What happened
Epoch AI’s model page lists Claude Fable 5.2 as an Anthropic model released on November 9, 2026. It is marked as a closed-weights model. The page gives it an Epoch Capabilities Index (ECI) score of 174 and places it first among 239 tracked models. ECI is presented alongside a benchmark scorecard rather than as a single-task evaluation.
In Epoch AI’s selected comparison table, Claude Fable 5.2 is ahead of Claude Opus 5.5 at 169, GPT-6 Astra at 169, and Gemini 4 Argon at 167 on the ECI. The table labels these as best scores and breaks out results in games, mathematics, multimodal tasks, science, and world knowledge. For Fable 5.2, the listed figures include 95% on FrontierMath Tiers 1–3 (v2), 100% on FrontierMath Tier 4 (v2), 96% on GPQA Diamond, and 94% on SimpleQA Verified.
The individual rows also show why the aggregate rank should not be read as universal superiority. Fable 5.2 and Gemini 4 Argon both record 96% on GPQA Diamond. On SpireClimb, Fable 5.2 records 34%, the same figure shown for GPT-6 Astra; Claude Opus 5.5 records 20%, while Gemini 4 Argon has no result listed. On Chess Puzzles, Fable 5.2 has 79%, against 76% for GPT-6 Astra, 67% for Opus 5.5, and 59% for Gemini 4 Argon. Epoch AI also lists 90% for Fable 5.2 on Mystery Game Puzzles, Earthborne Rangers (EBR-bench), and the Furniture Assembly Benchmark.
Epoch AI lists pricing of $8 per million input tokens and $40 per million output tokens for Claude Fable 5.2. Its metadata marks the knowledge cutoff, training parameters, training compute, training hardware, training time, and training dataset size as unknown. The entry does not describe an open-weights download, self-hosted deployment, or API availability terms; it identifies the model’s accessibility as closed weights. The ranking is therefore a published snapshot of measured benchmark results and listed model metadata, not an account of a particular customer deployment.
What the archive adds
New Runtime covered Claude Fable 5.2 on November 9, 2026, noting its formalization of Fermat’s Last Theorem in 30 hours (Introducing Claude Fable 5.2 and Claude Mythos 5.2). This ECI result is a new fact about that known model story, not a restatement of the formalization result. It adds a broad benchmark comparison and listed token pricing rather than another account of a single mathematical achievement.
Trend signal
This does not yet fit a pattern New Runtime tracks. An aggregate benchmark lead can guide a shortlist, but it does not by itself demonstrate a durable change in the architecture, economics, interoperability, or governance of agentic software.
Why it matters
For a team choosing a model, an aggregate score is useful mainly as a prompt to ask better evaluation questions. Claude Fable 5.2’s first-place ECI result may justify putting it into a model trial, especially for workloads that resemble the scorecard’s mathematics, puzzle, science, or multimodal tasks. It does not establish that it will be the best option for a particular codebase, tool environment, latency requirement, or budget.
The published breakdown offers a practical reminder to compare task-level evidence. Fable’s 100% score on FrontierMath Tier 4 differs in kind from its tie with Gemini 4 Argon on GPQA Diamond and its tie with GPT-6 Astra on SpireClimb. A blended index conceals those distinctions. Teams should construct an evaluation set from their own representative tasks, record success rates and failure modes, then compare those results with the listed $8 input and $40 output price per million tokens.
Closed weights are also a meaningful boundary. Organizations needing to inspect parameters, run models in their own infrastructure, or control the full serving stack cannot infer that path from this model entry. For those teams, the benchmark result may still be informative, but it is not a direct procurement answer.
Limits and uncertainty
Epoch AI’s page reports benchmark scores, a ranking, and model metadata, but it does not show how well those results predict a specific production workflow. The entry does not list confidence intervals, test costs, latency, reliability under repeated runs, or outcomes from deployed customer systems. Several metadata fields—including knowledge cutoff and training details—are marked unknown.
The listed prices are token prices, not a complete cost of ownership. They do not establish the cost of retries, orchestration, human review, tool calls, or context management in an agent workflow. Nor does the closed-weights label clarify API availability, regional access, rate limits, or contractual terms. The results have not been independently reproduced here, and the scorecard should be treated as comparative benchmark evidence rather than a guarantee of deployment performance.
Skill to build now
Build a small model-selection harness around a representative slice of your work this week. Include successful tasks, known edge cases, and tasks where failure is costly; capture quality, elapsed time, token use, and the reviewer effort required to accept an answer. Run the same harness on shortlisted models rather than extrapolating from a composite leaderboard. If your work uses tools or long context, keep those conditions intact, because a clean benchmark score may not represent the system you actually operate.
Reader takeaway
Claude Fable 5.2 now leads Epoch AI’s ECI table at 174, ahead of the selected comparison models. That is a credible reason to test it, not a complete decision rule. Use the scorecard to identify relevant capabilities, then validate quality, cost, operational constraints, and review burden on the work your team actually needs to complete.