Accuracy per dollar
Best value LLM APIs: accuracy vs price
The most accurate model is rarely the one to buy. Each dot below is an LLM: its GPQA Diamond score (graduate-level science questions, measured by Epoch AI) against the cheapest provider's price. The highlighted models are the ones no other model beats on both: nothing is cheaper and at least as accurate.
- Cheapest with 90%+$0.105 / 1MDeepSeek V4 Flash 0731 · 91.0% · DeepInfra
- Cheapest with 70%+$0.065 / 1Mgpt-oss-120b · 75.8% · W&B Inference (CoreWeave)
- Most accurate96.4%Gemini 4 Argon · from $4.00 / 1M · Google Gemini API
Accuracy: Epoch AI Benchmarking Hub (CC BY 4.0), Epoch's own runs, best reasoning setting, scores last changed Oct 17, 2026. Price: cheapest blended list price (3 input : 1 output tokens) across the providers we track, prices updated Oct 31, 2026; each provider name links its pricing page. How we compare.
The 10 best-value models, cheapest first
Going down the list, every step costs more and buys more accuracy. Any model not on this list has a cheaper alternative that scores at least as well.
| Model | GPQA Diamond | Cheapest / 1M | At |
|---|---|---|---|
| Llama-3.1-8B-Instruct | 27.0% | $0.0275 | Novita AI10 providers |
| Qwen2.5-7B-Instruct | 35.5% | $0.05 | SiliconFlow4 providers |
| gpt-oss-20bmedium reasoning | 60.8% | $0.055 | W&B Inference (CoreWeave)10 providers |
| Llama 4 Maverick 17B 128E Instruct FP8 | 67.0% | $0.0625 | Lambda6 providers |
| gpt-oss-120bhigh reasoning | 75.8% | $0.065 | W&B Inference (CoreWeave)18 providers |
| DeepSeek V4 Flash 0731max reasoning | 91.0% | $0.105 | DeepInfra10 providers |
| GPT-5.6 Lunamax reasoning | 91.6% | $0.45 | OpenAI3 providers |
| GLM-5.2max reasoning | 91.9% | $1.16 | DeepInfra13 providers |
| Gemini 3.8 Flashhigh reasoning | 95.4% | $1.50 | Google Gemini API2 providers |
| Gemini 4 Argonhigh reasoning | 96.4% | $4.00 | Google Gemini API2 providers |
Most accurate overall: Gemini 4 Argon at 96.4%, from $4.00 per 1M tokens.
Show the other 86 scored models
| Model | GPQA Diamond | Cheapest / 1M | At |
|---|---|---|---|
| GPT-6 Astramax reasoning | 95.8% | $20.00 | OpenAI3 providers |
| Claude Sonnet 5.5max reasoning | 95.6% | $4.00 | Anthropic4 providers |
| GPT-6.1 Solmax reasoning | 95.4% | $4.00 | OpenAI3 providers |
| Gemini 3.7 Flashhigh reasoning | 94.8% | $1.50 | Google Gemini API3 providers |
| Gemini 3.1 Pro Previewhigh reasoning | 94.4% | $4.50 | Google Gemini API2 providers |
| GPT-6 Solmax reasoning | 94.3% | $4.00 | OpenAI3 providers |
| Gemini 3.6 Flashhigh reasoning | 94.1% | $1.50 | Google Gemini API2 providers |
| Grok 4.6high reasoning | 94.0% | $3.00 | xAI3 providers |
| Claude Opus 5max reasoning | 93.9% | $10.00 | Anthropic5 providers |
| GPT-5.6 Solmax reasoning | 93.5% | $8.00 | OpenAI3 providers |
| GPT-5.6 Terramax reasoning | 93.3% | $4.50 | OpenAI3 providers |
