Which model is worth it?
Capability against what it costs you — dollars per token if you call the model, GPU memory if you host it. These are frontiers, not rankings: each table lists the models that nothing else beats on both axes at once. Where to sit on the curve is your call.
What the capability axis is
A single latent ability fitted across the boards that measure broadly the same thing. The number is on a log scale centred on the average model: 0 is average, higher is more capable. It is one dimension — roughly three quarters of the ordering across all axes — and it deliberately cannot represent an axis that measures something else, such as hallucination resistance. A model can sit low here and be the right choice for a job this number does not capture.
Value verdicts
A pick per budget, and who was left out of it and why. Each pick is the most capable frontier model that fits the budget; the list under it is every model in the budget that something both more capable and costing no more makes redundant.
Best under $1 / M tokens
Xiaomi: MiMo-V2.6-Pro$0.761 per million blended tokens · ability 3.73 from 19 boards. No cheaper model scores higher.
Left out: 75 models
- Xiaomi: MiMo-V2.5-Pro · $0.761 · 3.36 — beaten by Xiaomi: MiMo-V2.6-Pro
- Anthropic: Claude Haiku 5.5 · $0.400 · 3.19 — beaten by Z.ai: GLM 5.3 Flash
- DeepSeek: DeepSeek V4 Flash 0731 · $0.342 · 3.02 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: GPT-5.6 Luna · $0.950 · 2.88 — beaten by Xiaomi: MiMo-V2.6-Pro
- Tencent: Hy3 · $0.429 · 2.79 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M2.1 · $0.975 · 2.70 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.6-Flash · $0.245 · 2.67 — beaten by Z.ai: GLM 5.3 Flash
- Upstage: Solar Pro 3 · $0.487 · 2.66 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-6 Luna · $0.400 · 2.61 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M3 · $0.975 · 2.28 — beaten by Xiaomi: MiMo-V2.6-Pro
- StepFun: Step 3.7 Flash · $0.912 · 2.16 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.5 · $0.245 · 2.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.6 35B A3B · $0.787 · 1.86 — beaten by Xiaomi: MiMo-V2.6-Pro
- DeepSeek: DeepSeek V3.2 · $0.385 · 1.83 — beaten by Z.ai: GLM 5.3 Flash
- DeepSeek: DeepSeek V3.2 Exp · $0.375 · 1.73 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 235B A22B Instruct 2507 · $0.435 · 1.68 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.6V · $0.750 · 1.62 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Tencent: Hy3 preview · $0.495 · 1.58 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- MiniMax: MiniMax M2.7 · $0.975 · 1.56 — beaten by Xiaomi: MiMo-V2.6-Pro
- DeepSeek: DeepSeek V3.1 Terminus · $0.818 · 1.45 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-9B · $0.137 · 1.40 — beaten by Qwen: Qwen3.7 Flash
- Google: Gemma 4 26B A4B · $0.186 · 1.32 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: GPT-5.4 Nano · $0.988 · 1.31 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.6 Flash · $0.891 · 1.29 — beaten by Xiaomi: MiMo-V2.6-Pro
- Google: Gemma 4 31B · $0.350 · 1.07 — beaten by Z.ai: GLM 5.3 Flash
- StepFun: Step 3.5 Flash · $0.250 · 1.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.5-35B-A3B · $0.787 · 1.05 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-Flash · $0.211 · 1.00 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M2.5 · $0.780 · 0.97 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 Next 80B A3B Instruct · $0.848 · 0.95 — beaten by Xiaomi: MiMo-V2.6-Pro
- Upstage: Solar Pro 4 · $0.293 · 0.90 — beaten by Z.ai: GLM 5.3 Flash
