora. For most production workflows, Sonnet remains the value pick.
Full Model Lineup Comparison
Full model lineup comparison.
Here’s how the complete 2026 Claude lineup stacks up across every dimension that matters for production usage:
Model
Input $/M
Output $/M
Context
Max Out
Vision
Tool Use
Extended Think
Best For
◆ Fable 5
$10
$50
1M
32K
✓
✓
✓ Deep
Max-capability tasks, 1M+ context
◆ Opus 4.8
$5
$25
200K
32K
✓
✓
✓
Complex reasoning, agentic workflows
◆ Sonnet 4.6
$3
$15
200K
16K
✓
✓
✓
Production apps, content at scale
◆ Haiku 4.5
$1
$5
200K
8K
✓
✓
—
High-volume, latency-sensitive tasks
Prices are per million tokens. Cache read is 90% cheaper than standard input across all models. Batch API provides an additional 50% discount on both input and output.
Capability Matrix — What Each Model Can Do
Capability
Fable 5
Opus 4.8
Sonnet 4.6
Haiku 4.5
Full codebase analysis (>500K tokens)
✓ Native
⚠ Chunked
✗
✗
Extended thinking / chain-of-thought
✓ Deep
✓
✓
✗
Multi-step agentic orchestration
✓ Best
✓
Good
Limited
Computer use
✓
✓
✓
✗
MCP tool integration
✓
✓
✓
✓
Prompt caching
✓
✓
✓
✓
Batch API (50% discount)
✓
✓
✓
✓
PDF / document analysis
✓
✓
✓
Limited
Real-time streaming
✓
✓
✓
✓
Structured JSON output
✓
✓
✓
✓
Interactive Cost Calculator
Estimate your monthly API spend across the full model lineup. Enter your token volumes below — the calculator models prompt caching and Batch API discounts automatically.
Token Cost Calculator
Estimated Monthly Cost
$0.00
Which Claude Model Should You Use?
Which Claude model should you use?
Answer three questions to get a model recommendation tailored to your use case.
Model Picker — 3 Questions
1. How large is your context? (document/codebase size)
Under 50K tokens
50K–200K tokens
200K–1M tokens
2. How complex is the task?
Simple / structured (classify, extract, format)
Moderate (draft, summarize, QA)
Complex (reason, plan, code, orchestrate)
3. How cost-sensitive is this workload?
Very — high volume, every cent counts
Moderate — quality matters more than cost
Not sensitive — quality and capability first
How We Actually Use Each Model
These are real production workflows mapped to the right tier — built from running Claude in content operations, publishing automation, and knowledge management at scale. No hypotheticals.
Haiku 4.5 — High Volume
Daily SEO Refresh Pipeline
25-post-per-day SEO metadata refresh
Article classification and tag assignment
Structured data extraction from web pages
Keyword density checks across large post archives
Link validation and redirect flagging
Sonnet 4.6 — Production Default
Editorial Content at Scale
Desk article writing (1,200–2,500 words)
Content brief execution from keyword clusters
FAQ and schema markup generation
Cross-site content adaptation and localization
Monthly client update drafts and summaries
Opus 4.8 — Complex Reasoning
Workers & Deep Refreshes
Agentic Notion Workers (multi-step pipelines)
Deep content refresh with competitive gap analysis
Multi-database synthesis and reporting
Strategy documents requiring extended reasoning
Code generation for automation scripts
Fable 5 — Max Capability
Portfolio Audits & Strategy
Full-site content audits (500+ posts in single context)
Cross-domain strategy synthesis across large corpora
Complex multi-agent orchestration at the flagship tier
Long-horizon planning requiring deep reasoning depth
Codebase-wide analysis and architecture review
Routing principle: The right model is the cheapest one that reliably completes the task. Haiku handles volume. Sonnet handles production. Opus handles complexity. Fable 5 handles scale + complexity together — specifically the cases where you’d need Opus and more context than Opus can hold.
The Economics: Routed vs All-Fable
Smart model routing is where API costs get controlled. Here’s a real-world comparison of a mixed content-and-automation workload at scale — routed vs running everything on Fable 5.