| Kimi K3max reasoning | 93.1% | $5.40 | SiliconFlow9 providers |
| Gemini 3.5 Flashhigh reasoning | 92.8% | $3.38 | Google Gemini API3 providers |
| Grok 4.7xhigh reasoning | 92.7% | $3.00 | xAI3 providers |
| Qwen3.8 Maxxhigh reasoning | 92.7% | $2.48 | DeepInfra4 providers |
| DeepSeek V4 Pro 0813max reasoning | 91.7% | $1.63 | DeepInfra8 providers |
| Claude Opus 4.8max reasoning | 91.0% | $10.00 | Anthropic5 providers |
| GLM-5.3max reasoning | 90.9% | $1.68 | DeepInfra12 providers |
| MiniMax-M3 | 90.9% | $0.4125 | W&B Inference (CoreWeave)7 providers |
| DeepSeek V4 Prohigh reasoning | 90.9% | $1.63 | DeepInfra6 providers |
| Qwen3.7 Max | 90.9% | $1.88 | Novita AI4 providers |
| Kimi K2.6 | 90.8% | $1.34 | W&B Inference (CoreWeave)8 providers |
| GPT-5.5low reasoning | 90.7% | $11.25 | OpenAI3 providers |
| Claude Opus 5.5max reasoning | 90.6% | $8.00 | Anthropic4 providers |
| Claude Sonnet 5xhigh reasoning | 90.5% | $4.00 | Anthropic5 providers |
| GPT-6 Lunamax reasoning | 90.5% | $0.20 | OpenAI3 providers |
| GLM-5.3-Flashmax reasoning | 90.2% | $0.2375 | Z.ai11 providers |
| Claude Opus 4.7xhigh reasoning | 90.2% | $10.00 | Anthropic5 providers |
| GLM-5.1 | 89.9% | $1.66 | DeepInfra7 providers |
| DeepSeek-V4.1-Flashmax reasoning | 89.8% | $0.21 | DeepInfra8 providers |
| Claude Haiku 5.5max reasoning | 89.6% | $0.20 | Anthropic4 providers |
| Gemini 3 Flash Previewhigh reasoning | 89.4% | $1.13 | Google Gemini API2 providers |
| Grok 4.3high reasoning | 88.8% | $1.56 | xAI3 providers |
| Claude Opus 4.6max reasoning | 88.4% | $10.00 | Anthropic4 providers |
| Qwen3.6 Plus | 88.4% | $1.13 | Alibaba Cloud Model Studio2 providers |
| Inklingxhigh reasoning | 88.3% | $1.73 | DeepInfra3 providers |
| Kimi K2.7 Code | 87.9% | $1.36 | DeepInfra8 providers |
| Qwen3.7 Plus | 87.9% | $0.56 | Together AI3 providers |
| GLM-5 | 87.8% | $0.97 | DeepInfra5 providers |
| Qwen3.5 397B-A17Bnone reasoning | 86.4% | $1.09 | DeepInfra5 providers |
| Claude Fable 5max reasoning | 85.9% | $20.00 | Anthropic5 providers |
| Qwen3.6 27B | 85.9% | $1.03 | SiliconFlow4 providers |
| Gemini 2.5 Pro | 85.3% | $3.44 | Google Gemini API3 providers |
| Qwen3.6 35B-A3Bnone reasoning | 84.8% | $0.3125 | DeepInfra6 providers |
| Qwen3.5 35B-A3B | 83.5% | $0.355 | DeepInfra4 providers |
| Claude Sonnet 4.6high reasoning | 83.3% | $6.00 | Anthropic5 providers |
| GLM-4.7 | 83.3% | $0.7375 | DeepInfra5 providers |
| Gemini 3.5 Flash Litehigh reasoning | 83.3% | $0.85 | Google Gemini API2 providers |
| Gemini 3.1 Flash Litehigh reasoning | 81.8% | $0.5625 | Google Gemini API3 providers |
| qwen3-235b-a22b-thinking-2507 | 80.1% | $0.385 | Fireworks AI3 providers |
| Qwen3.5 9B | 79.0% | $0.1125 | DeepInfra3 providers |
| DeepSeek-R1-0528 | 76.3% | $0.30 | Lambda6 providers |
| Gemma 4 31B ITminimal reasoning | 75.8% | $0.16 | W&B Inference (CoreWeave)6 providers |
