Leaderboard

Name: GMP Bench — LLM Benchmark for Pharmaceutical Manufacturing
Creator: GMP Bench
License: https://creativecommons.org/licenses/by/4.0/

Compare model performance across GMP knowledge and task completion benchmarks. Click a model name to view detailed results.

Task type

Creator

Weights

#	Model	Overall	GMP Knowledge	Task Completion	Avg Latency	Total Tokens	# Evals
1	Claude Opus 4.6	95.5%	100.0%	90.9%	32.6s	69k	39
2	GPT-5.4	94.2%	100.0%	88.4%	8.3s	27k	39
3	Claude Sonnet 4.6	94.2%	100.0%	88.3%	27.6s	73k	39
4	Claude Haiku 4.5	93.0%	100.0%	86.1%	10.3s	56k	39
5	DeepSeek V4 Pro	92.9%	97.1%	88.6%	28.7s	48k	40
6	DeepSeek-R1	92.7%	97.1%	88.2%	19.8s	33k	39
7	MiniMax M2.7	92.4%	97.1%	87.6%	19.2s	68k	40
8	Gemini 3.1 Pro	91.2%	97.1%	85.2%	22.8s	54k	39
9	Qwen3.6 35B A3B	91.1%	94.3%	87.9%	12.0s	83k	40
10	GPT-5.4 mini	90.5%	100.0%	81.0%	2.6s	21k	39
11	Qwen3.6 27B	90.4%	94.3%	86.5%	33.6s	83k	40
12	Mistral Large 3 675B	90.1%	100.0%	80.2%	10.2s	30k	39
13	DeepSeek-V3.2	88.2%	97.1%	79.3%	19.2s	46k	79
14	Gemini 3 Flash	87.5%	100.0%	75.0%	3.9s	21k	39
15	Mistral Small 2603	87.1%	100.0%	74.2%	3.1s	26k	39
16	GPT-5.4 nano	86.1%	94.3%	77.8%	2.8s	23k	39
17	Gemma 4 26B A4B IT	85.3%	97.1%	73.5%	8.9s	25k	40
18	Llama 4 Maverick	84.1%	100.0%	68.3%	15.2s	25k	39
19	Gemma 4 31B IT	83.7%	97.1%	70.3%	24.1s	39k	40
20	DeepSeek V4 Flash	83.2%	94.3%	72.2%	40.6s	89k	40
21	Gemini 3.1 Flash-Lite	78.7%	97.1%	60.3%	2.4s	17k	39
22	DeepSeek-R1-Distill-Qwen-32B	77.8%	100.0%	55.6%	45.6s	33k	39
23	Llama 3.3 70B Instruct	77.7%	94.3%	61.2%	60.6s	23k	39
24	Qwen3.5-397B-A17B	77.4%	65.7%	89.2%	30.8s	81k	39
25	Llama 4 Scout	75.0%	97.1%	52.8%	4.5s	23k	39
26	Qwen3.5-35B-A3B	70.9%	54.3%	87.5%	30.1s	174k	39