The Number Every AI Lab Quoted Was Retired by the Lab That Made It Famous
Last Updated on October 29, 2026 by Editorial Team
Author(s): CodeInsights
Originally published on Towards AI.
OpenAI audited the hardest 138 problems in SWE-bench Verified and found 59.4% were broken. Then it found something worse: every frontier model could recite the answer.
On 23 February 2026, OpenAI published a post that no company writes about its own product category. The title was “Why we no longer evaluate SWE-bench Verified.”
The article argues that SWE-bench Verified became an un-auditable “prototype kilogram” for AI coding: not only were many failures due to broken or misleading tests, but models could also reproduce the exact gold patches or problem wording verbatim—suggesting memorization and leakage from training data. It explains that benchmark “defects” were largely test defects (narrow underspecified tests or wide unexpected requirements), and that replacing datasets didn’t fix the issue because the format itself incentivizes optimization against fixed evaluators. Examples from other benchmarks (like BrowseComp) show how models can game hidden structure or decrypt answer keys, gaining little real capability. The piece then introduces TestJack-style auditing that generates new tests to validate whether a proposed fix truly satisfies requirements, finding many reported successes don’t actually hold up under re-evaluation. Finally, it recommends building and auditing evaluations internally—using adaptive or dynamically generated tests—because leaderboards increasingly measure exposure to the benchmark rather than true software-writing ability.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ AI Engineering For Production — our new book, coming before the end of 2026. Join the waitlist.
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
![[Day 37/100] Build Your Own Agent Memory Layer with Postgres and pgvector](https://miro.medium.com/v2/resize:fit:700/1*-Doos2nmUT3FPZ-Qxbk3Hw.png "[Day 37/100] Build Your Own Agent Memory Layer with Postgres and pgvector")
