MODEL EVIDENCE / SOURCE-REPORTED RESULTS

Choose models for the work.

Choose the capabilities and benchmarks that matter to you. Select several, or leave both on All to explore every result.

STEMBench120 · audit snapshot 14 October 2026. Tool-assisted Codex and closed-book API runs have different resources. All ten runs report zero full-target successes. Fractional progress is not percentage solved.

Find modelConfigurationOrder

10 runs shown · 0 selected

Published source results; higher means more credited progress. Configurations differ. Select rows to compare a subset. Per-domain values retain zero-score attempts; six domains have 20 tasks each.
Model / configuration	Mean / 4	Biology	Chemistry	Physics	Astronomy	Earth science	Engineering
GPT-6 Astra xhighTools · Codex	2.809	2.709	2.777	3.025	2.723	2.726	2.894
GPT-6.1 SolTools · Codex	2.759	2.656	2.730	2.986	2.692	2.676	2.811
Muse WonderClosed-book API	2.683	2.615	2.703	2.714	2.647	2.659	2.759
Claude Opus 5.5Closed-book API	2.621	1.865	2.666	2.808	2.759	2.760	2.872
Hy4 PreviewClosed-book API	2.586	2.587	2.611	2.521	2.583	2.598	2.615
Mistral Large 4Closed-book API	2.501	2.459	2.432	2.431	2.597	2.486	2.602
DeepSeek V4.1 FlashClosed-book API	2.490	2.546	2.462	2.202	2.602	2.594	2.538
Gemini 4 ArgonClosed-book API	2.444	2.348	2.336	2.083	2.778	2.647	2.471
Qwen3.8-27B NVFP4Closed-book API	2.314	2.289	2.517	1.788	2.400	2.516	2.376
MiMo V2.6 ProClosed-book API	2.172	2.533	2.498	0.983	2.223	2.623	2.172

SOURCE SNAPSHOT / RETRIEVED 25 OCTOBER 2026

Published results and full methodology ↗Read the benchmark field notes →

What else belongs in a model comparison?

Find relevant benchmarks →

Quality & reliability

Use task success, proof validity, instruction compliance, factual error, calibration, and failure severity. Keep their original units.

Cost per successful task

Include input, output, reasoning, cache, and tool costs plus retries. Token price alone does not capture total work.

Latency & throughput

Separate time to first token, output rate, end-to-end task time, and tail latency. Pin provider, region, concurrency, and workload.

Context & modalities

Test effective retrieval and synthesis at realistic lengths. Label text, image, document, and tool inputs separately.

Deployment & access

Record model revision, provider, reasoning effort, context limits, rate limits, weight availability, and license constraints.

Post-training & robustness

Match the exact base checkpoint and inference budget. Check transfer, regressions, contamination, seed variation, and held-out tasks.

Results are dated source snapshots. Benchmark leaderboards collect the available reports without treating different evaluation settings as one matched experiment. Broad-market prices and provider speed are not yet ingested here. Explore the Artificial Analysis measurements and use this checklist to keep quality, economics, and deployment constraints distinct.