Cognition has released FrontierCode 2.0, a coding benchmark designed to test whether AI agents can handle the incomplete instructions and unstated constraints found in real software work. Even the top-scoring model, Claude Opus 5.5, reached 45%, leaving plenty of room between today’s agents and a clean bill of engineering health.
Cognition Watch analysis
What happened
FrontierCode 2.0 draws tasks from 55% more repositories than version 1.1, adds 43% more blocking criteria and raises the bar for correctness and design. Its reference solutions change 67% more files and 89% more lines on average, in codebases 25% larger, according to Cognition’s benchmark write-up.
The company says the benchmark tests functional correctness, interpreting and extending a user’s initial prompt, and making sound design and integration choices. Its results put Claude Opus 5.5 at 45%, GPT-6 Astra at 42.8% and GPT-6.1 Sol at 23.9%. Cognition says the benchmark is not close to saturated: the leading score is still below half.
Why it matters
Many coding tests hand an AI a detailed specification and check whether it follows it. FrontierCode is aimed at the messier bit of the job: working out what a user probably meant, respecting the conventions already in a codebase and writing changes that are thorough enough to merge.
That makes this a more pointed test of coding agents’ usefulness than a simple pass-or-fail coding puzzle. But the scores are Cognition’s, using its own tasks and grading rubric. They are evidence about performance on this benchmark, not a universal ranking of models or a guarantee of how one will perform in your repository.
Our read
The 45% headline is less a podium finish than a reminder that “can write code” and “can take an engineering task off your hands” remain different claims. FrontierCode’s emphasis on inferred constraints and thorough fixes is a worthwhile direction for evaluation. The next test is whether the benchmark’s results line up with what independent teams see in everyday development, rather than just producing a fresh leaderboard with more demanding homework.
What to watch
- Whether independent researchers reproduce the model rankings on FrontierCode 2.0.
- How scores change as coding agents improve, and whether the benchmark remains challenging.
- Whether performance on its tasks predicts success in real software teams.
Discussion spark: Should coding-agent benchmarks reward strict obedience to a prompt, or the judgement to infer what a working codebase needs?
Sources and evidence
- Source update (27 October 2026, 17:00 UTC)
not affiliated with or endorsed by Cognition