Coding

HumanEval: 2026 AI Leaderboard

164 Python programming problems: does the generated code pass unit tests?

What it tests

HumanEval is 164 handwritten Python programming problems with hidden unit tests. A model sees the function signature plus docstring and must generate a body that passes every test.

How it is scored

pass@1 -- the percentage of problems solved on the first attempt. Frontier models in 2026 sit in the 94-99% range, so this benchmark is effectively saturated for top-tier LLMs.

Why it matters

Still useful as a floor-check: any serious coding model should clear 90% here. For real-world discrimination, SWE-bench Verified and LiveCodeBench are the benchmarks that still separate the field.

Leaderboard (13 models)

Sorted by HumanEvalscore. Tier column shows the tool's overall AIToolTier rank, which blends this benchmark with pricing, features, and real-world usability.

#ModelTierHumanEval score
1Codex (OpenAI)
GPT-5.2-Codex (HISTORICAL -- this model was RETIRED 2026-07-23; scores retained for trend context only. Current Codex default is GPT-5.6 Sol, whose first-party Codex benchmarks are pending)
A95%
2ChatGPT
GPT-5.5 (launched 2026-04-23; scores below are the GPT-5.4 baseline -- GPT-5.5 launch benchmarks per OpenAI are logged in Known Issues, pending third-party verification)
A95%
3Claude (Anthropic)
Claude Opus 5 (launched 2026-07-24) is the current flagship and the default on Max -- Anthropic released COMPARATIVE claims only, no absolute scores: more than doubles Opus 4.8 on Frontier-Bench, within 0.5% of Fable 5 on CursorBench 3.2 at half the cost, 3x the next-best model on ARC-AGI 3, ~1.5x on Zapier AutomationBench, and beats Fable 5's best OSWorld 2.0 result at just over a third of the cost. Fable 5 (2026-06-09) holds vendor SWE-Bench Pro 80.3% (vs GPT-5.5 58.6%), #1 LMArena Elo 1510 and #1 Artificial Analysis Index 65 as of 6/11. Sonnet 5 (2026-06-30): OSWorld-Verified 78.5%. Legacy Opus-line reasoning-suite scores shown below as baseline pending third-party suites for Opus 5
A94%
4Qwen (Alibaba)
Qwen3.5-397B MoE
A92.5%
5Mistral AI
Mistral Medium 3.5 (vendor-published; third-party verification pending)
B92%
6DeepSeek
DeepSeek V4-Pro (SWE-bench + Arena Elo third-party verified post-launch; knowledge rows are V3.x baseline pending V4 figures). V4.1-Flash (2026-09-10) vendor-only so far: GPQA Diamond 90.9, DeepSWE 74.2, Terminal-Bench 2.1 90.6 -- third-party verification pending
A91.5%
7Muse Spark (Meta)
Muse Spark
A91%
8Grok
Grok 4.20 (baseline -- Grok 4.5 launched 2026-07-08; vendor-reported 4.5 scores in Known Issues pending third-party verification)
B90%
9Nemotron (Nvidia)
Llama-Nemotron Ultra 253B (prior gen -- Nemotron 3 Ultra 550B third-party scores pending)
B89.6%
10GLM / Z.ai (Zhipu AI)
GLM-5.2 (~753B MoE, launched 2026-06-13) -- vendor-published; third-party verification still settling
A89.1%
11Llama 4 (Meta)
Llama 4 Maverick (17B/400B MoE)
B88%
12Gemma 4 (Google)
Gemma 4 31B
A85%
13Falcon (TII)
Falcon 3 10B
B73.8%

About HumanEval

Creator
OpenAI, 2021
Unit
% (max 100)

Other benchmarks