- NVIDIA: Nemotron 3 Nano 30B A3B · $0.195 · 0.79 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 30B A3B Instruct 2507 · $0.247 · 0.73 — beaten by Z.ai: GLM 5.3 Flash
- NVIDIA: Nemotron 3 Super · $0.358 · 0.65 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.5 Air · $0.670 · 0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Arcee AI: Trinity Large Thinking · $0.662 · 0.57 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 Next 80B A3B Thinking · $0.937 · 0.50 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 4.7 Flash · $0.315 · 0.49 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemini 2.5 Flash Lite · $0.325 · 0.48 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M2 · $0.829 · 0.41 — beaten by Xiaomi: MiMo-V2.6-Pro
- Meta: Llama 4 Maverick · $0.536 · 0.40 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: gpt-oss-120b · $0.487 · 0.35 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 14B · $0.210 · 0.34 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 27B · $0.282 · 0.29 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 32B · $0.230 · 0.26 — beaten by Z.ai: GLM 5.3 Flash
- Inception: Mercury 2 · $0.625 · 0.23 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-5 Nano · $0.313 · 0.22 — beaten by Z.ai: GLM 5.3 Flash
- DeepSeek: DeepSeek V3 · $0.747 · -0.01 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 30B A3B · $0.405 · -0.08 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 12B · $0.125 · -0.11 — beaten by Qwen: Qwen3.7 Flash
- Qwen: Qwen3 VL 30B A3B Instruct · $0.487 · -0.25 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 4 Scout · $0.250 · -0.25 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen-Plus · $0.650 · -0.33 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-4.1 Nano · $0.325 · -0.42 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 Coder 30B A3B Instruct · $0.228 · -0.56 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.3 70B Instruct · $0.265 · -0.57 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 Coder Next · $0.630 · -0.59 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-4o-mini · $0.487 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Mistral Small 3.1 24B · $0.504 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- IBM: Granite 4.2 8B · $0.203 · -0.78 — beaten by Z.ai: GLM 5.3 Flash
- Perplexity: Sonar · $1.00 · -0.83 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 VL 8B Instruct · $0.625 · -0.84 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Reka Flash 3 · $0.175 · -0.84 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 2 27B · $0.650 · -0.93 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 70B Instruct · $0.400 · -0.97 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 4B · $0.087 · -1.02 — beaten by OpenAI: gpt-oss-20b
- Microsoft: Phi 4 · $0.123 · -1.21 — beaten by Qwen: Qwen3.7 Flash
- Mistral: Mistral Small 3 · $0.072 · -1.23 — beaten by OpenAI: gpt-oss-20b
- Mistral: Saba · $0.500 · -1.48 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Cohere: Command R (08-2024) · $0.487 · -1.51 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Nous: Hermes 3 70B Instruct · $0.700 · -1.55 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- WizardLM-2 8x22B · $0.620 · -1.67 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 8B Instruct · $0.072 · -2.02 — beaten by OpenAI: gpt-oss-20b
- Meta: Llama 3.2 3B Instruct · $0.260 · -2.57 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.2 1B Instruct · $0.158 · -3.43 — beaten by Qwen: Qwen3.7 Flash
Best under $5 / M tokens
Meta: Muse Spark 1.3$3.50 per million blended tokens · ability 4.44 from 26 boards. No cheaper model scores higher.
