📅 Data snapshot: November 2026

Using GitHub Copilot? Hit the Copilot filter above to see models listed as available there. Copilot uses per-token AI credit billing; the $/task column estimates a separate direct-API task cost. See GitHub's rate card for Copilot-specific rates.
Why some cells say "not tested". Each benchmark only scores the models its maintainers choose to run. LiveBench 2.0 started in October 2026 and has tested 32 models so far. Arena Code covers almost everything because anyone can vote. SWE-bench Verified has published one independent run since February 2026, so only older models have a score there. "Not tested" means the source has no result for that model, not that the model scored badly. The table sorts by LiveBench 2.0, then by Arena.
Model Family Copilot $/task LiveBench 2.0 Arena Code SWE-bench
Claude Fable 5.1Anthropic✓$1.0085.71537not tested
GPT-6 AstraOpenAI✓$1.0085.61542not tested
Claude Opus 5.5Anthropic✓$0.4085.11551not tested
GPT-6.1 SolOpenAI✓$0.2084.11541not tested
Claude Sonnet 5.5Anthropic✓$0.2080.01529not tested
Gemini 4 ArgonGoogle✓$0.20‡79.31550not tested
GPT-6.1 TerraOpenAI✓$0.1077.8not testednot tested
Muse Wonder 1.0Meta--74.21545not tested
DeepSeek V4.1 FlashDeepSeek-$0.03*72.91523not tested
Kimi K3.1 ProMoonshot--71.3not testednot tested
Grok 4.7xAI✓$0.1670.11494not tested
Claude Haiku 5.5Anthropic✓$0.0169.81489not tested
MiMo V2.6 ProXiaomi--68.51538not tested
GPT-5.6 TerraOpenAI✓$0.2267.81517not tested
Grok 4.6xAI✓$0.1667.81509not tested
MiniMax M3.1 FlashMinimax--66.9not testednot tested
GPT-6 LunaOpenAI✓$0.0164.61506not tested
Kimi K3Moonshot✓-64.61540not tested
Muse Spark 1.3Meta--63.01534not tested
GLM-5.3 FlashZhipu--62.11526not tested
Qwen3.8 MaxAlibaba--61.61525not tested
Gemini 3.8 FlashGoogle✓$0.075†59.21530not tested
Qwen3.8 27BAlibaba--58.41498not tested
Claude Fable 5Anthropic✓$1.00not tested1550not tested
Claude Opus 5Anthropic✓$0.50not tested1534not tested
Claude Sonnet 5Anthropic✓$0.20not tested1520not tested
Claude Opus 4.8Anthropic✓$0.50not tested1534not tested
Claude Opus 4.7Anthropic-$0.50not tested1551not tested
Claude Opus 4.6Anthropic-$0.50not tested154875.6%
Claude Opus 4.5Anthropic-$0.50not tested153076.8%
Claude Sonnet 4.6Anthropic✓$0.30not tested1528not tested
Claude Sonnet 4.5Anthropic-$0.30not tested151971.4%
Claude Haiku 4.5Anthropic✓$0.10not tested148266.6%
GPT-6 SolOpenAI✓$0.20not tested1519not tested
GPT-5.6 SolOpenAI✓$0.40‡not tested1533not tested
GPT-5.6 LunaOpenAI✓$0.02not tested1502not tested
GPT-5.5OpenAI-$0.55not tested1519not tested
GPT-5.4OpenAI-$0.28not tested1520not tested
GPT-5.4 miniOpenAI-$0.08not tested1495not tested
GPT-5.4 nanoOpenAI✓$0.02not tested1460not tested
GPT-5.2OpenAI-$0.23not tested148972.8%
GPT-5 miniOpenAI-$0.03not tested143256.2%
Gemini 3.6 FlashGoogle-$0.075†not tested1518not tested
Gemini 3.1 ProGoogle-$0.22not tested1521not tested
Gemini 3.5 Flash-LiteGoogle-$0.04not tested1502not tested
Gemini 3 FlashGoogle-$0.06not tested150575.8%
Gemini 2.5 Pro (retiring Dec 1)Google-$0.16not tested146453.6%
Gemini 2.5 Flash (retiring Dec 1)Google-$0.04not tested142428.7%
Grok 4.5xAI✓$0.16not tested1514not tested
DeepSeek V4 Pro 0813DeepSeek-$0.106*not tested1505not tested
Muse Spark 1.2Meta--not tested1543not tested
GLM-5.3Zhipu--not tested1520not tested
GLM-5.2Zhipu--not tested1514not tested
GLM-5Zhipu-$0.05not tested149872.8%
GLM-4.7Zhipu-$0.05not tested1484not tested
Kimi K2.5Moonshot--not tested150470.8%
MiniMax M3Minimax--not tested1495not tested
MiniMax M2.5Minimax-$0.07not tested144675.8%

Column guide

Column What it means
Copilot Available in GitHub Copilot (✓ = listed, - = not listed). Copilot usage is billed in AI credits based on tokens used.
$/task Estimated direct-API cost for 50K input + 10K output tokens; it is not a Copilot bill estimate.
LiveBench 2.0 Weighted overall score (0-100) across seven suites: long agentic work, coding, design, browser use, data analysis, research and reasoning. Release 2026-10-25 (source). Not comparable with the old LiveBench 1 scores this page showed until October.
Arena Code Rating from blind human votes on coding prompts, with Arena’s default style control. Published October 28, 2026 (source). Where Arena lists several effort settings for a model, the highest-rated one is shown; hover a score to see which, and how many votes it has. Current models sit between about 1490 and 1550, and the 95% confidence intervals are roughly ±6 to ±18 points, so many neighbouring models are statistically tied.
SWE-bench % of 500 real GitHub issues fixed in the SWE-bench Verified standardized runs (source). Most runs date from February 2026; none of the models released since then has a run.

DeepSeek V4.1 Flash and V4 Pro have peak and off-peak rates; this estimate uses peak pricing. Arena’s DeepSeek V4.1 Flash score is for the September 10 build, LiveBench’s for the October 17 build. †Gemini 3.6/3.8 Flash promotional API pricing runs through Dec 31, 2026; after that it doubles to $0.15/task. ‡Promotional pricing until Nov 30, 2026: Gemini 4 Argon then rises to $0.40/task; OpenAI has not published GPT-5.6 Sol’s rate after the promotion. Claude Haiku 5.5’s $0.01 applies to prompts up to 100K tokens. Gemini 2.5 Pro and 2.5 Flash are deprecated; from Dec 1, 2026 Google routes their requests to Gemini 3.8 Flash and 3.5 Flash-Lite.

Aider’s polyglot scores are no longer in this table: its leaderboard hasn’t added a model since October 2025. The archived results are on the benchmarks page, along with ProgramBench and the LiveBench per-suite scores.

Data sources: LiveBench 2.0 (release 2026-10-25) · Arena Code (published Oct 28, 2026) · SWE-bench Verified (standardized runs to Feb 2026, plus one on Sep 1, 2026) · ProgramBench (leaderboard updated Oct 22, 2026) · GitHub Copilot (AI credit billing since Jun 2026)
API pricing: Anthropic · OpenAI · Google · DeepSeek · xAI · Zhipu (GLM)