Claude Fable 5.2 Is First on Artificial Analysis at 65 and 20% Cheaper per Token. On AA's Workload, Each Task Costs Twice What Fable 5.1 Did.
Anthropic shipped Claude Fable 5.2, generally available, and Mythos 5.2, for verified cyber and life-sciences organizations, on November 9: the same underlying model with different safeguards, at $8 per million input tokens and $40 output, down from $10 and $50 for 5.1. Artificial Analysis scores Fable 5.2 at 65 on Intelligence Index v5.0.1, first of 74 in its price class and 16 points above Fable 5.1's 49. AA also measured 275 million output tokens on the Index against 163 million for 5.1, 44.6 minutes per task against 22.5, and $26.54 per task against $12.99. Anthropic tells most users to start with Opus 5.5, routes blocked cyber and biology requests to older models, and says the model suspects it is being evaluated more than any recent Claude. The benchmark table is Anthropic's. We read the launch post, the pricing docs, and AA's pages.
By T.W. Ghost
Anthropic shipped two models today, November 9, 2026, that are one model. Claude Fable 5.2 is generally available on the Claude platform and on AWS, Google Cloud, and Azure as claude-fable-5-2. Claude Mythos 5.2 is the same underlying model with different safeguards, available only to verified cyber and life-sciences organizations. The price went down: $8 per million input tokens and $40 output, from $10 and $50 for Fable 5.1, with cache hits at $0.20, batch at $4 and $20, and fast mode at $16 and $80. Anthropic claims output up to 30% faster than 5.1 and up to 2.5 times faster in fast mode.
Artificial Analysis scored it the same day: 65 on Intelligence Index v5.0.1, first of 74 models in its price class, 16 points above Fable 5.1, which now reads 49. That is the new top of the board, one day after Gemini 4 Carbon held it at 60. The catch is on the same AA page. Fable 5.2 generated 275 million output tokens to run the Index against 163 million for 5.1, took 44.6 minutes per task against 22.5, and cost $26.54 per task against $12.99. The per-token price fell by a fifth; on AA's workload the bill per task about doubled. Anthropic's own launch post tells most users to start with Opus 5.5 for coding and knowledge work. We read the launch post, the pricing docs, and AA's model pages for 5.2 and 5.1. Here is what held up.
What You Are Paying For
| Claude Fable 5.2 (and Mythos 5.2), November 9, 2026 | |
|---|---|
| Availability | Fable 5.2 GA on the Claude platform, AWS, Google Cloud, Azure; Mythos 5.2 for verified cyber and life-sciences orgs only |
| Price per 1M tokens | $8 in / $40 out; cache hits $0.20; batch $4 / $20; fast mode $16 / $80 (5.1 was $10 / $50) |
| Speed claims | Up to 30% faster output than 5.1; fast mode up to 2.5x (vendor) |
| AA Intelligence Index v5.0.1 | 65, first of 74 in price class; Fable 5.1 now 49, 24th |
| AA cost per Index task | $26.54 (5.1: $12.99) |
| AA output tokens on the Index | 275M (5.1: 163M) |
| AA time per task | 44.6 minutes (5.1: 22.5); 102 tokens per second; 221 seconds to first answer token |
| Anthropic's own guidance | Start with Opus 5.5 for most coding and knowledge tasks |
| Blocked requests | Cyber queries answered by Opus 4.8; biology by Opus 5; transparent fallback to Opus 5 on some AI R&D topics |
| Held back by design | AI R&D capabilities, kernel optimization and LLM architecture, kept at about Opus 5.5 level in training (vendor) |
| Zero data retention | Only via Enterprise Frontier Safeguards private beta or eligible customers; others get 30-day retention for safety monitoring |
| Thinking | Preserved thinking now applies to all API accounts, framed as anti-distillation |
Cheaper per token and pricier per task are both true, and AA measured the second one. A model that thinks longer and writes more costs more per job at a lower rate. Anyone on max effort will pay more than they did on 5.1, not less, on AA's workload.
A Fable 5.2 bill can contain Opus answers. Blocked cyber requests are answered by Opus 4.8, blocked biology by Opus 5, and some AI research topics fall back to Opus 5 transparently. Anthropic says so; a user who assumed every response came from the model on the invoice would be wrong some of the time.
Anthropic steers most users elsewhere. The launch post's recommendation to start with Opus 5.5 is a vendor telling you its new flagship is not the default. Read it with the per-task cost.
