Best AI models for coding.

A short function, a debugging session and a change across a repository ask different things of a model. This shortlist emphasizes coding and agentic coding results, with reasoning and instruction following also contributing. Choose a difficulty level around the scope of the change you need.

The list weighs the published results that matter for coding and keeps the models close enough to the best one for a balanced task. Price picks the provider after that; it never lifts a weaker model into the list. 382 available models were considered and 42 have comparable evidence.

Each price is the cheapest provider read in the last three days, per M tokens, blended 3 in : 1 out.

Published coding results depend on the benchmark and evaluation setup. They do not predict success on your repository or guarantee that a provider reproduces the measured configuration.

The ranking uses the same Balanced settings as the Ask page. The Capability methodology has the sources, the exact identity rules, the task weights and the difficulty ranges.