Left out: 137 models
- Qwen: Qwen3.8 Max (0902) · $5.00 · 4.28 — beaten by Meta: Muse Spark 1.3
- Meta: Muse Spark 1.2 · $3.50 · 4.07 — beaten by Meta: Muse Spark 1.3
- Meta: Muse Spark 1.1 · $3.50 · 3.97 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 3.6 Flash · $3.00 · 3.86 — beaten by Google: Gemini 3.7 Flash
- Z.ai: GLM 5.2 · $1.97 · 3.48 — beaten by Z.ai: GLM 5.3
- Qwen: Qwen3.7 Max · $3.69 · 3.44 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 3 Flash Preview · $2.38 · 3.36 — beaten by Z.ai: GLM 5.3
- Xiaomi: MiMo-V2.5-Pro · $0.761 · 3.36 — beaten by Xiaomi: MiMo-V2.6-Pro
- SpaceXAI: Grok 4.5 · $5.00 · 3.31 — beaten by Meta: Muse Spark 1.3
- DeepSeek: DeepSeek V4 Pro 0813 · $1.65 · 3.20 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Anthropic: Claude Haiku 5.5 · $0.400 · 3.19 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 5.1 · $2.52 · 3.18 — beaten by Z.ai: GLM 5.3
- MoonshotAI: Kimi K2.6 · $2.72 · 3.11 — beaten by Z.ai: GLM 5.3
- Z.ai: GLM 5 Turbo · $3.30 · 3.04 — beaten by Google: Gemini 3.8 Flash
- DeepSeek: DeepSeek V4 Flash 0731 · $0.342 · 3.02 — beaten by Z.ai: GLM 5.3 Flash
- SpaceXAI: Grok 4.6 · $5.00 · 2.97 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-5.6 Luna · $0.950 · 2.88 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.7 Plus · $1.04 · 2.81 — beaten by Xiaomi: MiMo-V2.6-Pro
- Tencent: Hy3 · $0.429 · 2.79 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 5 · $2.65 · 2.72 — beaten by Z.ai: GLM 5.3
- Google: Gemini 3.5 Flash Lite · $1.95 · 2.71 — beaten by Z.ai: GLM 5.3
- MiniMax: MiniMax M2.1 · $0.975 · 2.70 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.6-Flash · $0.245 · 2.67 — beaten by Z.ai: GLM 5.3 Flash
- Upstage: Solar Pro 3 · $0.487 · 2.66 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- SpaceXAI: Grok 4.7 · $5.00 · 2.64 — beaten by Meta: Muse Spark 1.3
- Mistral: Mistral Large 4 · $1.74 · 2.64 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- OpenAI: GPT-6 Luna · $0.400 · 2.61 — beaten by Z.ai: GLM 5.3 Flash
- StepFun: Step 5 Preview · $2.28 · 2.59 — beaten by Z.ai: GLM 5.3
- Qwen: Qwen3.5 397B A17B · $2.83 · 2.44 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.4 Mini · $3.56 · 2.42 — beaten by Meta: Muse Spark 1.3
- Qwen: Qwen3.6 Plus · $1.54 · 2.40 — beaten by Xiaomi: MiMo-V2.6-Pro
- MoonshotAI: Kimi K2.7 Code · $2.68 · 2.37 — beaten by Z.ai: GLM 5.3
- MiniMax: MiniMax M3 · $0.975 · 2.28 — beaten by Xiaomi: MiMo-V2.6-Pro
- Google: Gemini 3.1 Flash Lite · $1.19 · 2.27 — beaten by Xiaomi: MiMo-V2.6-Pro
- SpaceXAI: Grok 4.3 · $2.19 · 2.24 — beaten by Z.ai: GLM 5.3
- Thinking Machines: Inkling · $3.29 · 2.24 — beaten by Google: Gemini 3.7 Flash
- MoonshotAI: Kimi K2.5 · $2.00 · 2.23 — beaten by Z.ai: GLM 5.3
- MoonshotAI: Kimi K2 Thinking · $2.02 · 2.22 — beaten by Z.ai: GLM 5.3
- Z.ai: GLM 4.7 · $1.62 · 2.20 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.8 27B · $1.71 · 2.17 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- StepFun: Step 3.7 Flash · $0.912 · 2.16 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.5 · $0.245 · 2.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.6 27B · $1.56 · 2.06 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 5V Turbo · $3.30 · 1.98 — beaten by Google: Gemini 3.8 Flash
- SpaceXAI: Grok Build 0.1 · $1.75 · 1.98 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Google: Gemini 3.1 Flash Lite Preview · $1.19 · 1.92 — beaten by Xiaomi: MiMo-V2.6-Pro
- NVIDIA: Nemotron 3 Ultra · $1.95 · 1.89 — beaten by Z.ai: GLM 5.3
