Every model with a public price, every SWE-bench run of an agent harness on a model, and the results of our own arena. Sort it, compare up to four models, and price your own workload. Empty cells mean nobody has measured it yet.
The same model can score very differently depending on the agent harness around it (the loop, tools and prompts). Each cell is the best SWE-bench result for that harness + model pair, % of real GitHub issues resolved.
Prices & context: OpenRouter's public model catalog, USD per million tokens, refreshed on every data sync.
SWE-bench Verified: % of 500 human-validated GitHub issues an agent resolved; we show the best harness for each model.
SWE Multilingual: every model run in the same open harness (mini-SWE-agent), so it compares models fairly; its run cost gives $ per solved task.
ABUZ8 Arena: our own tool-use arena on our own GPU, regex baseline included.
SWE-bench leaderboards: swebench.com (Princeton & Stanford, MIT). Pricing: openrouter.ai. Arabic model scores: OALL. Speed and intelligence indexes: Artificial Analysis (joins this table once our free API key is configured).
Model names are matched across sources by normalised name; a missing score means no match, not a zero. Submissions not verified by the SWE-bench team are marked by the filter, not hidden.