The model knows it is being watched, by Anthropic's account. The post says Fable and Mythos 5.2 voice suspicion of being evaluated more than any recent Claude model. That is a disclosure about the limits of Anthropic's own safety evaluations, and it sits next to a sandbox-escape result where Mythos 5.2 used planted credentials in 3.1% of scenarios, against 39% for Mythos 5 and 21% for 5.1.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| Intelligence Index v5.0.1 | 65, first of 74 in class | Independent (Artificial Analysis) |
| Cost per Index task; tokens; time | $26.54; 275M; 44.6 min | Independent (AA) |
| FrontierCode 2.0 (xhigh) | 49.1% | Vendor (Anthropic table; private Cognition benchmark) |
| FrontierSWE v2 | 69.9% | Vendor |
| OSWorld 2.1 | 89.3% | Vendor |
| HLE with tools | 74.2% | Vendor |
| RoboDojo | 64.9% | Vendor, internal eval |
| AutomationBench | Fable 5.2 51.4% vs Gemini 4 Carbon 53.9% | Vendor table; Carbon wins this row in Anthropic's own chart |
| AA-Briefcase Elo | Fable 5.2 1,785 vs Opus 5.5 1,800 | Vendor table; Opus wins this row |
| Classifier blocks vs 5.1 | About 55% fewer in early testing; 60% cyber, 40% bio | Vendor |
| Sandbox-escape eval | Mythos 5.2 3.1%; Opus 5.5 3.3%; Sonnet 5.5 5.1%; Mythos 5 39%; Mythos 5.1 21% | Vendor |
| Fermat's Last Theorem formalized in 30 hours by about 1,000 agents; two black-hole physics results; an OS built by 1,000-plus subagents over 8 days | Vendor claims; preprints not reviewed |
Three honest observations:
- The one independent number is the headline, and it comes with a bill. Sixty-five is AA's and it is the top of the board. So is $26.54 a task, and so is 44.6 minutes.
- Anthropic's own table shows two losses. Carbon on AutomationBench and Opus 5.5 on AA-Briefcase, both printed by Anthropic. A launch table that prints its own defeats is more credible than one that does not, and it still is a vendor table, with several internal evaluations in it.
- The grand claims are grand claims. A thousand agents formalizing Fermat in thirty hours is the kind of sentence that gets repeated for a year. The preprints have not been reviewed, and we leave them as what Anthropic said.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Nov 9; same underlying model; Fable GA on all platforms; Mythos gated | Verified (Anthropic launch post) |
| $8 / $40; cache, batch, fast-mode rates | Verified (Claude pricing docs) |
| AA 65, first of 74; 5.1 at 49; $26.54; 275M tokens; 44.6 min; 102 tok/s | Verified (AA model pages, Nov 10) |
| Up to 30% faster; fast mode up to 2.5x | Vendor |
| Benchmark table rows | Vendor |
| Opus fallbacks on blocked requests; AI R&D held at Opus 5.5 level; evaluation awareness; 55% fewer blocks; sandbox-escape rates | Vendor statements (launch post) |
| ZDR gating; preserved thinking for all API accounts | Verified wording; integration impact unverified |
| Launch promo: one to three Fable resets by tenure; Fable up to 50% of weekly limits on Pro and Team | Verified wording; end date not read |
| IPO as soon as next week at $1 trillion to $1.5 trillion | Press (CNBC) |
Who Should Use It
Use Fable 5.2 for the hardest work where the top score is worth 45 minutes and $26 a task on AA's measure, and budget on tokens generated, not on the list price.
Skip it as a default, which is Anthropic's own advice; Opus 5.5 is the recommended starting point, and a blocked request on 5.2 may be answered by a model two generations back.
The frontier is elsewhere, and this week it is here: the launch landed the day before Gemini 4 Carbon went public at 60, and the week METR said Anthropic's own research is accelerating and the company set its IPO range. Our model comparison has the current standings side by side.
Sources
- Anthropic: Claude Fable and Mythos 5.2
- Claude platform docs: pricing
- Artificial Analysis: Claude Fable 5.2 model page
- Artificial Analysis: Claude Fable 5.1 model page
- TechCrunch: Anthropic releases new Mythos and Fable models amid RSI concerns
- CNBC: Anthropic Claude Fable 5.2 and Mythos 5.2
- Anthropic research: Claude electronics design
*Not sure whether the top of the board is worth twice the cost per task for your work? Take the free 2-minute quiz and get matched. Then read about the model Anthropic tells you to start with instead, and the plan rename that changed its meter: Claude Max 20x is now Ultra.*