Anthropic's Claude Fable 5.2 Beats GPT-6 Astra With a Rare Five-Point Jump

Anthropic’s newly released Claude Fable 5.2 scored 174 on the Epoch Capabilities Index, according to Epoch AI’s model record. The result puts it five points above OpenAI’s GPT-6 Astra, the previous leader at 169. Epoch’s benchmark history shows that leading frontier models typically gain less than one ECI point per update, making this margin unusually large within the index.

Epoch calculated the preliminary score before launch using only its own benchmark suite, as the group noted when publishing the pre-launch results. The index changes as external evaluations arrive, so Fable 5.2’s position and score may shift.

A Score Built From More Than 50 Tests

The Epoch Capabilities Index combines more than 50 benchmarks into a general capability scale. Epoch calibrated Claude 3.5 Sonnet at 130 and GPT-5 at 150, then positioned other models relative to those reference points. ECI points represent relative placement on that scale, which makes 174 useful for comparing models rather than measuring a percentage success rate.

Epoch recently replaced its frequentist least-squares method with a Bayesian item response theory model. The method estimates underlying model capability while accounting for differences in benchmark difficulty. Scores for models above GPT-4 moved by less than one ECI point on average after the change, and the new model can estimate head-to-head win probabilities alongside point scores.

The Gains Cluster in Agents and Code

Anthropic released Fable 5.2 alongside Mythos 5.2, a less-restricted model aimed at cybersecurity work, according to the launch announcement. The company says both models generate output up to 30% faster than their 5.1 predecessors. API pricing fell 20% to $8 per million input tokens and $40 per million output tokens. An optional fast mode runs at roughly 2.5 times the standard speed while doubling token prices.

Anthropic reports its largest gains on coding and agentic benchmarks, which test how well a model completes multistep work with limited human direction.

Effort settings control how much inference compute the model can spend on an answer. At Xhigh effort, Fable 5.2 reaches 49.1% on FrontierCode 2.0, the highest result among released models Anthropic has tested. At Medium effort, it matches Opus 5.5’s top score on Anthropic’s WANDR search benchmark at about 40% of the cost per task.

Separate evaluators also report strong results, although each suite measures different tasks and calculates costs differently.

Artificial Analysis places Fable 5.2’s Max-effort cost per task at roughly 2.6 times Opus 5.5’s cost and twice Fable 5.1’s. Those figures show why token prices alone cannot determine the cheaper model for a production workload.

The Acceleration Hypothesis Needs Evidence

Epoch says Fable 5.2 sits above the trajectory predicted by the pre-Astra trend for reasoning models. That observation concerns the rate of capability growth: the model landed above the line formed by earlier releases instead of extending the previous pace.

Anthropic links the result to AI-assisted research and development. Its launch materials claim that 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in 30 hours and that the model solved two open problems in black-hole physics during testing.

Epoch has not yet published an analysis connecting AI-assisted research to the five-point ECI gain. The preliminary score, Anthropic’s research demonstrations, and the underlying cause remain separate pieces of evidence until that work arrives.

Choose the Effort Level Before Comparing Costs

Production teams can use Medium effort as the initial comparison point because it is the default in Claude apps and Claude Code. Anthropic says Fable 5.2 completes difficult tasks with fewer tokens than Fable 5.1 at this setting, allowing lower token use to compound the 20% API price reduction.

• Medium: Compare quality, latency, and cost against the current production model under the default setting.
• Xhigh or Max: Reserve additional compute for tasks where higher completion rates justify the higher cost per query.
• Fast mode: Test workloads where lower latency can offset doubled token prices.
• Cost accounting: Measure cost per successful task, including retries and tool calls, rather than relying on token prices alone.

VentureBeat’s analysis likewise notes that cost-per-task comparisons depend heavily on the selected effort setting. The Max configuration that leads some benchmarks may produce a different economic result from the Medium configuration used in many applications.

Fable 5.2 trails several models on narrower business-task evaluations. Its AA-Briefcase 1.1 preference Elo is 1785, below Opus 5.5 at 1800. On GDPval-AA 2.1, Fable scores 1831, behind Opus at 1866 and Muse Wonder Full at 1857. Anthropic continues to recommend Opus 5.5 as its everyday model, while positioning Fable 5.2 for demanding reasoning, coding, and long-horizon agent tasks where higher completion rates can justify the added compute.