Workload
Monthly Volume
Routed Model
Routed Cost
All-Fable 5 Cost
Savings
SEO metadata batch refresh
750 posts/mo
Haiku 4.5 + Batch
$1.20
$18.75
93% less
Article drafting
90 articles/mo
Sonnet 4.6
$8.10
$67.50
88% less
Agentic worker runs
200 runs/mo
Opus 4.8
$22.50
$45.00
50% less
Full-site portfolio audits
4 audits/mo
Fable 5
$24.00
$24.00
—
Total
—
Routed
$55.80
$155.25
64% less
Stacking Discounts: Caching + Batch API
Two discount mechanisms compound independently:
Prompt caching: Cache your system prompt and shared context once. Subsequent requests pay ~10% of the input price for cache reads. On Fable 5, that’s $1.00/M instead of $10.00/M on cached tokens — a 90% reduction on your largest cost lever.
Batch API: Submit requests asynchronously (results within 24 hours) for a flat 50% discount on both input and output. Works on all four models. Best for non-real-time workloads like overnight refreshes, audits, or bulk classification.
Stacked: Caching + Batch combined can bring effective Fable 5 input cost from $10/M to ~$0.50/M on cached tokens — making it economically viable for high-volume tasks that previously only fit Haiku’s budget.
Claude Fable 5 sits above Opus 4.8 in the lineup. The primary difference is context window size — Fable 5 offers 1 million tokens vs Opus 4.8’s 200K — and the depth of extended reasoning for highly complex tasks. Opus 4.8 remains the right choice for most complex agentic workflows at half the cost. Fable 5 is best when you need both maximum context and maximum reasoning depth simultaneously, or when a task has routinely hit the limits of what Opus can do reliably.
Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — 2× Opus 4.8 ($5/$25), 3.3× Sonnet 4.6 ($3/$15), and 10× Haiku 4.5 ($1/$5). Prompt caching drops the effective input cost to $1.00/M on cache reads, and the Batch API adds a 50% discount on all tokens for non-real-time workloads. Stacking both discounts makes Fable 5 viable for higher-volume use cases than the base price suggests.
Claude Fable 5 has a 1-million-token context window — approximately 750,000 words or roughly 1,500 pages of text. This is 5× the context window of Opus 4.8, Sonnet 4.6, and Haiku 4.5 (all 200K). In practice, a 1M context window lets you pass entire codebases, long research corpora, or full document archives in a single API call without chunking or retrieval workarounds. For more on context window mechanics, see our full context window guide.
Yes. Claude Fable 5 is available through the Anthropic API using the model ID claude-fable-5-20260101 (check the Anthropic documentation for the exact identifier). It supports the same API surface as the rest of the Claude family — streaming, tool use, prompt caching, vision, the Batch API, and MCP server integration. Access requires an Anthropic API account with Fable 5 enabled on your usage tier.
Fable 5 is available in Claude.ai on the Pro and Team plans. The interface lets you select it from the model picker when starting a conversation. Like Opus, Fable 5 in claude.ai has message limits that reset on a rolling window — it’s designed for individual complex tasks rather than high-volume API workloads. For production-scale usage, the API with the Batch API discount is the more economical path.
Yes — and Fable 5’s extended thinking is the deepest in the lineup. Where Opus 4.8 supports extended thinking for complex reasoning tasks, Fable 5 uses a more capable reasoning engine designed for tasks that require longer chains of inference, more working memory, and more reliable self-correction. It’s particularly effective on math, logic, long-horizon planning, and tasks where the model needs to hold and manipulate many interdependent concepts simultaneously.
For most content production — articles, blog posts, social copy, summaries, SEO content — Sonnet 4.6 is the right call. It produces high-quality output at 3.3× less cost than Fable 5, and for typical content lengths (500–3,000 words), the quality difference is minimal. Reach for Fable 5 when you need to synthesize across a very large corpus (e.g., auditing 200+ posts simultaneously), when the content requires deep domain reasoning that benefits from extended thinking, or when the task involves both large-context ingestion and complex output generation in a single pass.
Three levers in order of impact: (1) Model routing — only use Fable 5 when the task genuinely requires it; route everything else to Opus, Sonnet, or Haiku based on complexity and volume. (2) Prompt caching — structure your system prompt and shared context so it can be cached; cache reads cost $1.00/M instead of $10.00/M on Fable 5. (3) Batch API — submit non-real-time workloads via the Batch API for a flat 50% discount. Stacking all three — routing + caching + batch — can reduce effective per-task costs by 85–95% compared to unoptimized Fable 5 calls.
More Claude Guides from Tygart Media
We run Claude in production every day. These are the guides that come from using it, not just writing about it.
