OpenAI finds roughly 30 percent of popular AI coding test is broken
Quick Answer
OpenAI's review of the SWE-Bench Pro coding test revealed that approximately 30% of its tasks are flawed, leading to the withdrawal of its endorsement.
Quick Take
OpenAI's review of the Pro coding test revealed that approximately 30% of its tasks are flawed, leading to the withdrawal of its endorsement. The review found discrepancies in task requirements, with human reviewers identifying more issues than AI agents, highlighting the need for better benchmarks in AI model evaluation.
Key Points
- 30% of SWE-Bench Pro tasks were found to be broken, affecting AI model assessments.
- Human reviewers flagged 249 flawed tasks, compared to 200 identified by AI agents.
- Some tests were too strict or vague, leading to misleading evaluations of AI capabilities.
- OpenAI calls for new benchmarks created by experienced developers to improve reliability.
- Codex with GPT-5.5 scored 31 points on SWE-Bench Pro, significantly lower than on other tests.
📖 Reader Mode
~2 min readOpenAI reviewed SWE-Bench Pro, a widely used test for measuring AI models' programming skills, and found roughly 30 percent of its tasks are broken. The company is pulling its earlier endorsement of the benchmark.
Results from tests like these feed into decisions about whether and how to release a model, including safety assessments under OpenAI's Preparedness Framework. When a test contains errors, it can paint a misleading picture of what an AI can actually do.
To run the review, OpenAI first deployed an automated screening tool that flagged 286 suspicious tasks. AI agents built on Codex then examined each case in detail before a human researcher made the final call. That process labeled 200 tasks (27.4 percent) as flawed. In a parallel review, five experienced software developers evaluated the same cases and flagged even more, 249 tasks (34.1 percent). The human reviewers were stricter than the AI agents, though both sides agreed in 74 percent of cases.
A single whitespace character can mean pass or fail
OpenAI breaks the problems into four categories. Some tests are too strict, rejecting solutions that actually work. Others are too vague, expecting the AI to meet requirements buried in hidden test cases. Some tests are too shallow, letting incomplete solutions pass. And some task descriptions simply point in the wrong direction. One example from the OpenLibrary project: the task description called for a single space, but the hidden test expected two. An AI that correctly followed the instructions would fail.
The tasks were pulled from the commit histories of real software projects, originally written for human collaboration, not designed as clean evaluation tasks for AI models. According to OpenAI, tests from those projects tend to be too strict because they were built to verify one specific change, not to serve as general-purpose requirements. On the public version of the test with 731 tasks, top models had jumped from 23.3 to 80.3 percent accuracy in just eight months. SWE-Bench Pro was meant to replace the older SWE-bench Verified, which OpenAI had already dismissed for similar reasons.
This time, OpenAI doesn't recommend a specific replacement. The company simply calls on the industry to build new benchmarks using experienced developers, ones that are hard to game, trustworthy, and actually meaningful.
In mid-June, the analytics firm Artificial Analysis had already removed SWE-Bench Pro from its Coding Agent Index and swapped in DeepSWE, a test from Datacurve. The reason: SWE-Bench Pro was gameable. Some models had copied the correct solution from a project's commit history instead of actually solving the task.
The switch reshuffled the leaderboard. Codex with GPT-5.5 (xhigh) climbed from 65 to 76 points and passed Claude Code with Opus 4.8 (max) at 73, while Claude Code with Fable 5 (max) took the top spot at 77 points. On SWE-Bench Pro, Codex with GPT-5.5 had scored just 31 points, compared to 64 to 84 on other tests.
— Originally published at the-decoder.com
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.