| Gemma 4 26B A4B ITminimal reasoning | 73.2% | $0.1375 | DeepInfra7 providers |
| qwen3-235b-a22b | 70.7% | $0.27 | DeepInfra3 providers |
| DeepSeek-R1 | 69.2% | $0.92 | SiliconFlow8 providers |
| deepseek-v3-0324 | 67.6% | $0.30 | Lambda5 providers |
| Qwen3 32B | 65.7% | $0.13 | DeepInfra7 providers |
| Qwq 32B | 65.3% | $0.185 | Nscale5 providers |
| Deepseek R1 Distill Qwen 32B | 64.1% | $0.15 | Nscale5 providers |
| Qwen3-14B | 63.8% | $0.12 | Nebius Token Factory4 providers |
| Qwen3 30B A3B | 61.7% | $0.15 | Nebius Token Factory3 providers |
| GLM-4.7-Flash | 60.5% | $0.145 | DeepInfra4 providers |
| Qwen3-8B | 56.8% | $0.06 | SiliconFlow2 providers |
| DeepSeek-V3 | 56.5% | $0.4375 | SiliconFlow8 providers |
| Phi-4 | 56.1% | $0.0875 | DeepInfra2 providers |
| deepseek-r1-distill-llama-70b | 55.7% | $0.30 | DeepInfra6 providers |
| Qwen3-30B-A3B-Instruct-2507 | 55.6% | $0.1425 | SiliconFlow3 providers |
| qwen3-4b | 52.3% | $0.12 | Nebius Token Factory2 providers |
| Llama 4 Scout 17B 16E Instruct | 51.8% | $0.0625 | Lambda7 providers |
| Qwen2.5-72B-Instruct | 49.1% | $0.1975 | Nebius Token Factory6 providers |
| mistral-large-2407 | 49.0% | $3.00 | Amazon Bedrock2 providers |
| Gemma-3-27b-it | 47.7% | $0.10 | DeepInfra5 providers |
| Mistral Small 3.1 | 47.5% | $0.15 | Google Cloud Vertex AI3 providers |
| Llama-3.3-70B-Instruct | 47.4% | $0.1975 | Nebius Token Factory13 providers |
| qwen2.5-32b-instruct | 46.1% | $0.095 | Nebius Token Factory2 providers |
| deepseek-r1-distill-qwen-14b | 44.7% | $0.07 | Nscale3 providers |
| llama-3.1-70b-instruct | 44.2% | $0.1975 | Nebius Token Factory7 providers |
| WizardLM 2 8x22B | 43.4% | $0.48 | DeepInfra2 providers |
| llama-3-2-90b-vision-instruct | 41.0% | $0.90 | Fireworks AI3 providers |
| Gemma 3 12B IT | 39.5% | $0.0625 | Novita AI3 providers |
| mistral-large-2402 | 38.8% | $6.00 | Amazon Bedrock2 providers |
| deepseek-r1-distill-qwen-1.5b | 33.6% | $0.09 | Nscale2 providers |
| mixtral-8x7b-instruct-v0.1 | 30.6% | $0.40 | DeepInfra3 providers |
| Gemma 3 4B IT | 23.2% | $0.05 | Amazon Bedrock2 providers |
| mistral-7b-instruct-v0.3 | 15.2% | $0.1117 | OVHcloud AI Endpoints2 providers |
Before you pick by this chart
GPQA Diamond measures hard science reasoning. It says little about coding, writing, tool use or your own prompts, so treat it as one filter, not a verdict. Scores use the best reasoning setting Epoch ran; lower settings are cheaper per answer but usually less accurate, and reasoning models spend extra output tokens, which a blended per-token price does not show. The score belongs to the model, not to a provider: a provider serving a quantized version can do worse. Only 96 of 272 models have an exact Epoch run; the others are not shown here, which says nothing about their quality.
Keep comparing
Picked a model?
Sort all LLMs by accuracy - Check what you pay today - Cheapest LLM APIs