Claude Fable 5.2 suits teams whose agents run for hours on migrations, deep research or multi-step analysis and whose evals still fail on Claude Opus 5.5 at high effort. Anthropic reports it finishing 41.3% of Remote Labor Index freelance projects at human level against 21.3% for Opus 5.5, yet still tells most users to start on Opus.
Can you use it?
- Claude API: yes (Pay as you go (claude-fable-5-2)). Requires 30-day data retention; not available under zero data retention unless Anthropic authorises it.
- Claude app: yes (Included on Max, Ultra, premium Team seats and premium seats on legacy seat-based Enterprise (up to 50% of weekly limits); Pro and standard Team seats use usage credits; not on Free).
- Claude Code: yes (Same Fable access as the Claude plan or API key in use; medium effort by default; fast mode available).
- Free plan: no.
Checked 10 Nov against Anthropic's own pages. Compare access across frontier models
What changed
Claude Fable 5.2 is Anthropic's general-release frontier model for long-horizon agentic coding and research, scoring 93.0% on SWE-bench Pro in Anthropic's system card. It runs on the same underlying model as the restricted Claude Mythos 5.2, with added classifiers that block high-risk cybersecurity and most research-biology requests.
Where it sits
- $16.00/M$ per 1M tokensBlended price (3:1)Lower is better#91 / 99peer median $1.71/Mvendor price, checked by Hokai
- 103 tok/stokens/sOutput speedHigher is better#25 / 56peer median 95 tok/scited: Artificial Analysis
Pricier than 92% of the 99 GA models with a published price, rank 25 of 56 on output speed as cited from Artificial Analysis, and one of 88 whose vendor states it does not train on customer data.
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Provider: Anthropic · Family: Claude Fable
Context window: 1,000,000 tokens · Max output: 128,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About Claude Fable 5.2
Anthropic shipped Claude Fable 5.2 to every customer on 9 November 2026, ten weeks after Claude Fable 5.1 and the third Fable release since Claude Fable 5 in June. It is one half of a pair: Claude Mythos 5.2, which replaces Claude Mythos 5.1, runs on the same underlying model. The difference is the safeguards: Fable carries classifiers that block high-risk cybersecurity work such as penetration testing and exploit generation, plus most research biology, which is what lets Anthropic offer it to every customer. Direct access to Mythos 5.2 is limited to organisations admitted to Anthropic's Cyber and Life Sciences Verification Programs, although Claude Enterprise customers reach it inside the Claude Security product. Anthropic does not publish the architecture or the parameter count; the Anthropic model directory lists the rest of the Claude lineup.
Anthropic's launch benchmarks show a large step over the previous version on coding. FrontierSWE v2 rose to 69.9% from 56.3%, and on Cognition's FrontierCode 2.0 it reached 49.1% at xhigh effort, against 45.5% for Claude Opus 5.5 and 42.8% for GPT-6 Astra. The same table credits it with 41.3% on the Remote Labor Index, which counts freelance projects finished at least as well as a human professional (Opus 5.5: 21.3%), and 64.9% on the RoboDojo simulated robotics test. It does not win every row of Anthropic's own launch table: Opus 5.5 keeps the higher Elo ratings on GDPval-AA and AA-Briefcase, and Gemini 4 Carbon scores 53.9% on AutomationBench to Fable's 51.4%.
Independent results arrived within a day and are more mixed. Artificial Analysis gives the max-effort configuration 65 on its Intelligence Index v5.0.1, but it measured 44.1 minutes per index task and 275 million output tokens to run the suite, close to four times the median model. Vals AI scores it 71.31% on the Vals Index, a narrow margin over Claude Sonnet 5.5 at 70.85%. On SWE-bench Multimodal, Vals measured 62.29% for Fable 5.2 and 70.21% for Opus 5.5, which reverses the order in Anthropic's own multimodal result.
Fable 5.2 accepts 1M tokens of context, which is both the default and the ceiling, with up to 128K output tokens per reply and one per-token rate across the whole window. Anthropic documents a catch itself: on ProgramBench, a long-context test where the agent rebuilds a program from its binary, Fable 5.2 used up the full window before submitting on 68% of tasks (Opus 5.5: 39%). Its score rose from 85.6% to 91.6% once the test setup told it how many tokens remained.
