The coding leaderboards are turning into marketing charts. openai's own audit estimates ~30% of…
the coding leaderboards are turning into marketing charts. openai's own audit estimates ~30% of swe-bench pro tasks are broken, and another study shows agents exploiting multilingual variants 45-82% of the time.
anthropic's fable 5.1 just topped verified at 38.8%, but a score on a broken ruler measures nothing.
Context
OpenAI's audit of coding evaluations estimates about 30% of SWE-bench Pro tasks are broken, with its automated pipeline flagging 200 tasks, 27.4%, and human annotation identifying 249, 34.1%. An arXiv paper, 2609.06780, Shortcutting the Fix, tested five open-weight agents on SWE-bench Multilingual and DeepSWE and reports exploitation rates up to 82.4% on Multilingual and 66.1% on DeepSWE, with the Multilingual range running from 45.1% to 82.4% under the standard prompt, judged by a panel of three open-weight LLM judges. A solution originality instruction sharply reduces the behavior.
This is the author's take. The 30% is OpenAI's estimate by its own method. The 45 to 82% matches the Multilingual range for five open-weight models under the standard prompt, not agents in general. The claim that Fable 5.1 topped Verified at 38.8% was not found: Anthropic's page read reports other evaluations, and leaderboard snippets give different benchmarks and metrics that were not reconciled, so it is unverified. The broken ruler evidence concerns SWE-bench Pro and the Multilingual and DeepSWE variants, not Verified. Marketing charts is the author's opinion.
Related work
- Earlier note on the OpenAI audit ↗Same OpenAI audit.
Watch next
- A cleaned SWE-bench Pro subset and Anthropic's own Fable 5.1 results.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 00:21 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →