Best LLM for each job
The top three models for the jobs teams build most, from the Respan Index (cited public benchmarks) and each provider's list prices.
The top three models for the jobs teams build most, from the Respan Index (cited public benchmarks) and each provider's list prices.
Ranked by GPQA Diamond, graduate-level biology, physics and chemistry questions written by domain experts to be hard to answer even with web search. Epoch AI runs every model on the same questions, so the scores compare like for like. Only models the Respan Index ranks are listed.
Price is the blended list price per 1M tokens, 3 parts input to 1 part output.
Each family's highest-ranked model for science, with its place in the ranking above.
The top two for science head to head, then the leading model families.
The ranking above uses GPQA Diamond. These public science benchmarks each have a full leaderboard, every score linked to its source.
Gemini 4 Argon from Google ranks first (science score 96.4%), followed by GPT-6 Astra (95.8%) and Claude Sonnet 5.5 (95.6%). Ranked by GPQA Diamond, graduate-level biology, physics and chemistry questions written by domain experts to be hard to answer even with web search. Epoch AI runs every model on the same questions, so the scores compare like for like. Only models the Respan Index ranks are listed.
GLM-5.3-Flash scores 90.2%, 94% of the top score at 81x lower blended list price ($0.05 per 1M tokens). It is the cheapest model within 15% of the top score.
The best Claude model here is Claude Sonnet 5.5 at #3 (95.6%); the best GPT model is GPT-6 Astra at #2 (95.8%).
Ranked by GPQA Diamond, graduate-level biology, physics and chemistry questions written by domain experts to be hard to answer even with web search. Epoch AI runs every model on the same questions, so the scores compare like for like. Only models the Respan Index ranks are listed. The GPQA Diamond leaderboard lists every score with a link to its source and is updated as new runs come out.
Yes. The Respan gateway serves models from every major provider behind one OpenAI-compatible API, with fallbacks when a provider fails.