BenchLeader

Best models for coding agents

An agent works for minutes at a time without anyone watching, so the wait before the first token is close to irrelevant. What decides it is whether the model writes code that passes, drives tools without losing the thread, and holds a large repository in context — and what that costs over a long session, since an agent burns tokens at a rate a chat never does.

As of 15 Nov 2026, Muse Wonder Full at max effort fits this best, with coding of 70 and tool use of 64.

Ranked on one measure instead: Best LLM for coding (November 2026) · Best LLM for agents (November 2026)

How this is weighted

  • 35%Codingthe work is writing code that actually runs
  • 30%Tool usean agent spends most of its turns calling tools and reading what came back, not composing prose
  • 20%Long contexta real repository does not fit in a small window, and attention that degrades over length shows up as forgotten decisions
  • 15%Price per 1Msessions run long, so a price difference that is invisible in a chat is the whole budget here

246 of 339 ranked models qualify: the rest either miss one of the measurements the weighting depends on, or fall outside a limit above. We would rather leave a model out than score it on the dimensions it happens to have.

Benchmark scores above are pooled across a model’s reasoning-effort settings, weighted by how many benchmarks each rests on. Publishers rarely run the same board at every setting, so a single setting’s category score often comes from one result — and one result moves too much to decide an order. Latency and price stay per-setting, because those genuinely differ.

#ModelFitCodingTool useLong contextPrice per 1MContext windowOpenRouter share
1Muse Wonder FullmaxMeta93706467–1M–
2Muse SparkMeta87656564–262k–
3Qwen3.8-Flash-NextAlibaba84616465$0.230256k–
4Gemini 4 CarbonhighGoogle83716964$4.001.0M0.24%
5NewClaude Fable 5.2maxAnthropic83758166$16.001M0.36%
6Mimo v2 OmniXiaomi82626561–256k–
7Kimi K3.1 PromaxMoonshot AI81656468$4.001.0M1.9%
8GPT-6.1 SolmaxOpenAI80656466$4.001.1M1.9%
9GPT-6.1 TerramaxOpenAI80636464$2.001.1M0.59%
10Claude Fable 5Anthropic78716366$20.001M0.01%
11Claude Fable 5.1maxAnthropic78676367$20.001M0.08%
12Gemini 4 ArgonhighGoogle78646464$4.001.0M0.43%
13Claude Opus 5.5maxAnthropic78696266$8.001M3.3%
14GPT-6 AstramaxOpenAI78706465$20.001.1M0.14%
15GPT-5.6 SolmaxOpenAI77666564$8.001.1M0.11%
16Inkling 2 SmallThinking Machines77605566$0.4751.0M0.48%
17Grok 4.6highSpaceXAI76596465$3.00500k0.03%
18Claude Sonnet 5.5maxAnthropic76616564$4.001M1.2%
19MiMo-V2.6-ProXiaomi76595568$0.5481.0M0.46%
20Muse Spark 1.3maxMeta75605966$2.001.0M0.03%
21Qwen3.6 MaxmaxAlibaba75567265$2.92262k–
22GPT-5.3-CodexxhighOpenAI75576566$4.81400k<0.01%
23Claude Opus 5maxAnthropic74656365$10.001M0.09%
24NewGemini 4 HydrogenhighGoogle746362–$3.001.0M0.05%
25Qwen3.8 MaxmaxAlibaba72615965$3.001M0.10%
26GPT-5.2-CodexOpenAI72556866$4.81272k<0.01%
27Muse WonderMeta72626265$8.001.0M0.06%
28Hy4Tencent726157–$1.25960k9.0%
29o3OpenAI72576271$3.50200k<0.01%
30Gemini 3 ProGoogle72586162–––
31Gemini 3.7 FlashhighGoogle71625365$1.501.0M0.14%
32Claude Opus 4.7Anthropic71636463$10.001M<0.01%
33GPT-6 SolmaxOpenAI71625964$4.001.1M0.12%
34DeepSeek V4 FlashmaxDeepSeek70535865$0.1751.0M1.1%
35Step 3.5 Flash 2603StepFun70586256$0.150256k–
36Muse Spark 1.1Meta70655463$2.001.0M<0.01%
37MiniMax M3.1 FlashMiniMax70555565$0.2621M1.6%
38Kimi K3maxMoonshot AI69605567$6.001.0M0.55%
39GLM 5.3 FlashZhipu AI69555565$0.2381M5.6%
40Claude Opus 4.5thinkingAnthropic69586762$10.00200k<0.01%

How to read this

Fit is a percentile blend, not a score out of a hundred: each dimension is ranked against every other model that could be judged here, then combined with the weights above. It says how well a model matches this job compared with the alternatives — a model can fit voice work superbly and sit well down the quality leaderboard, which is the point of ranking by job rather than by index.

The last column is not part of the fit score. It is the share of a week’s tokens that went to each model on OpenRouter, which is what people chose rather than what measured well — often a different answer. Read it as one marketplace’s slice: traffic sent straight to a maker’s own API never appears, and a model you run on your own hardware never appears at all, so a small share is not evidence that nobody uses it.

Disagree with the weighting? That is a reasonable thing to do, which is why it is printed rather than hidden. To set your own constraints instead, use the model finder.

Cite as: BenchLeader, “Best models for coding agents”, data as of 15 Nov 2026, https://www.benchleader.com/use/coding-agents.