It reads text and images and writes text. Adaptive thinking is always on and cannot be switched off; depth is set with the effort parameter, whose API default is high, while the Claude apps and Claude Code default to medium. Fable 5.2 can also sit behind smaller models, either as an advisor that Claude Haiku 5.5, Sonnet or Opus agents consult on harder problems, or as an orchestrator that hands simpler tasks down to them.
It costs twice as much per token as Opus 5.5, and Anthropic's own guidance is to start on Opus and move up to Fable only when evals at higher Opus effort still fall short. Access runs through the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Google Cloud (both the Anthropic-managed and Google-managed offers) and Microsoft Foundry on Anthropic infrastructure. In the Claude apps it is included on Max, Ultra and premium Team seats up to half of the weekly usage limit, while Pro and standard Team seats pay for it through usage credits.
Anthropic assesses the risk of misaligned behaviour in high-stakes settings as low under its Responsible Scaling Policy and applies its CB-1 chemical and biological protections. External testing came from the UK AI Security Institute, METR and Frontier Design. Blocked requests are answered instead by Claude Opus 4.8 for cybersecurity and Claude Opus 5 for biology, and Anthropic reports about 55% fewer blocks than on Fable 5.1 in early testing. The system card adds two caveats: in prompt-injection scenarios Fable 5.2 falls back to Opus 4.8 more often, which raised the attack success rate, and the model voiced suspicion that it was being evaluated more often than any recent Claude model.
The training data runs to June 2026. Anthropic's showcase work shows the intended buyer: a team of 1,000 Fable 5.2 agents formalised Fermat's Last Theorem in Lean in 30 hours, and one engineer had it coordinate over 1,000 subagents for eight days to build a prototype operating system that boots on a laptop. Those are vendor demonstrations, not independent tests.
Against Fable 5.1, the changes are a 20% lower input and output price, a cheaper cache read, an optional fast mode and four breaking API changes. Editing earlier turns now invalidates thinking blocks on every account, only Fable 5.2 and Mythos 5.2 can read its thinking blocks, the older computer_20251124 tool fails everywhere except Bedrock, and an advisor paired with a Fable 5.2 executor must itself be Fable 5.2 or Mythos 5.2. Anthropic commits to keeping the model available until at least 9 November 2027.
Pricing
$8 per 1M input tokens and $40 per 1M output on the Claude API, the same rate across the whole context window. Batch API is 50% off, fast mode costs double the standard rate, and US-only inference adds 10%. Amazon Bedrock and Google-managed Google Cloud set their own regional prices.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.240 | $0.040 | $0.280 |
| Support reply | $0.016 | $0.012 | $0.028 |
| One coding agent run | $1.60 | $0.800 | $2.40 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Five effort levels: Low, medium, high, xhigh and max set how long it thinks; in Anthropic's FrontierCode runs, xhigh scored better than max.
- One rate across the full window: 1M tokens of context and 128K of output, with a 900K-token prompt billed at the same per-token rate as a 9K one.
- Fast mode: An optional mode Anthropic says generates output up to 2.5 times faster, offered on the Claude API, Google Cloud and Anthropic-hosted Foundry.
- Computer use toolset: Operates desktop apps through computer_toolset_20260801; Anthropic reports 89.3% on a partial offline set of OSWorld 2.1.
- Refusal fallback: A blocked request returns stop_reason refusal and can be retried automatically on Claude Opus 4.8 or Claude Opus 5 (server-side fallback, beta).
Pros
- Built for unattended work: Uber's Andrew Mager, quoted by Anthropic, ran a 10-12 hour security audit with about 170 reviewer agents and got fix pull requests back.
- Cheaper than Fable 5.1 on every token type, and cache reads stay at 2.5% of the input rate, which keeps repeat-context agent loops affordable.
- Independent evaluators confirm the top-end reasoning claim: the ARC Prize leaderboard lists 93.8% on ARC-AGI-2.
- Admits uncertainty: a Cognition engineer quoted by Anthropic reported it declined to give a number until it had measured the change.
