gpt-oss-120B
Aug 2025 · Open weight · Current · supported evidence
OpenAI’s open-weight line. Fine to self-host; do not confuse it with the 5.6 API stack.
Intelligence
35
AA Index
API price
Unlisted
Input / output per 1M
Context
128K
Speed unlisted
Capability
Intelligence Index35
Coding index50
Agentic index22
Arena Elo (offset)55
Use it when
- ▸Self-host
- ▸Fine-tunes
- ▸Air-gapped
Skip if
- –You wanted GPT-5.6 quality
Benchmarks
GPQA Diamond Graduate-level science questions designed so Google search is not enough. Still one of the cleanest knowledge/reasoning splits. | 80.9% |
|---|---|
Humanity’s Last Exam Expert-written questions across fields. Harder than MMLU; the current differentiator for “does this model actually know things.” | — |
MMLU-Pro Harder, less-saturated successor to MMLU. Classic MMLU is above 90% for every flagship and no longer ranks the field. | 90% |
SWE-bench Verified 500 human-validated GitHub issues. Score swings 5–15 points by harness — treat vendor numbers as an upper bound. | 62.4% |
SWE-bench Pro Harder, contamination-resistant coding eval. Currently the best public split between “can code” and “can maintain a repo.” | — |
Terminal-Bench 2.1 End-to-end tasks in a real terminal. Better proxy for coding agents than HumanEval, which is fully saturated. | — |
ARC-AGI-2 Abstract visual puzzles. Rewards generalization over memorization. GPT-5.6 Sol currently leads the published set. | — |
AIME 2025 American Invitational Mathematics Examination. Contest math; reasoning-mode models dominate. | 97.9% |
MMMU College-level multimodal understanding across diagrams, charts, and exam figures. | — |
Modalities
text · tools
117B
Reasoning mode
reasoning
Hybrid and reasoning models spend tokens thinking. That raises GPQA and agents, and also raises latency and bill.