Key details

  1. Released 9 November 2026; model ID claude-fable-5-2.
  2. Fable and Mythos 5.2 share an underlying model; Mythos requires approved access.
  3. Standard API $8 input and $40 output per million tokens; Fable 5.1 was $10/$50.
  4. Fast mode $16/$80 per million tokens with up to 2.5x output generation speed claimed.
  5. Context 1M tokens; maximum output 128K; adaptive thinking always on.
  6. Opus 5.5 remains $4/$20 per million tokens and is Anthropic's recommended default for most work.
  7. Fable 5.2 requires 30-day retention unless Anthropic expressly authorises otherwise.
  8. Anthropic commits not to retire Fable 5.2 before 9 November 2027.

What builders should take away

  1. Benchmark dollars per accepted task against Opus 5.5 and Fable 5.1, including effort and retries.
  2. Validate request parameters, tool calls, preserved thinking and computer-use integration before changing model IDs.
  3. Check data-retention and verification restrictions before routing sensitive workloads.
  4. Use fast mode only when the latency benefit justifies its doubled token rate.

What changed

On 9 November 2026 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model with different safeguards and access. Fable is generally available through the Claude API and cloud partners; Mythos requires verification for specialized cybersecurity and life-science work. Standard prices fall from Fable 5.1's $10/$50 to $8/$40 per million input/output tokens. Optional fast mode costs $16/$80 and promises up to 2.5x output-token generation speed. Both have 1M-token context and 128K maximum output, with always-on adaptive thinking.

Why it matters

Builders using high-end coding and long-running research agents can revisit routing and cost-per-completed-task assumptions. A 20% token-rate cut does not guarantee a 20% job-cost cut, especially where effort, retries and token consumption differ. Fable still costs twice as much per uncached input/output token as Opus 5.5, which Anthropic itself recommends as the starting point for most work. Restricted Mythos access, mandatory 30-day retention unless specifically authorised otherwise, and breaking API conditions mean a blind model-ID swap is unsafe.

The economics are task-dependent, not just a rate-card reduction

Fable 5.2 is $8 input/$40 output per million tokens, versus $10/$50 for 5.1 and $4/$20 for Opus 5.5. Cache reads cost $0.20 per million; five-minute writes $10 and one-hour writes $16. Batch API input/output is discounted 50%. Fast mode doubles token rates to $16/$80 for up to 2.5x faster output generation according to Anthropic; end-to-end latency may differ.

Two access policies for one underlying model

Fable is the general-use model with stronger cybersecurity and biology safeguards. Mythos has the same base model and prices but is accessible through verification programmes. Mythos is not simply an unrestricted public API model. Anthropic's claims of research and cybersecurity improvements derive mainly from its own evaluations and named early testers.

Migration requires API and compliance checks

The model identifier is claude-fable-5-2. Anthropic documents always-on adaptive thinking, incompatible forced tool-choice settings and sampling parameters, and a 30-day data-retention requirement without ordinary zero-data-retention eligibility. The migration guide also details preserved-thinking and computer-use tool compatibility. Integration owners should run representative request and tool-call tests rather than assume older payloads work unchanged.

Vendor benchmark gains need real-world validation

Anthropic reports 49.1% on FrontierCode 2.0 at Xhigh effort against 32.4% for Fable 5.1, but effort settings, task harness and costs are important. Customer quotations are useful experience reports, not independent reproducible benchmark studies. Test completion quality, retry rates and total spend on the actual workload.

What to watch next

  • Independent cost-per-task and long-running agent evaluations.
  • Fast-mode actual end-to-end latency and rate-limit behaviour.
  • Any changes to Mythos verification or retention terms.

Still unclear

  • Anthropic's benchmark and speed comparisons are primarily vendor-produced.
  • Early customer examples are not broadly reproducible measurements.
  • Task-level cost and performance depend on prompt, effort, caching, retries and deployment.

Sources

Direct reading behind this dossier.

4 sources

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment