Direct answer: a living AI model leaderboard rots. On the public Claude 4.6 vs GPT-5 leaderboard we replaced the promise that standings were “continuously updated by our autonomous tracker” with a September 2026 editor’s note. The table stayed. The date stayed. The page is now a snapshot of the race on 30 May 2026 — not a live scoreboard.
If an answer engine cites an undated comparison as current, it will be wrong within a release cycle. Date the page. Point prices at a desk you actually re-verify. Leave last season’s Elo alone unless you re-measured it.
What changed on the public pages
Three live URLs now say the same thing in different words:
- Claude 4.6 vs GPT-5: The 2026 Leaderboard — editor’s note that the autonomous tracker retired after the 30 May 2026 update. The LMSYS box still reads Last Updated: 2026-05-30, with Claude 4.6 Sonnet (Elo 1345), GPT-5 Early Preview (Elo 1338), and Claude 4.6 Haiku (Elo 1312). Those numbers were not rewritten this month.
- Claude 4.6 Roadmap: Guide for Technical Founders — September 2026 editor’s note at the top of the body: the roadmap is a May 2026 generation snapshot. Stop 2 now points at the leaderboard snapshot for where Claude 4.6 stood against GPT-5 in May 2026.
- Anthropic Slashes Claude 4.6 Haiku API Pricing by 40% — September 2026 outcome note. The phrase “verified Claude pricing page” goes to the live desk at /claude-ai-pricing/, not to a second rate card buried in the old post.
Titles and slugs did not move. The old URLs are the record.
Why a “living” leaderboard fails the reader
Model names, list prices, and context windows moved several times between that May snapshot and late September 2026. The live Claude desk we keep — last verified 23 September 2026 against Anthropic’s own pages — is a different document from a May Elo box. Mixing them on one URL trains both humans and models to treat last season’s table as this week’s answer.
That is an AEO problem and a GEO problem. Answer engines prefer a short, dated sentence they can lift. Generative engines prefer a page that does not contradict the next page in your cluster. A tracker you cannot staff is worse than a snapshot you labeled.
The method: snapshot the race, live-link the meter
- Put the as-of date in the first screen. “Last Updated: 2026-05-30” in the box. “Editor’s note (September 2026)” above it. Two dates, two jobs: when the numbers were taken, when you admitted they froze.
- Do not invent a new leaderboard you did not run. We did not scrape LMSYS tonight. We did not publish a September Elo column. If the measurement is not in the room, the number does not go on the page.
- Send current prices to one desk. Token rates and seats live on Claude Pricing (September 2026). The Haiku cut post now points there. A short companion on the 22 September Opus list sits at Claude Opus 5.5 Cuts Token Price 20% — $4 / $20 List. Quote the desk. Do not fork a third table.
- Keep the historical URL. Changing the slug of a comparison page that already earned links and citations throws the evidence away. Date the body. Leave the address.
- Say what the page is not. A May snapshot is not a buying guide for today’s default model. A 40% Haiku headline from an earlier cycle is not the live list. The pricing desk is the live list.
What we would not claim
- We would not claim a September 2026 LMSYS or Arena ranking. We did not pull one for this note.
- We would not claim Claude 4.6 is still the current default. The live desk names current Opus, Sonnet, Haiku, and Fable lines separately. Read that page for names and rates.
- We would not restated a full rate card here. Rates move. The desk is dated 23 September 2026 and cites claude.com/pricing and platform.claude.com.
- We would not describe internal work orders, agents, or edit packets. Those are not the public record. The public record is the three URLs above.
How to write the editor’s note so an answer engine can use it
Keep the note to one sentence that answers “is this current?” Then link the current desk.
Pattern we used: Editor’s note (Month Year): the living tracker was retired after [snapshot date] — the standings below are preserved as a snapshot of the race at that point.
That sentence is the AEO payload. It is short. It names the freeze date. It tells a model not to treat the table as live. GEO follows if the rest of the cluster agrees: founder roadmap says snapshot, pricing post says “go to the desk,” desk says last verified.
FAQ
Q: Should I delete an old model comparison?
A: No, if the URL already has links or citations. Label it a snapshot, keep the slug, and point current buyers at the live desk.
Q: When does a snapshot need a new page instead of an editor’s note?
A: When the question changed. “Where did Claude 4.6 sit vs GPT-5 in May 2026?” is a history question. “What does Claude cost this week?” is a desk question. Different questions get different URLs.
Q: Can I keep a living tracker?
A: Only if someone re-measures on a published cadence and the page shows the last run date. A tracker you do not run is a lie with a progress spinner.
Q: Where are current Claude prices?
A: tygartmedia.com/claude-ai-pricing/, last verified 23 September 2026 against Anthropic’s public pricing pages. Confirm Anthropic before you quote a customer.
Q: Does this mean the May Elo numbers were wrong?
A: It means they are dated. They remain the numbers printed on that page for 30 May 2026. They are not a September 2026 ranking.
Sources
- Live pages on tygartmedia.com, read 24 September 2026: leaderboard snapshot, founder roadmap snapshot note, Haiku pricing outcome note, Claude pricing desk (desk last verified 23 September 2026).
- Official Anthropic seats and API tables linked from the pricing desk: claude.com/pricing and platform.claude.com pricing.
Related on Tygart Media: May 2026 leaderboard snapshot · founder roadmap snapshot · Haiku pricing note · live Claude pricing desk · Opus 5.5 list-price note · the site has to be the answer.
I write pages like this so AI search cites a dated sentence instead of last season’s table — then I do the same for operators who publish. That’s what Tygart Media does.