Anthropic’s Claude Fable 5.2 scored 54.6% on OWE-Bench, a new benchmark of 26 long-horizon engineering projects that Mercor released on 12 November, Mercor’s announcement says. OpenAI’s GPT-6 Astra scored 44.2% and Claude Opus 5.5 41.6%.
By our arithmetic, Fable 5.2 led GPT-6 Astra by 10.4 points while spending $105.08 on tokens per attempt, 44.8% of the $234.63 that Astra spent, according to Mercor’s cost figures.
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 OWE-Bench.
The key insight: As we read Mercor’s figures, the top score on OWE-Bench came with less than half the token spend of the runner-up, so on this test the leading model was not the most expensive one to run; Mercor’s own runs also show no model near the ceiling.
Mercor calls the benchmark OpenWorldEngineer. Its 26 tasks span four domains: software engineering (8 tasks), physical engineering (5), machine learning (7) and scientific computing (6), from GPU kernels to aircraft design, the announcement says.
Agents get 4 to 18 hours per attempt, with a median of 8, for 257 agent-hours across one pass over all tasks. Every model makes eight attempts per task, with web search over a copy of the web frozen at 1 October 2026, the benchmark site says.
Each task starts with a domain expert recruited through Mercor and “can only be accepted if approved by three human reviewers”. Scores run from 0% for the untouched baseline to 100% at the task’s ceiling, so an agent that gets halfway earns 50% rather than a zero.
OWE-Bench scores from Mercor’s 12 November announcement, with token spend per attempt where Mercor gives it. 0% is the untouched baseline and 100% the task ceiling; the cost share is our arithmetic.
The Leaderboard
Mercor’s published scores: Claude Fable 5.2 54.6%, GPT-6 Astra 44.2%, Claude Opus 5.5 41.6%, GPT-6.1 Sol 41.2% and Claude Sonnet 5.5 39.0%. Two smaller models, Inkling and Beam, scored 1.1% and 3.3%. “No model is close to the benchmark ceiling”, the announcement says.
Fable 5.2 leads all four domains. In software engineering it scored 62.2% against 49.7% for Claude Opus 5.5. In physical engineering (59.6% vs 57.3%) and scientific computing (40.2% vs 39.4%), Mercor says GPT-6 Astra “is within the margin of error”.
What Each Point Costs
Mercor gives token spend per attempt for four models. By our arithmetic, that works out to about $1.92 per percentage point for Fable 5.2, $5.31 for GPT-6 Astra, $1.27 for GPT-6.1 Sol and $0.99 for Claude Sonnet 5.5.
Fable 5.2 used an average of 211 million tokens per attempt against 89 million for Astra, about 2.4 times as many, and still spent less. Claude Opus 5.5 used 253 million tokens; Mercor gives no dollar cost for it.
Where Agents Failed
Mercor read 624 runs from Claude Opus 5.5, Gemini 4 Argon and GPT-6.1 Sol. “Most failed runs shared one problem: the model checked its code on simpler tests than the actual ones in the tasks”, it writes. It found exploit attempts in 39 of 2,912 trials, 1.3%, and set exploits that raised a score to zero.
Mercor’s Position
Mercor recruits domain experts for AI work. Nvidia joined its $750M round at a $20B valuation, we reported on 10 November, and Mercor led $3.58B of AI rounds in three days, by our count on 11 November. The benchmark is Mercor’s own: its experts designed the tasks and it runs the leaderboard. All 26 tasks are public; 10 of the graders stay private because they cannot be separated from the hidden data, Mercor says.
The Structural Read
As we read the cost figures, Claude Sonnet 5.5 at 39.0% and $38.65 per attempt is the cheapest point on the board, and GPT-6 Astra the most expensive at $5.31 per point by our arithmetic.
As we read the method, scores are graded by execution on hidden inputs, and 10 graders stay private, so outside teams can run the tasks but not fully re-grade them.
Mercor, Introducing OWE-Bench, 12 November 2026
“Most failed runs shared one problem: the model checked its code on simpler tests than the actual ones in the tasks.”
Three Implications
FOR BUYERS OF AI AGENTS Mercor publishes token cost per attempt next to each score, and the cheapest and most expensive runs differ by about six times.
FOR MERCOR The benchmark adds a public leaderboard to a company that recruits domain experts for AI work.
FOR OUR COVERAGE We will add scores for the remaining models and any outside reproduction if they are published.
The framework argues: “Universal harnesses are likely to dominate highly digital, modular, and verifiable work.”
As we read it, OWE-Bench is built from verifiable work: each task is graded by an automated verifier on hidden inputs. The framework predates the benchmark and does not discuss it.
What Is Not Established
Mercor gives no overall score for Gemini 4 Argon or DeepSeek V4.1 Flash in the announcement, and no dollar cost for Claude Opus 5.5. The scores are Mercor’s own runs; no outside lab has reported reproducing them.
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.
On Mercor’s new 26-project engineering benchmark, Anthropic’s Fable 5.2 scored 54.6%, 10.4 points ahead of OpenAI’s GPT-6 Astra, at 44.8% of Astra’s token cost per attempt by our arithmetic. No model came close to the ceiling.
A note on sourcing. We describe the benchmark from Mercor’s announcement, its post on X and the benchmark site, all of 12 November 2026. Scores and token costs are Mercor’s own figures; the gaps, cost shares and cost per point are our arithmetic. Mercor’s funding comes from our earlier coverage.
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.
