BroadRouter

Open-weight models for science reasoning: published GPQA Diamond scores and API prices

Updated

Per Epoch AI's GPQA Diamond results (checked 2026-11-16; setting as listed by Epoch, no suffix), the highest published score among 10 open-weight models is 0.909 on a 0 to 1 scale, for MiniMax M3. Its lowest listed API prices in our snapshot of 2026-11-15 are USD 0.23 input and USD 0.96 output per 1M tokens. BroadNet did not run these benchmarks. BroadRouter is BroadNet's site.

"Best" here means the highest published score on one benchmark, per one source and one setting; the tables are the sources' numbers, sorted. BroadNet did not run these benchmarks; vendor-reported numbers use each vendor's own settings. BroadRouter is BroadNet's site. Prices are the lowest standard prices listed in our snapshot (newest check 2026-11-15), read from each provider's own pricing page.

GPQA Diamond is the hardest subset of GPQA, a set of graduate-level multiple-choice questions in biology, physics and chemistry written by domain experts. GPQA paper (Rein et al., 2023).

2 tables below, 19 open-weight models in all. Each table holds one benchmark from one source at one setting, with 4 or more models; a model can appear in more than one table. Scores from different tables are not comparable, so they are never merged or averaged.

GPQA Diamond per Epoch AI, setting as listed by Epoch, no suffix

Source: Epoch AI benchmark data, licensed CC BY 4.0, checked 2026-11-16. Epoch ran these itself (setting as listed by Epoch, no suffix). Values are Epoch's mean score on a 0 to 1 scale, rounded to three decimals (hover a value for the exact figure). Sorted by published score, higher first.

GPQA Diamond per Epoch AI, setting as listed by Epoch, no suffix: published score and lowest listed API prices per model
ModelPublished scoreLowest listed inputLowest listed outputPrices checked
MiniMax M30.909USD 0.23 CoreWeave ForgeUSD 0.96 CoreWeave Forge2026-11-15
Kimi K2.60.908USD 0.39 ChutesUSD 1.80 Chutes2026-11-15
GLM 5.10.899USD 0.90 ChutesUSD 2.90 Chutes2026-11-15
Kimi K2.7 Code0.879USD 0.60 CoreWeave ForgeUSD 3 CoreWeave Forge2026-11-15
GLM 50.878USD 0.60 GMI Cloud, promoUSD 1.92 GMI Cloud, promo2026-11-15
Qwen3.6 27B0.859USD 0.25 ChutesUSD 2 Chutes2026-11-15
NVIDIA Nemotron 3 Ultra0.854USD 0.50 CoreWeave ForgeUSD 2.15 CoreWeave Forge2026-11-15
Qwen3.6 35B A3B0.838USD 0.10 CoreWeave ForgeUSD 0.89 CoreWeave Forge2026-11-15
GLM 4.70.833USD 0.40 DeepInfraUSD 1.75 DeepInfra2026-11-15
Llama 3.3 70B0.474USD 0.10 DeepInfraUSD 0.32 DeepInfra2026-11-15

GPQA Diamond per Epoch AI, max reasoning setting

Source: Epoch AI benchmark data, licensed CC BY 4.0, checked 2026-11-16. Epoch ran these itself (max reasoning setting). Values are Epoch's mean score on a 0 to 1 scale, rounded to three decimals (hover a value for the exact figure). Sorted by published score, higher first.

GPQA Diamond per Epoch AI, max reasoning setting: published score and lowest listed API prices per model
ModelPublished scoreLowest listed inputLowest listed outputPrices checked
DeepSeek V4.1 Flash 10170.932USD 0.098 DeepInfra, promoUSD 0.294 DeepInfra, promo2026-11-15
Kimi K30.931USD 1.85 ChutesUSD 9.50 Chutes2026-11-15
GLM 5.20.919USD 0.5625 DeepInfra, promoUSD 1.80 DeepInfra, promo2026-11-15
DeepSeek V4 Pro 08130.917USD 0.85 ChutesUSD 2.60 DeepInfra2026-11-15
DeepSeek V4 Flash 07310.910USD 0.06 DeepInfraUSD 0.18 DeepInfra2026-11-15
GLM 5.3 · on BroadNet0.909USD 0.19 AkashML, promoUSD 2.20 GMI Cloud, promo2026-11-15
GLM 5.3 Flash0.902USD 0.075 GMI Cloud, promoUSD 0.25 GMI Cloud, promo2026-11-15
DeepSeek V4.1 Flash · on BroadNet0.898USD 0.13 CoreWeave ForgeUSD 0.42 DeepInfra, promo2026-11-15
DeepSeek V4 Pro0.896USD 0.957 GMI Cloud, promoUSD 1.914 GMI Cloud, promo2026-11-15

"Lowest listed": the lowest standard, non-free price per 1M tokens for the model in our snapshot, with the provider that lists it; input and output can come from different providers. Free rows and batch, flex, priority and fast tiers are on each model page but not here. "promo" marks a promotional price. BroadNet's own prices are not counted.

Epoch AI rows: Epoch AI, Capabilities & benchmarking, licensed CC BY 4.0 (checked 2026-11-16). BroadRouter only sorts Epoch's published values and adds prices.

Open models by use case · All models: API prices compared · How we collect prices

Questions

Which open-weight model has the highest published GPQA Diamond score per Epoch AI?

In Epoch AI's results (checked 2026-11-16), setting as listed by Epoch, no suffix: MiniMax M3 at 0.909 on a 0 to 1 scale, of 10 models; max reasoning setting: DeepSeek V4.1 Flash 1017 at 0.932 on a 0 to 1 scale, of 9 models. The settings differ, so the tables are not compared with each other. BroadNet did not run these benchmarks, and one benchmark says little about your own workload. The full table, with sources.

What do these models cost per 1M tokens?

Lowest listed standard prices in our snapshot (newest check 2026-11-15): MiniMax M3 USD 0.23 input and USD 0.96 output per 1M tokens; DeepSeek V4.1 Flash 1017 USD 0.098 input and USD 0.294 output per 1M tokens. Of the 19 models on this page with a price, the lowest listed input price is USD 0.06 per 1M tokens, for DeepSeek V4 Flash 0731 at DeepInfra. Input and output can come from different providers; each model page ranks its providers.

Does BroadNet serve any of these models?

Yes. BroadRouter is BroadNet's own site, and BroadNet Inference serves GLM 5.3 as broadnet/glm-5.3 at USD 0.78 input and USD 3.46 output per 1M tokens and DeepSeek V4.1 Flash as broadnet/deepseek-v4.1-flash at USD 0.175 input and USD 0.52 output per 1M tokens (rate card synced 2026-11-14). They are marked "on BroadNet" in the tables. BroadNet's prices are not counted in the lowest listed prices, and the scores are the published ones for the open weights, not a test of BroadNet's service.

Did BroadNet run these benchmarks?

No. Every score is as published by Epoch AI (checked 2026-11-16). Epoch AI ran its results itself and publishes them under CC BY 4.0. Each table keeps one benchmark, one source and one setting; we sort the published values and add prices, and compute no average or combined score.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity