Open-weight models by use case, per published benchmarks
Updated
A use case gets a page here when 4 or more open-weight models have a published result on the same benchmark, from the same source at the same setting. On 2026-11-16 that holds for 1 use case: science reasoning (GPQA Diamond), from Epoch AI. Each model shows its lowest listed API price. BroadNet did not run these benchmarks. BroadRouter is BroadNet's site.
Use cases with a page
- Science reasoning: 19 open-weight models in 2 tables (GPQA Diamond per Epoch AI, setting as listed by Epoch, no suffix, 10 models; GPQA Diamond per Epoch AI, max reasoning setting, 9 models).
Use cases without a page yet
These have published rows in our data, but no benchmark with 4 or more open-weight models from one source at one setting, so we do not rank them:
- Coding: 44 published rows for 21 models across 27 benchmark names.
- Math: 6 published rows for 5 models across 2 benchmark names.
- Knowledge and factuality: 12 published rows for 12 models across 5 benchmark names.
- Agents and tool use: 6 published rows for 6 models across 6 benchmark names.
- Vision and documents: 4 published rows for 3 models across 3 benchmark names.
- Instruction following: 3 published rows for 3 models across 2 benchmark names.
- Speech to text: 11 published rows for 2 models across 11 benchmark names.
Each model's own results, with sources, are on its price page under "Published benchmark results".
Epoch AI rows: Epoch AI, Capabilities & benchmarking, licensed CC BY 4.0.
All models: API prices compared · Prices by model maker · How we collect prices
Questions
What does "best" mean on these pages?
Only the highest published score on one benchmark, per one source, at one setting. We sort the published values in each table and add prices; we compute no average or combined score, and we never put two sources or two settings in one ranking. One benchmark says little about your own workload.
Why are there so few use cases?
A page needs 4 or more open-weight models with a result on the same benchmark, from the same source at the same setting. On 2026-11-16 only 1 use case meets that rule. Published rows exist for coding (44 rows across 27 benchmark names), math (6 rows across 2 benchmark names), knowledge and factuality (12 rows across 5 benchmark names), agents and tool use (6 rows across 6 benchmark names), vision and documents (4 rows across 3 benchmark names), instruction following (3 rows across 2 benchmark names) and speech to text (11 rows across 11 benchmark names), but makers name and set up their benchmarks differently, so no benchmark there reaches 4 comparable models yet.
Where do the scores come from?
From the model makers' own model cards and from Epoch AI's benchmark data (licensed CC BY 4.0, checked 2026-11-16), each with its source link and check date. BroadNet did not run these benchmarks. We do not use Artificial Analysis or OpenRouter benchmark data.
Which open-weight models have the highest published scores?
For science reasoning, per Epoch AI (setting as listed by Epoch, no suffix, checked 2026-11-16): MiniMax M3 at 0.909 on a 0 to 1 scale on GPQA Diamond, of 10 models. All science reasoning tables.