- Qwen: Qwen3.5 Plus 2026-04-20 · $1.43 · 1.88 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.6 35B A3B · $0.787 · 1.86 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5.1-Codex-Mini · $1.56 · 1.85 — beaten by Xiaomi: MiMo-V2.6-Pro
- DeepSeek: DeepSeek V3.2 · $0.385 · 1.83 — beaten by Z.ai: GLM 5.3 Flash
- DeepSeek: DeepSeek V3.2 Exp · $0.375 · 1.73 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 235B A22B Instruct 2507 · $0.435 · 1.68 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.6V · $0.750 · 1.62 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3.5-122B-A10B · $1.63 · 1.58 — beaten by Xiaomi: MiMo-V2.6-Pro
- Tencent: Hy3 preview · $0.495 · 1.58 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- MiniMax: MiniMax M2.7 · $0.975 · 1.56 — beaten by Xiaomi: MiMo-V2.6-Pro
- Anthropic: Claude Haiku 4.5 · $4.00 · 1.51 — beaten by Meta: Muse Spark 1.3
- DeepSeek: DeepSeek V3.1 Terminus · $0.818 · 1.45 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 VL 235B A22B Instruct · $1.48 · 1.41 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-9B · $0.137 · 1.40 — beaten by Qwen: Qwen3.7 Flash
- Qwen: Qwen3.5-27B · $2.02 · 1.39 — beaten by Z.ai: GLM 5.3
- DeepSeek: R1 0528 · $1.74 · 1.37 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Z.ai: GLM 4.5 · $1.80 · 1.36 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Thinking Machines: Inkling Small · $1.01 · 1.33 — beaten by Xiaomi: MiMo-V2.6-Pro
- Google: Gemma 4 26B A4B · $0.186 · 1.32 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemini 2.5 Flash · $1.95 · 1.32 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.4 Nano · $0.988 · 1.31 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.6 Flash · $0.891 · 1.29 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 235B A22B Thinking 2507 · $2.74 · 1.24 — beaten by Z.ai: GLM 5.3
- DeepSeek: R1 · $2.05 · 1.10 — beaten by Z.ai: GLM 5.3
- OpenAI: o4 Mini High · $3.58 · 1.10 — beaten by Meta: Muse Spark 1.3
- Google: Gemma 4 31B · $0.350 · 1.07 — beaten by Z.ai: GLM 5.3 Flash
- StepFun: Step 3.5 Flash · $0.250 · 1.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.5-35B-A3B · $0.787 · 1.05 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5 Mini · $1.56 · 1.04 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-Flash · $0.211 · 1.00 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M2.5 · $0.780 · 0.97 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 Next 80B A3B Instruct · $0.848 · 0.95 — beaten by Xiaomi: MiMo-V2.6-Pro
- Upstage: Solar Pro 4 · $0.293 · 0.90 — beaten by Z.ai: GLM 5.3 Flash
- NVIDIA: Nemotron 3 Nano 30B A3B · $0.195 · 0.79 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 30B A3B Instruct 2507 · $0.247 · 0.73 — beaten by Z.ai: GLM 5.3 Flash
- MoonshotAI: Kimi K2 0905 · $2.02 · 0.70 — beaten by Z.ai: GLM 5.3
- OpenAI: o3 Mini High · $3.58 · 0.69 — beaten by Meta: Muse Spark 1.3
- MoonshotAI: Kimi K2 0711 · $1.87 · 0.68 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- NVIDIA: Nemotron 3 Super · $0.358 · 0.65 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: GPT-4.1 Mini · $1.30 · 0.64 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 4.5 Air · $0.670 · 0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Arcee AI: Trinity Large Thinking · $0.662 · 0.57 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 Next 80B A3B Thinking · $0.937 · 0.50 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 4.7 Flash · $0.315 · 0.49 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemini 2.5 Flash Lite · $0.325 · 0.48 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M1 · $1.79 · 0.41 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- MiniMax: MiniMax M2 · $0.829 · 0.41 — beaten by Xiaomi: MiMo-V2.6-Pro
