Best LLMs for coding

Coding ability is the hardest capability to summarise in one number, because 'writing code' is really four separate skills: producing a correct patch, applying that patch to an existing file without mangling it, navigating an unfamiliar repository, and recovering when a test fails. Each of the benchmarks below measures a different one, which is why we weight rather than pick.

SWE-bench Verified carries the most weight because its pass criterion cannot be argued with: the model's patch either makes the project's own hidden tests go green or it does not. It is the closest public proxy for 'can this model close a real ticket'. Its limits are worth knowing — it is Python-only, drawn from a narrow set of large repositories, and results shift by ten points or more depending on the agent harness wrapped around the model. Compare scores from the same harness or not at all.

Aider Polyglot fills the language gap and, more importantly, measures edit-format compliance across C++, Go, Java, JavaScript, Python and Rust. A model that reasons brilliantly but emits malformed diffs will feel broken inside an IDE regardless of its SWE-bench number, and this is the benchmark that exposes that failure. Terminal-Bench then adds the long-horizon dimension: planning, running commands, reading stderr and course-correcting without a human in the loop. LiveBench acts as the contamination-resistant control, since its questions rotate and cannot have been memorised.

A practical way to use the table: shortlist the top three, then check price and throughput before committing. Coding workloads are token-hungry, and a model two points behind at a third of the cost is usually the better production choice. Reasoning models also emit large volumes of hidden thinking tokens, so measure cost per merged pull request rather than cost per token — the ordering often changes when you do.

Finally, treat gaps under roughly three points as ties. These evaluations have a few hundred items each, which puts several points inside the noise band. If two models are that close, the deciding factors should be latency, context window, tool-calling reliability and how they behave on your own repository — not the leaderboard.

More on how these numbers are produced in the methodology and what each evaluation measures in the benchmark guide. Spotted a score that disagrees with its source? Tell us.