Summary:
Anthropic releases Fable 5.2 with lower listed API prices. Three sourced benchmarks help frame evaluation, while Mythos 5.2 remains restricted to verified users.Anthropic released Claude Fable 5.2 on November 9, pairing a broadly available model for demanding technical work with a more restricted Mythos 5.2 configuration. For teams evaluating an upgrade, the important questions are which version they can access, what the workload costs and whether the reported gains match their own tasks.
Fable is public; Mythos requires verification
Fable 5.2 is available through Anthropic’s platform and its cloud distribution channels. Mythos 5.2 uses the same underlying model with more permissive safeguards for sensitive cybersecurity and biology work; direct access requires the relevant verification program. Public availability does not mean every plan includes unlimited usage.
Anthropic lists standard API prices of US$8 per million input tokens and US$40 per million output tokens. Optional fast mode costs US$16 and US$80 respectively. These are token rates, not a fixed project price: longer reasoning, repeated attempts and tool use can change the total.
Fable 5.2 has also been added to ChatbotsPlace’s Fable 5.2 chat. Check the service’s current account and usage conditions before starting a substantial workload.
Three benchmarks, three measurement contexts
The comparison below uses the November 9 system card’s named model versions. All values are percentages, and higher is better within a row. Bold identifies the largest reported value among these four models; it does not establish a statistically significant difference.
| Evaluation | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| FrontierSWE v2 | 69.9 | 56.3 | 62.3 | 65.5 |
| AutomationBench | 51.4 | 31.4 | 42.5 | 41.4 |
| Chartography, no tools | 73.4 | 44.8 | 64.4 | 71.0 |
FrontierSWE v2: Proximal evaluates 34 long engineering and research tasks in its own agent harness, with a 20-hour budget per trial and five trials per task. Models use their maximum reasoning effort. That makes the results useful for difficult, extended work, but different models’ maximum settings are not equal-cost budgets.
AutomationBench: Zapier’s private 1.0.6 set contains 657 tasks per effort level, covering multistep business workflows. The system card attributes the runs to Zapier and uses the default API fallbacks for the Claude comparison. The published scores measure task completion, not savings achieved by a deployed business.
Chartography: This 100-task chart-reading test grades answers against expert-set ranges. The card reports Anthropic’s Claude runs and Surge AI’s published Astra result. It uses Gemini 3.5 Flash grading for comparability; earlier cards using another grader should not be mixed into this table. Tool-enabled scores are separate and are not included here.
Choose a test that resembles the job
A coding team can compare accepted fixes, review effort and cost on a small set of its own difficult issues. An operations team should test complete workflows, including policy compliance and recovery from missing information. A research team working with charts should include the actual chart types it encounters and decide whether tools are allowed before comparing answers.
Use the same task inputs, permissions and success criteria for each model. Record failed attempts as well as successful ones, and measure total spend rather than only the advertised token price. The three benchmarks above should not be averaged into a universal score: they test different abilities and use different evaluation procedures.
Teams evaluating sensitive work should also distinguish model capability from access approval. For context, our GPT-6 Cyber coverage explains another release where specialized access conditions matter.
Sources
- Anthropic’s November 9 launch announcement: availability and listed API prices.
- Claude Fable 5.2 and Mythos 5.2 system card, November 9, 2026, sections 8.1, 8.7, 8.13.1 and 8.14.5: benchmark results and methods.
- ChatbotsPlace Fable 5.2: model-specific chat destination.
The comparison image is an original PromptsZone chart of the cited results, not a product screenshot.