- Meta: Llama 4 Maverick · $0.536 · 0.40 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: gpt-oss-120b · $0.487 · 0.35 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 14B · $0.210 · 0.34 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 27B · $0.282 · 0.29 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.5V · $1.50 · 0.26 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 32B · $0.230 · 0.26 — beaten by Z.ai: GLM 5.3 Flash
- Inception: Mercury 2 · $0.625 · 0.23 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-5 Nano · $0.313 · 0.22 — beaten by Z.ai: GLM 5.3 Flash
- DeepSeek: DeepSeek V3 · $0.747 · -0.01 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Mistral Medium 3 · $1.60 · -0.06 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 30B A3B · $0.405 · -0.08 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 12B · $0.125 · -0.11 — beaten by Qwen: Qwen3.7 Flash
- Qwen: Qwen3 VL 30B A3B Instruct · $0.487 · -0.25 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 4 Scout · $0.250 · -0.25 — beaten by Z.ai: GLM 5.3 Flash
- Mistral: Mistral Medium 3.1 · $1.60 · -0.30 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen-Plus · $0.650 · -0.33 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-4.1 Nano · $0.325 · -0.42 — beaten by Z.ai: GLM 5.3 Flash
- Mistral Large 2407 · $5.00 · -0.51 — beaten by Meta: Muse Spark 1.3
- Qwen: Qwen3 Coder 30B A3B Instruct · $0.228 · -0.56 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.3 70B Instruct · $0.265 · -0.57 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 Coder Next · $0.630 · -0.59 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Devstral 2 2512 · $1.60 · -0.61 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-4o-mini · $0.487 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Mistral Small 3.1 24B · $0.504 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- IBM: Granite 4.2 8B · $0.203 · -0.78 — beaten by Z.ai: GLM 5.3 Flash
- Perplexity: Sonar · $1.00 · -0.83 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 VL 8B Instruct · $0.625 · -0.84 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Reka Flash 3 · $0.175 · -0.84 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 2 27B · $0.650 · -0.93 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 70B Instruct · $0.400 · -0.97 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 4B · $0.087 · -1.02 — beaten by OpenAI: gpt-oss-20b
- Microsoft: Phi 4 · $0.123 · -1.21 — beaten by Qwen: Qwen3.7 Flash
- Mistral: Mistral Small 3 · $0.072 · -1.23 — beaten by OpenAI: gpt-oss-20b
- Mistral: Saba · $0.500 · -1.48 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Cohere: Command R (08-2024) · $0.487 · -1.51 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Nous: Hermes 3 70B Instruct · $0.700 · -1.55 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- WizardLM-2 8x22B · $0.620 · -1.67 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 8B Instruct · $0.072 · -2.02 — beaten by OpenAI: gpt-oss-20b
- Mistral: Mixtral 8x22B Instruct · $5.00 · -2.02 — beaten by Meta: Muse Spark 1.3
- Mistral Large · $5.00 · -2.42 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-3.5 Turbo · $1.25 · -2.54 — beaten by Xiaomi: MiMo-V2.6-Pro
- Meta: Llama 3.2 3B Instruct · $0.260 · -2.57 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.2 1B Instruct · $0.158 · -3.43 — beaten by Qwen: Qwen3.7 Flash
Best under $25 / M tokens
Google: Gemini 4 Argon$8.00 per million blended tokens · ability 5.06 from 22 boards. No cheaper model scores higher.
