Top score on 225 Exercism exercises across 6 languages. Uses high reasoning setting.

Second highest on Aider polyglot. High reasoning mode.

Third on Aider polyglot with 32k think budget.

o3 high reasoning effort matches GPT-5 low at 81.3%. Cost $21 vs GPT-5 low at $10.

GPT-5 at low reasoning effort. Full effort hits 88%. Cost only $10 at low effort.

Fourth on Aider polyglot leaderboard. Strong multi-language code editing.

Up 7pp from Opus 4.6 (72.0%). Strong multi-language code editing.

Best open-weight Aider Polyglot score. 22x cheaper than GPT-5 per run. Evaluated by aider.chat leaderboard on 225 Exercism exercises.

32k thinking budget. Gap to GPT-5 highlights agentic coding is Anthropic weakness vs OpenAI.

Chat mode variant, 4 points below Reasoner mode. Cost-efficient at $0.88/run.

Outperforms Qwen3-Coder-480B (60.4) on Aider polyglot with 3B active params, showing code editing strength.

Open-weight MoE, 1T params / 32B active. Competitive cost at $1.24 but below DeepSeek V3 performance.