How we choose

Every morning at 6:30 (Rome time) we repeat the choice on freshly downloaded prices and scores.

The rule

Each priority has a monthly budget. Within it we look at the Artificial Analysis Coding Index: models within 3 points of the best are treated as close enough, a choice of ours rather than a measure of uncertainty. Among them, the one OpenCode users keep using most the following week wins (the cheaper on a tie); without usage figures, the cheapest wins. The model for hard problems is chosen the same way with the larger budget, and named only if it scores more than 3 points higher. If the runner-up is within 3 points, we show it as an alternative.

PriorityEvery dayHard problems
Spend less5 USD50 USD
Balanced20 USD100 USD
Best results50 USD250 USD

Budgets in USD per month for bug fixing. For other kinds of work they grow with the tokens: see the «Budget» column in the table alongside.

How we estimate the monthly cost

Kind of workInputOutputFrom cacheBudget
Small edits4 M0.4 M20 M0.49×
Bug fixing8 M0.9 M40 M1.00×
New features12 M1.8 M55 M1.54×
Refactoring15 M2.2 M70 M1.90×
Understanding a codebase20 M0.6 M90 M2.09×

Tokens per month for one person using a coding agent every day, times the provider's price, fees included. They are stated assumptions, not measurements of your usage, and they are the same for every model: one that reasons at length may use more and cost more than the estimate. The «Budget» column says how much that work costs compared with bug fixing, as the median of today's offers. A missing price never becomes zero: if the cache price is missing, those tokens cost as much as input.

What we guarantee

  • Recent data. Only data read in the last 7 days counts, OpenCode usage figures included. Artificial Analysis does not publish when it ran a test: the date is when we read the score.
  • No retired models. When the maker retires a model we no longer show it, nor its dated builds or resellers' variants.
  • The configuration that was measured. The same model performs differently depending on its reasoning effort. We only count what you can set, without being able to guarantee the result on your project: with OpenAI's own API we write it into the configuration, with Anthropic's it is "high", OpenCode's default; elsewhere we count the lowest measured score.
  • New versions. If a new version is not yet measured on code, but its Intelligence Index is at least that of the previous one, it uses the previous one's score, flagged as provisional. In 71 past cases the new version was at least as good 67 times: a fact about the past, not a guarantee for the next one. The list is published.

When an offer is excluded

  • The price is more than 48 hours old.
  • The provider was available less than 90% in the last half hour, or less than 98% over the last day. Direct sellers often do not publish it: then the figure is missing, and we do not treat it as good.
  • It is bought straight from a cloud platform (Amazon Bedrock, Google Vertex, Azure): that needs a cloud account and extra permissions. Through OpenRouter the account is OpenRouter's.
  • The context is too small for the chosen kind of work.
  • OpenCode does not recognise the model-provider pair: the command would not start. SiliconFlow is the exception: OpenCode does not ship with it, and the configuration we publish adds it.
  • The price changed more than 5 times since the previous update: it waits for confirmation.

Limits

  • We cover the providers our sources reach, not the whole market.
  • A benchmark measures standard tasks: a serious signal, not a guarantee about your code.
  • Every score comes from Artificial Analysis: the other benchmarks we know publish measurements that are months old.
  • OpenRouter does not publish which providers retain what you send. On the card you can turn privacy on: the configuration asks OpenRouter to use only providers that do not retain it, and if the one shown does not guarantee that it picks another, possibly at another price. It is off by default. For direct purchases the seller's own policy applies.
  • OpenCode does not switch to the model for hard problems on its own: pick it by hand with /models.

All the details, including how models are matched, in METHODOLOGY.md