On November 9, 2026, Anthropic released two new models: Claude Fable 5.2 and Claude Mythos 5.2. Fable 5.2 generates answers up to 30% faster than the previous version, Fable 5.1, and is 20% cheaper: $8 per million input tokens and $40 per million output tokens. According to the company, the update outperforms Claude Opus 5.5 in nearly all benchmarks and is already available to developers via API and to Claude subscribers without a separate account.

image
image

What happened

On November 9, 2026, Anthropic presented an updated generation of models for complex technical tasks: the publicly available Claude Fable 5.2 and Claude Mythos 5.2, the latter of which is issued only to verified organizations in cybersecurity and life sciences. In standard mode, Fable 5.2 costs $8 and $40 per million input and output tokens, respectively, while Fast Mode, priced at $16/$80, responds up to 2.5 times faster than the standard mode. The model is available in the API via the identifier claude-fable-5-2 on Claude Platform, AWS Bedrock, Google Cloud, and Microsoft Foundry. Reported benchmark results: 49.1% on FrontierCode 2.0 (Xhigh), 89.3% on OSWorld 2.1, 74.2% on Humanity's Last Exam, 74.6% on Terminal-Bench-Science, and 69.9% on FrontierSWE v2 — according to these figures, the model outperforms Claude Opus 5.5 in nearly all metrics.

Context

To correctly interpret these figures, two key points are important. First is the comparative baseline: the previous flagship, Fable 5.1, was approximately 20% more expensive and slower at a comparable billing level, meaning the new generation simultaneously reduces price and response time while reportedly improving quality — a combination that was rare in flagship model releases until recently. Second is the quality control mechanism: adaptive thinking in Fable 5.2 is always enabled and is adjustable via the effort parameter, making the reasoning level a separate regulator for balancing cost and depth rather than a fixed model characteristic. A separate layer is the access division within a single generation: Fable remains publicly available with enhanced safeguards, while Mythos, as a carrier of high-risk capabilities, operates on a gated model with organizational verification, setting a new format for handling dual-use capabilities.

Why this matters for the industry

For the industry, the central shift is the economics of agentic pipelines. At a price of $8/$40 per million tokens, heavy reasoning loads on long context finally align in unit economics without special workarounds, while the batch API with a 50% discount covers offline volumes such as mass document processing. Anthropic states that at the High effort level, Fable 5.2 achieves the level of Opus 5.5 at approximately 50% of the cost, and on Humanity's Last Exam — at less than 10%; if these estimates are confirmed by independent measurements, migration from more expensive flagships will become a routine operation rather than a project requiring a separate budget. A 20% price reduction for the flagship amid increased speed also exerts pricing pressure on other providers: competitors' price lists and benchmark tables are now recalculated relative to a new baseline set by a single vendor.

Why this matters for users

Claude subscribers do not need to connect anything: Fable 5.2 is already available in Pro, Max, Ultra, and Team plans — Pro seats and standard Team seats receive it within a volume of up to 50% of the weekly limit, and paid subscribers received one to three additional resets of Fable limits. Those working via API have access to claude-fable-5-2 with a 1M token context, a maximum output of 128K, and Fast Mode for latency-critical scenarios. The most practical way to verify the benefit is to run an A/B test against the current production model on your own traffic: measure p50/p95 latency, quality on your own eval set, and the actual cost of the task at each effort level.

What is still unknown / limitations

All capability figures are vendor self-tests: the provided materials do not describe the harness, pass@k, scaffolds, or benchmark contamination checks, so results like 49.1% on FrontierCode 2.0 should be viewed as a declaration pending independent reproductions. The phrasing "outperforms Claude Opus 5.5 in nearly all metrics" is methodologically vague: it is not disclosed on which specific tests Opus 5.5 is still ahead or at what effort level the comparison was conducted — High effort is not equal to the default settings of real pipelines. The estimates of "Opus 5.5 level at 50% of the cost" and "less than 10% on Humanity's Last Exam" are vendor-origin enrichment data, not independent measurements, and building product unit economics on them before your own measurement is premature. Finally, aggregated averages across benchmarks hide variance: for agentic tasks, tail failures and stability over long horizons are decisive, and these metrics are not disclosed in Anthropic's statements; independent comparisons are expected to appear in the coming months.

Sources

Author

Look at AI, editorial team