Left out: 172 models
- Anthropic: Claude Opus 5.5 · $16.00 · 4.79 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Opus 5 · $20.00 · 4.52 — beaten by Google: Gemini 4 Argon
- OpenAI: GPT-6.1 Sol · $8.00 · 4.36 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Opus 4.7 · $20.00 · 4.31 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Opus 4.6 · $20.00 · 4.30 — beaten by Google: Gemini 4 Argon
- OpenAI: GPT-5.6 Sol · $16.00 · 4.29 — beaten by Google: Gemini 4 Argon
- Qwen: Qwen3.8 Max (0902) · $5.00 · 4.28 — beaten by Meta: Muse Spark 1.3
- MoonshotAI: Kimi K3 · $9.03 · 4.24 — beaten by Google: Gemini 4 Argon
- Meta: Muse Spark 1.2 · $3.50 · 4.07 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 3.1 Pro Preview · $9.50 · 4.06 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Opus 4.8 · $20.00 · 4.01 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Sonnet 5.5 · $8.00 · 3.99 — beaten by Google: Gemini 4 Argon
- Meta: Muse Spark 1.1 · $3.50 · 3.97 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-5.5 · $23.75 · 3.89 — beaten by Google: Gemini 4 Argon
- Google: Gemini 3.6 Flash · $3.00 · 3.86 — beaten by Google: Gemini 3.7 Flash
- Google: Gemini 3.5 Flash · $7.13 · 3.70 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-5.4 · $11.88 · 3.66 — beaten by Google: Gemini 4 Argon
- Z.ai: GLM 5.2 · $1.97 · 3.48 — beaten by Z.ai: GLM 5.3
- Qwen: Qwen3.7 Max · $3.69 · 3.44 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 3 Flash Preview · $2.38 · 3.36 — beaten by Z.ai: GLM 5.3
- Xiaomi: MiMo-V2.5-Pro · $0.761 · 3.36 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5.6 Terra · $9.50 · 3.36 — beaten by Google: Gemini 4 Argon
- SpaceXAI: Grok 4.5 · $5.00 · 3.31 — beaten by Meta: Muse Spark 1.3
- Anthropic: Claude Opus 4.5 · $20.00 · 3.29 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Sonnet 4.6 · $12.00 · 3.28 — beaten by Google: Gemini 4 Argon
- Anthropic: Claude Sonnet 5 · $8.00 · 3.24 — beaten by Google: Gemini 4 Argon
- DeepSeek: DeepSeek V4 Pro 0813 · $1.65 · 3.20 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Anthropic: Claude Haiku 5.5 · $0.400 · 3.19 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 5.1 · $2.52 · 3.18 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-6 Sol · $8.00 · 3.11 — beaten by Google: Gemini 4 Argon
- MoonshotAI: Kimi K2.6 · $2.72 · 3.11 — beaten by Z.ai: GLM 5.3
- Z.ai: GLM 5 Turbo · $3.30 · 3.04 — beaten by Google: Gemini 3.8 Flash
- DeepSeek: DeepSeek V4 Flash 0731 · $0.342 · 3.02 — beaten by Z.ai: GLM 5.3 Flash
- SpaceXAI: Grok 4.6 · $5.00 · 2.97 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-5.6 Luna · $0.950 · 2.88 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5.3-Codex · $10.94 · 2.85 — beaten by Google: Gemini 4 Argon
- Qwen: Qwen3.7 Plus · $1.04 · 2.81 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5.2-Codex · $10.94 · 2.80 — beaten by Google: Gemini 4 Argon
- Tencent: Hy3 · $0.429 · 2.79 — beaten by Z.ai: GLM 5.3 Flash
- Anthropic: Claude Sonnet 4.5 · $12.00 · 2.73 — beaten by Google: Gemini 4 Argon
- Z.ai: GLM 5 · $2.65 · 2.72 — beaten by Z.ai: GLM 5.3
- Google: Gemini 3.5 Flash Lite · $1.95 · 2.71 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.1 · $7.81 · 2.71 — beaten by Meta: Muse Spark 1.3
- MiniMax: MiniMax M2.1 · $0.975 · 2.70 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.6-Flash · $0.245 · 2.67 — beaten by Z.ai: GLM 5.3 Flash
