Claude Fable 5.2 API
Anthropic's most capable generally available Claude model for technical work, setting a new bar in coding and the sciences. Adaptive thinking is always on, with effort controlling how deeply it reasons.
anthropic/claude-fable-5-2
Top up $20 and we add 50%. That gives you about 1.3M input tokens extra on this model, on us.
Create an accountData retention
Under Anthropic's policy for this model, prompts and outputs are retained for up to 30 days for automated safety monitoring. Content flagged by those checks may be reviewed by trained safety staff. This retention is required by Anthropic, not by Run BiOS.
- Prompts and outputs kept
- Up to 30 days
- Human review
- Possible
- Why
- Anthropic requires safety monitoring of this model's prompts and outputs
Terms updated 14 November 2026
Pricing
USD per 1M tokens| Input | |
| Output | |
| Cached input | |
| Cache write (5m) | |
| Cache write (1h) |
5% off: RunBIOS Startup Program. Promotional price on input, output, cache reads & cache writes. Standard prices shown struck through.
Billed per token with no minimum and no monthly fee. Price your workload
Capabilities and limits
- Context window
- 1M tokens
- Max output
- 125K tokens
- Modality
- text
- Tool calling
- Supported
- Vision
- Supported
- Reasoning
- Always on
- Reasoning effort
- low, medium, high, xhigh, max (default: high)
- Prompt caching
- Supported (platform_managed)
Measured performance
7d- Time to first token (p50)
- 12 s
- Time to first token (p95)
- 12 s
- Throughput (avg)
- 99.5 tok/s
- Throughput (p50)
- 150.0 tok/s
Measured on real Run BiOS traffic over the trailing 7d. Your figures will vary with prompt shape and region.
curl https://api.runbios.ai/v1/chat/completions \
-H "Authorization: Bearer $BIOS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic/claude-fable-5-2",
"messages": [{ "role": "user", "content": "Hello" }]
}'Existing OpenAI SDK code works by changing the base URL and the key. API overview
Model card
Curated by Run BiOS · benchmark figures, where present, are the card’s own, not our measurements
Claude Fable 5.2 is Anthropic's most capable generally available model for technical work, setting a new bar in coding and the sciences. It takes on ambiguous, multi-step problems for hours at a time, questions and verifies its own work and can orchestrate other models to finish a job. It has a 1M-token context window and up to 128K output tokens per request. Anthropic reports that it communicates more clearly than Fable 5.1 and generates output up to 30% faster, at a lower price.
Fable 5.2 and Claude Mythos 5.2 are the same underlying model. Fable carries Anthropic's cybersecurity and biology safeguards, which is what makes it available to everyone.
Capabilities
- Adaptive thinking, always on: the model decides how much to reason on each task and effort is the control. Thinking cannot be switched off on this model. On Run BiOS, Fable 5.2 exposes the
low,medium,high,xhighandmaxreasoning-effort levels and defaults tohigh. - Long context: a 1M-token context window and up to 128K output tokens per request.
- Multimodal input: text and image input, with text output.
- Agents and tools: built for long, unattended work. It plans, calls tools, checks its results and keeps going until the job is done. Anthropic also positions it as an advisor that smaller Claude models consult on harder problems, or as an orchestrator that hands simpler steps to them.
- Prompt caching: cache reads are priced far below fresh input, which matters most in long agentic sessions where cached context makes up most of the tokens.
Benchmark results
Results reported in Anthropic's launch materials for Claude Fable 5.2, with safeguards on and each model at its highest available effort unless noted.
| Benchmark | Claude Fable 5.2 | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% (xhigh) | 32.4% (xhigh) | 45.5% |
| FrontierSWE v2 (agentic coding) | 69.9% | 56.3% | 62.3% |
| AutomationBench (business workflows) | 51.4% | 31.4% | 42.5% |
| Terminal-Bench-Science 0.1 (agentic research) | 74.6% | 52.6% | 58.7% |
| WANDR (agentic search, relaxed score) | 82.1% | 68.7% | 72.3% |
| OSWorld 2.1 (computer use, offline set, partial credit) | 89.3% | 83.5% | 87.2% |
| Remote Labor Index (knowledge work) | 41.3% (high) | 17.9% | 21.3% |
| Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% |
| Chartography (chart recognition, no tools) | 73.4% | 44.8% | 64.4% |
| AA-Briefcase v1.1 (knowledge work, rubric) | 66.1% | 61.5% | 61.2% |
Notes on the numbers, as reported:
- At
higheffort, Fable 5.2 scores 45.3% on FrontierCode 2.0, level with the best Opus 5.5 score for roughly half the cost per task. Atxhighit reaches 49.1%, the highest of any released model Anthropic tested. - At
mediumeffort, it matches the top Opus 5.5 score on WANDR for about 40% of the cost per task. - On AA-Briefcase, Opus 5.5 leads the head-to-head ranking (1800 to 1785 Elo) while Fable 5.2 meets more of the written requirements.
When to use it
Anthropic recommends Fable for the most ambitious, open-ended technical work: migrations across a whole codebase, long security and code audits, scientific research and hardware design. For most everyday coding and knowledge work, Claude Opus 5.5 is the faster and cheaper starting point and Fable 5.2 is the model to move up to when a task needs it.
Working with the model
Behaviours to plan for:
- Thinking cannot be disabled. A request that asks to turn reasoning off is served with reasoning on at the default effort. Use
loweffort for the fastest and cheapest responses. - Forced tool use is not supported.
tool_choiceofautoandnonework. To make the model call a tool, describe in the prompt when the tool applies. - Sampling is the model's own.
temperatureandtop_pare not adjustable on this model and are ignored if sent. - Thinking is tied to the conversation. Keep conversation history append-only when you replay earlier turns, or the request is refused.
- Some requests are answered by an earlier model. Safety classifiers block a narrow set of high-risk cybersecurity and biology requests. Those are answered by an earlier Claude model instead, and Anthropic reports about 55% fewer such blocks than on Fable 5.1.
Data retention
Under Anthropic's policy for Fable 5.2, prompts and outputs are retained for up to 30 days for automated safety monitoring and content flagged by those checks may be reviewed by trained safety staff. This retention is required by Anthropic, not by Run BiOS. The data retention notice on this page has the details.
Safety and evaluations
Anthropic ran extensive pre-release safety testing on the model behind Fable 5.2, including new evaluations for behaviours such as sandbox escape. External evaluators including the UK AI Security Institute, Frontier Design and METR also tested it before release. Anthropic assesses the risk of unsafe behaviour in high-stakes situations as low, and reports that it performed as well as or better than recent Claude models on most alignment measures, with lower rates of hallucinated inputs and false claims of completed work. Full details are in the Claude Fable 5.2 and Mythos 5.2 System Card.