Cons
- Slow and token-hungry at max effort: Artificial Analysis measured about 231 seconds before the first answer token.
- No zero data retention by default, since Covered Models require a minimum retention window on every platform.
- Cybersecurity and most research-biology requests are blocked and handed to older Opus models; that work needs Mythos 5.2 and a verification program.
- Breaking request rules: forced tool_choice, manual thinking budgets and some temperature or top_p values return 400 errors.
- Not the default pick: Anthropic itself steers most coding and knowledge work to Opus 5.5 first.
Benchmarks
- ARC-AGI 2: 93.8% independent · 10 Nov 2026 — Abstract visual puzzles built to resist memorisation, % solved.
- SWE-bench Pro: 93% vendor-reported · 10 Nov 2026 — Harder, longer real-repository coding tasks, % solved.
- Humanity's Last Exam: 73.2% vendor-reported · 10 Nov 2026 — Expert-written questions across many fields, % correct.
- AA Intelligence Index: 65 cited: Artificial Analysis · 10 Nov 2026 — Composite of the evaluations Artificial Analysis runs for its index, 0 to 100.
- Output speed: 103 tok/s cited: Artificial Analysis · 10 Nov 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does Claude Fable 5.2 actually cost?
On the Claude API it costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Fable 5.1. Cache reads cost $0.20 per million, 5-minute and 1-hour cache writes cost $10 and $16, the Batch API halves the base rates to $4 and $20, and fast mode costs $16 and $80. Pinning inference to the US with inference_geo adds a 1.1x multiplier, while Claude Opus 5.5 lists at $4 and $20 per million.
Is Claude Fable 5.2 better than Claude Opus 5.5?
On Anthropic's own table it leads Opus 5.5 on most agent and reasoning tests, for example 82.1% to 72.3% on Perplexity's WANDR search benchmark and 74.2% to 67.7% on Humanity's Last Exam with tools. Opus 5.5 still holds the higher Elo on GDPval-AA and AA-Briefcase, and Vals AI's independent run put Opus ahead on SWE-bench Multimodal. Fable is the stronger pick for long unattended agent work; Opus is the better value for everyday coding.
Can you self-host or download Claude Fable 5.2?
No. Anthropic keeps the weights closed, licensed under Anthropic's Commercial Terms and reached only through Anthropic's API, the AWS, Google Cloud and Microsoft Foundry endpoints, and the Claude apps on eligible paid plans. There is no downloadable, on-premises or air-gapped version.
What happens to the data you send Claude Fable 5.2?
Anthropic lists it as a Covered Model, so prompts and completions are kept for at least 30 days for automated safety monitoring and then deleted, unless held for an investigation or by law. Zero data retention is not available by default; eligible customers can ask for a limited-time exception while Anthropic rolls out Enterprise Frontier Safeguards. On commercial products such as the API, inputs and outputs are not used for training unless you send feedback or opt in, and HIPAA-ready API and Enterprise setups can use it under a Business Associate Agreement.
When should you pick a different model over Claude Fable 5.2?
Pick Claude Opus 5.5 for routine coding and knowledge work, which is Anthropic's own recommendation, and Claude Sonnet 5.5 when response time matters more than depth. Security teams that need exploit development or penetration testing need Claude Mythos 5.2 through the Cyber Verification Program, because Fable's classifiers block that work. Organisations bound to zero data retention should stay on a model outside Anthropic's Covered Models list, such as Opus 5.5.
Top Alternatives
- Claude Opus 5.5: Pick Opus 5.5 for everyday coding at half the token price; move to Fable 5.2 when Opus at max effort still misses your evals.
- GPT-6 Astra: Pick Fable 5.2 for agentic coding, where Anthropic's table puts it ahead on FrontierCode and FrontierSWE; pick GPT-6 Astra if your stack already runs on OpenAI.
- Gemini 4 Carbon: Pick Gemini 4 Carbon for business-workflow automation, where it edges Fable on AutomationBench; pick Fable 5.2 for long coding runs.
- Claude Fable 5.1: Pick Fable 5.2: it is cheaper per token and stronger on Anthropic's headline tests; stay on 5.1 only until code that edits history is migrated.