Anthropic’s Claude Fable 5.2 (Max) ranks first on Arena’s Agent Arena leaderboard with a net improvement score of 30.35%, according to the board’s Overall view dated 11 November, which we read that evening. OpenAI’s highest entry, GPT-6 Astra (Max), is sixth at 14.54%, behind four Claude models.
By our arithmetic, Fable 5.2’s score is about 2.1 times GPT-6 Astra’s, and Anthropic holds four of the top five places: Fable 5.2 first, Claude Opus 5.5 (Max) second at 19.52%, Claude Sonnet 5.5 (Max) fourth at 15.95% and Claude Fable 5.1 (Max) fifth at 14.82%. Google’s Gemini 4 Carbon (High) is third at 16.98%.
The page lists 59 models and 2,810,198 total sessions, and says the ranking is “based on signals like task success, user satisfaction, and steerability”. OpenAI holds sixth to eighth, with GPT-6.1 Sol (Max) at 12.46% and GPT-6.1 Terra (Max) at 12.00%. Gemini 4 Argon (High) is ninth and Meta’s Muse Wonder Full (Max) tenth at 8.96%.
By our arithmetic the top ten splits four Anthropic, three OpenAI, two Google and one Meta. Each score carries a 95% interval on the page: Fable 5.2’s is ±3.93 points, the widest in the top ten, on 4,731 sessions, the fewest. Even so, its range of about 26.4% to 34.3% sits clear of Opus 5.5’s range of about 17.5% to 21.6%.
The fifth-to-sixth step is different. Fable 5.1 at 14.82% ±2.06 and GPT-6 Astra at 14.54% ±2.02 are 0.28 points apart, and their intervals overlap almost entirely, by our arithmetic, so the board does not separate those two.
Agent Arena’s overall net improvement scores for the top ten of 59 models, with the 95% intervals the board shows, as read on the evening of 11 November 2026. Colours mark the lab. The 2.1x ratio and the overlap of fifth and sixth place are our arithmetic.
Business Pill · Benchmarks Explained
An evergreen Business Engineer explainer on why a test score is not your own result. It teaches the general idea only and does not discuss Arena’s board.
The key insight: As we read the board, the frontier ranking on agent work and the cost ranking point in opposite directions: Fable 5.2 leads with a score about 2.1 times OpenAI’s best entry, while its median task cost of $7.34 is about 5.1 times Gemini 4 Carbon’s, so the cheaper models score more points per dollar, by our arithmetic.
What a Task Costs
The board also lists a median cost per task. Fable 5.2’s is $7.34, against $5.47 for Opus 5.5, $1.45 for Gemini 4 Carbon and $1.07 for GPT-6.1 Sol. By our arithmetic, Fable 5.2 costs about 5.1 times Gemini 4 Carbon per task and about 6.9 times GPT-6.1 Sol.
Divide each score by its median task cost and the order changes: about 11.7 points per dollar for Gemini 4 Carbon and 11.6 for GPT-6.1 Sol, against about 4.1 for Fable 5.2 and 3.6 for Opus 5.5, by our arithmetic. The page lists Fable 5.2’s price at $8 per million input tokens and $40 per million output tokens.
Arena’s methodology post of 4 June says that rather than pairwise votes, “rankings are calculated using a methodology we call causal tracing”. It randomises the components of each agent session and estimates each choice’s effect on signals such as task success rates, tool error recovery and tool hallucinations. It says it typically measures that effect against a uniform distribution over components.
As we read it, that makes the score a relative measure: a model’s net improvement is its estimated effect compared with an even mix of the field, not a pass rate.
The Structural Read
As we read the methodology, a net improvement score is an effect measured against a uniform mix of components, so a lead of 10.83 points over second place, by our arithmetic, describes how much Fable 5.2 moves an agent session relative to the field rather than a share of tasks completed.
As we read the intervals, the board separates its first place clearly but not its fifth and sixth: Fable 5.1 and GPT-6 Astra sit 0.28 points apart with overlapping 95% intervals, so OpenAI’s sixth place rests on a gap the board itself cannot resolve.
Arena, Agent Arena methodology post, 4 June 2026
“rankings are calculated using a methodology we call causal tracing”
Three Implications
FOR ANTHROPIC The board lists four Claude models in its top five, with Fable 5.2 on 4,731 sessions, the fewest in the top ten.
FOR OPENAI Its three entries in the top ten sit sixth to eighth, from 14.54% to 12.00%.
FOR BUYERS The board lists median task costs from $0.72 to $7.62 across the top ten, with one model unlisted.
The framework says: “The surface, not the model, governs autonomy.”
As we read it, Arena’s method is built around that split: it treats the agent as a multi-component system and scores the orchestrator model as one component among several, which is why its board measures models inside a harness rather than on a fixed test. The framework predates this board and does not discuss it.
What Is Not Established
The leaderboard updates as sessions arrive, so these ranks are a snapshot from the evening of 11 November. A ratio of two net improvement scores, and points per dollar of task cost, are our arithmetic, not measures Arena publishes.
Fable 5.2’s lead rests on fewer sessions than any other top-ten model. We did not contact Arena, Anthropic or OpenAI for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Arena’s Agent Arena board of 11 November ranks Claude Fable 5.2 (Max) first at 30.35%, about 2.1 times OpenAI’s best entry, GPT-6 Astra (Max), which is sixth behind four Claude models. By our arithmetic, Fable 5.2 also costs about 5.1 times Gemini 4 Carbon per task, and fifth and sixth place sit within each other’s intervals.
A note on sourcing. We read Arena’s Agent Arena leaderboard page (Overall view, dated 11 November 2026) twice on the evening of 11 November and transcribed the top rows; the board changes as sessions arrive. We read Arena’s 4 June 2026 methodology post in full. Ratios, intervals, per-dollar figures and counts by lab are our arithmetic. We did not contact Arena, Anthropic or OpenAI for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
