UnifyBench

RankingHumanity's Last Exam

Humanity's Last Exam

Data updated 22 Oct 2026

Bucket
Hard reasoning
Unit
percent
Direction
Higher is better
Version
full
Display harness
HLE no tools
Board
https://lastexam.ai/

Compare published benchmark results with category weights →

Models

Results are ordered by score in the display harness, followed by models with no result. Other harnesses are listed separately below. Missing results remain unknown.

326–350 of 471 entries

ModelScoreEvidenceSource-recorded date
Phi-3-small-128k-instructMicrosoft———
Phi-3-small-8k-instructMicrosoft———
Phi-3-vision-128k-instructMicrosoft———
Phi-3.5-mini-instructMicrosoft———
Phi-3.5-MoE-instructMicrosoft———
Phi-3.5-vision-instructMicrosoft———
phi-4Microsoft———
Phi-4-mini-flash-reasoningMicrosoft———
Phi-4-mini-instructMicrosoft———
Phi-4-mini-reasoningMicrosoft———
Phi-4-multimodal-instructMicrosoft———
Phi-4-reasoningMicrosoft———
Phi-4-reasoning-plusMicrosoft———
Phi-4-reasoning-vision-15BMicrosoft———
Phi-mini-MoE-instructMicrosoft———
Phi-tiny-MoE-instructMicrosoft———
Pixtral 12BMistral———
Pixtral LargeMistral———
Qianfan-VL-3BBaidu———
Qianfan-VL-70BBaidu———
Qianfan-VL-8BBaidu———
QVQ-maxQwen———
QVQ-plusQwen———
Qwen-coder-plusQwen———
Qwen-coder-turboQwen———

Other harnesses

These runs use other harnesses. Capability scoring matches configurations separately; model details identify which source records contribute.

ModelScoreHarnessEvidenceSource-recorded date
Claude Fable 5Anthropic63.8%HLE with toolslab self-report

Same full HLE with web search, restricted web fetch, programmatic tools and code execution. Kept separate from no-tools model capability; production safeguard fallback applies.

2026-09-01
Claude Fable 5.1Anthropic65%HLE with toolslab self-report

Same full HLE with web search, restricted web fetch, programmatic tools and code execution. Kept separate from no-tools model capability; production safeguard fallback applies.

2026-09-01
Claude Opus 5Anthropic63.6%HLE with toolslab self-report

Same full HLE with web search, restricted web fetch, programmatic tools and code execution. Kept separate from no-tools model capability.

2026-09-01
Claude Opus 5Anthropic64.7%HLE with toolslab self-report

Tool-enabled result is explicitly incompatible with the preferred no-tools unit.

2026-07-24