Model

Developer

Release

GPQA Diamond

Context window

Notes on willingness for diagnostic biotech

GPT-5 / GPT-5 Pro

OpenAI

Aug 2025

88.4–89.4% (Pro, no tools)

256K–400K

Generally engaged, medical-domain stress-tested

GPT-5.2 / GPT-5.2 Pro

OpenAI

Late 2025

92.4–93.2%

400K

State-of-the-art reasoning

Gemini 2.5 Pro

Google DeepMind

Mar 2025

~86.4%

1,048,576 (1M) tokens

Strong AI Co-Scientist multi-agent (Nature 2025)

Grok 4 / Grok 4 Heavy

xAI

Jul 2025

87–88.9% (Heavy)

256K (API)

Less restrictive on scientific content

Claude Opus 4.5

Anthropic

Late 2025

~87%

200K

Documented over-restriction, 0.23 % benign refusal vs. 0.05% Sonnet 4.5

DeepSeek V3 / V3.1 / R1

DeepSeek-AI

Dec 2024 – 2026

68.4 → 81 (R1-0528)

128K

Open-weights MIT license

Llama 4 Scout / Maverick

Meta

2026

~75–80 %

10M (Scout)

Open weights, best for self-hosted long context

Mistral Large 3

Mistral AI

2026

~75%

128K

EU-based, Apache 2.0 (Large 3 / Small 4)

Qwen 3 / 3.5

Alibaba

2026

81.7 (Qwen 3.5-9B), 90+ flagship

Up to 1M

Apache 2.0, multilingual (29–201 languages)