Skip to content
All insights

AI & Automation

Claude Fable 5.2 Benchmarks, API Pricing and Migration

Read Claude Fable 5.2 benchmarks from Anthropic and Epoch AI, check API prices and retention, and plan a document workflow pilot with a review rubric.

Claude Fable 5.2 is a candidate for document work that needs to reconcile several sources and explain its reasoning. Anthropic released it on November 9, 2026, at $8 per million input tokens and $40 per million output tokens on the Claude API. If your policy requires zero data retention, check eligibility before sending production documents. Anthropic documents a 30-day requirement for this model unless it expressly authorizes an exception.

For document-heavy SaaS, start with one review workflow and grade it against verified source material. The numbers below come from Anthropic and Epoch AI, and the cost scenario is illustrative.

Key takeaways

  • Anthropic lists a 1 million-token context window, a 128,000-token maximum output and always-on adaptive thinking.
  • Epoch AI’s human-graded evaluation found better performance from Fable 5.2 than GPT-6.2 Astra on its workplace tasks. Both models fell short of Epoch’s full-time employee standard.
  • Check data eligibility and request compatibility before testing with customer documents.
Applying this to your product?Plan a document workflow pilot

Claude Fable 5.2 benchmarks and what they measure

Anthropic’s launch table gives two points of comparison with Fable 5.1 and Opus 5.5. FrontierCode 2.0 tests coding agents. OSWorld 2.1 measures computer-use tasks and the public figures below cover Anthropic’s partial offline set. Read the settings with the scores. The FrontierCode results use Xhigh for the Fable versions and the highest available effort for Opus 5.5.

Anthropic’s November 9 launch results, with safeguards enabled
EvaluationFable 5.2Fable 5.1Opus 5.5
FrontierCode 2.049.1% (Xhigh)32.4% (Xhigh)45.5% (highest available effort)
OSWorld 2.1 (partial offline set)89.3%83.5%87.2%

Anthropic’s release and benchmark methodology

How Fable 5.2 performed on workplace tasks

Epoch AI evaluated Fable 5.2 and GPT-6.2 Astra on an updated set of open-ended tasks drawn from its own work. Fable averaged 82% across the tasks evaluated, versus 65% for Astra. On the data center analysis task, Fable reached 98%, compared with 88% for Opus 5.5 and 65% for Astra. Epoch graded those outputs against its own workplace standards and considered the two Claude results to be at contractor level on that task.

These percentages describe rubric scores for a small task suite with one run per model per task and human grading. Use them to shortlist Fable for document work, then measure material errors and review effort in your own pilot. A supplier review can look polished while missing the one exception that changes a purchasing decision.

Epoch AI workplace evaluation and task-level limitations

Build source-grounded answers with production RAG

Claude API pricing and migration checks

Use the model ID claude-fable-5-2 for the direct Claude API. Standard rates in USD per million tokens are $8 for uncached input, $40 for output, $10 for a five-minute cache write, $16 for a one-hour cache write and $0.20 for cache reads. Provider-specific deployments may use different IDs and terms.

The API defaults to high effort and adaptive thinking stays enabled. Check any adapter that sends sampling parameters or forces a specific tool choice. Omit temperature, top_p and top_k, and test the documented compatible tool-choice settings. Keep the retention decision in the rollout checklist, with the provider, environment and data classification recorded. Use approved synthetic or redacted packets while that review is open.

Claude API model contract, prices and retention

FROM READING TO DOING

Can a reviewer trace every material claim?

Share the types of documents and decisions your team reviews. We can help design a pilot around evidence, data handling and approval.

Budget repeated reviews of one document packet

Consider ten independent reviews that reuse the same 80,000-token prefix, each with 20,000 new input tokens and 10,000 billed output tokens. At standard rates, ten uncached calls cost $12.00. With a five-minute write on the first call and nine hits, the first call costs $1.36 and the next nine cost $0.576 each, for about $6.54 total.

This estimate assumes exact eligible prefix reuse while the cache remains available. If each review changes the shared prefix, measure the observed hit rate and update the budget. Track the actual number of accepted reviews and include the cost of corrections and reruns.

Claude prompt caching requirements and pricing

Pilot Fable on a supplier renewal review

Build a small set of synthetic supplier packets from a contract, renewal quote and statement of work. Give reviewers the expected answer and its source location. Include at least one conflict, one missing field and one exception buried in an appendix. Compare Fable with your current model at matched data access, and time the human review of each output.

  • Require a document and page or clause for every material claim.
  • Leave absent fields unresolved and send conflicting clauses to a reviewer.
  • Reconcile line items and totals, including currency and term length.
  • Require human approval before a consequential decision and log errors that would have changed it.

Set review gates for AI workflows

Frequently asked questions

Can I use Claude Fable 5.2 with zero data retention?

Anthropic’s model documentation specifies 30-day retention and says zero data retention requires express authorization. Confirm eligibility and the terms for the deployment surface you plan to use before processing sensitive inputs.

Should I switch from Opus 5.5 to Fable 5.2?

Put both on the same document packets and grade material errors, reviewer minutes and cost per approved report. Use the results to choose which tasks warrant switching.

YOUR NEXT STEP

Let’s work through your next step.

Tell us what you’re building, what you’ve tried, and where you need a hand. We’ll work with you to define a practical way forward.

Keep reading.

Choose which optional tools can run. You can reopen these settings from the footer. These choices do not subscribe you to marketing emails.

Cookie Policy