- Upstage: Solar Pro 3 · $0.487 · 2.66 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- SpaceXAI: Grok 4.7 · $5.00 · 2.64 — beaten by Meta: Muse Spark 1.3
- Mistral: Mistral Large 4 · $1.74 · 2.64 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- OpenAI: GPT-6 Luna · $0.400 · 2.61 — beaten by Z.ai: GLM 5.3 Flash
- StepFun: Step 5 Preview · $2.28 · 2.59 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.1-Codex-Max · $7.81 · 2.55 — beaten by Meta: Muse Spark 1.3
- Qwen: Qwen3.5 397B A17B · $2.83 · 2.44 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.2 · $10.94 · 2.44 — beaten by Google: Gemini 4 Argon
- OpenAI: GPT-5.4 Mini · $3.56 · 2.42 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 2.5 Pro Preview 06-05 · $7.81 · 2.40 — beaten by Meta: Muse Spark 1.3
- Qwen: Qwen3.6 Plus · $1.54 · 2.40 — beaten by Xiaomi: MiMo-V2.6-Pro
- MoonshotAI: Kimi K2.7 Code · $2.68 · 2.37 — beaten by Z.ai: GLM 5.3
- Google: Gemini 2.5 Pro · $7.81 · 2.30 — beaten by Meta: Muse Spark 1.3
- MiniMax: MiniMax M3 · $0.975 · 2.28 — beaten by Xiaomi: MiMo-V2.6-Pro
- Google: Gemini 3.1 Flash Lite · $1.19 · 2.27 — beaten by Xiaomi: MiMo-V2.6-Pro
- SpaceXAI: Grok 4.3 · $2.19 · 2.24 — beaten by Z.ai: GLM 5.3
- Thinking Machines: Inkling · $3.29 · 2.24 — beaten by Google: Gemini 3.7 Flash
- MoonshotAI: Kimi K2.5 · $2.00 · 2.23 — beaten by Z.ai: GLM 5.3
- MoonshotAI: Kimi K2 Thinking · $2.02 · 2.22 — beaten by Z.ai: GLM 5.3
- Z.ai: GLM 4.7 · $1.62 · 2.20 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.8 27B · $1.71 · 2.17 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- StepFun: Step 3.7 Flash · $0.912 · 2.16 — beaten by Xiaomi: MiMo-V2.6-Pro
- Xiaomi: MiMo-V2.5 · $0.245 · 2.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.6 27B · $1.56 · 2.06 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5 · $7.81 · 2.06 — beaten by Meta: Muse Spark 1.3
- Z.ai: GLM 5V Turbo · $3.30 · 1.98 — beaten by Google: Gemini 3.8 Flash
- SpaceXAI: Grok Build 0.1 · $1.75 · 1.98 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- OpenAI: o3 · $6.50 · 1.94 — beaten by Meta: Muse Spark 1.3
- Google: Gemini 3.1 Flash Lite Preview · $1.19 · 1.92 — beaten by Xiaomi: MiMo-V2.6-Pro
- NVIDIA: Nemotron 3 Ultra · $1.95 · 1.89 — beaten by Z.ai: GLM 5.3
- Qwen: Qwen3.5 Plus 2026-04-20 · $1.43 · 1.88 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.6 35B A3B · $0.787 · 1.86 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5.1-Codex-Mini · $1.56 · 1.85 — beaten by Xiaomi: MiMo-V2.6-Pro
- DeepSeek: DeepSeek V3.2 · $0.385 · 1.83 — beaten by Z.ai: GLM 5.3 Flash
- Anthropic: Claude Sonnet 4 · $12.00 · 1.80 — beaten by Google: Gemini 4 Argon
- Mistral: Mistral Medium 3.5 · $6.00 · 1.80 — beaten by Meta: Muse Spark 1.3
- DeepSeek: DeepSeek V3.2 Exp · $0.375 · 1.73 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 235B A22B Instruct 2507 · $0.435 · 1.68 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.6V · $0.750 · 1.62 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3.5-122B-A10B · $1.63 · 1.58 — beaten by Xiaomi: MiMo-V2.6-Pro
- Tencent: Hy3 preview · $0.495 · 1.58 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- MiniMax: MiniMax M2.7 · $0.975 · 1.56 — beaten by Xiaomi: MiMo-V2.6-Pro
- Anthropic: Claude Haiku 4.5 · $4.00 · 1.51 — beaten by Meta: Muse Spark 1.3
- DeepSeek: DeepSeek V3.1 Terminus · $0.818 · 1.45 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 VL 235B A22B Instruct · $1.48 · 1.41 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-9B · $0.137 · 1.40 — beaten by Qwen: Qwen3.7 Flash
- Qwen: Qwen3.5-27B · $2.02 · 1.39 — beaten by Z.ai: GLM 5.3
- DeepSeek: R1 0528 · $1.74 · 1.37 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Z.ai: GLM 4.5 · $1.80 · 1.36 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- Thinking Machines: Inkling Small · $1.01 · 1.33 — beaten by Xiaomi: MiMo-V2.6-Pro
- Google: Gemma 4 26B A4B · $0.186 · 1.32 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemini 2.5 Flash · $1.95 · 1.32 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-5.4 Nano · $0.988 · 1.31 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.6 Flash · $0.891 · 1.29 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 235B A22B Thinking 2507 · $2.74 · 1.24 — beaten by Z.ai: GLM 5.3
- OpenAI: GPT-4.1 · $6.50 · 1.21 — beaten by Meta: Muse Spark 1.3
- DeepSeek: R1 · $2.05 · 1.10 — beaten by Z.ai: GLM 5.3
- OpenAI: o4 Mini High · $3.58 · 1.10 — beaten by Meta: Muse Spark 1.3
- Google: Gemma 4 31B · $0.350 · 1.07 — beaten by Z.ai: GLM 5.3 Flash
- StepFun: Step 3.5 Flash · $0.250 · 1.06 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.5-35B-A3B · $0.787 · 1.05 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-5 Mini · $1.56 · 1.04 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3.5-Flash · $0.211 · 1.00 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M2.5 · $0.780 · 0.97 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 Next 80B A3B Instruct · $0.848 · 0.95 — beaten by Xiaomi: MiMo-V2.6-Pro
- Upstage: Solar Pro 4 · $0.293 · 0.90 — beaten by Z.ai: GLM 5.3 Flash
- NVIDIA: Nemotron 3 Nano 30B A3B · $0.195 · 0.79 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 30B A3B Instruct 2507 · $0.247 · 0.73 — beaten by Z.ai: GLM 5.3 Flash
- MoonshotAI: Kimi K2 0905 · $2.02 · 0.70 — beaten by Z.ai: GLM 5.3
- OpenAI: o3 Mini High · $3.58 · 0.69 — beaten by Meta: Muse Spark 1.3
- MoonshotAI: Kimi K2 0711 · $1.87 · 0.68 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- NVIDIA: Nemotron 3 Super · $0.358 · 0.65 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: GPT-4.1 Mini · $1.30 · 0.64 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 4.5 Air · $0.670 · 0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Cohere: Command A · $8.13 · 0.60 — beaten by Google: Gemini 4 Argon
- Arcee AI: Trinity Large Thinking · $0.662 · 0.57 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 Next 80B A3B Thinking · $0.937 · 0.50 — beaten by Xiaomi: MiMo-V2.6-Pro
- Z.ai: GLM 4.7 Flash · $0.315 · 0.49 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemini 2.5 Flash Lite · $0.325 · 0.48 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M1 · $1.79 · 0.41 — beaten by DeepSeek: DeepSeek V4 Pro 0813
- MiniMax: MiniMax M2 · $0.829 · 0.41 — beaten by Xiaomi: MiMo-V2.6-Pro
- Meta: Llama 4 Maverick · $0.536 · 0.40 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: gpt-oss-120b · $0.487 · 0.35 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Qwen: Qwen3 14B · $0.210 · 0.34 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 27B · $0.282 · 0.29 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.5V · $1.50 · 0.26 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 32B · $0.230 · 0.26 — beaten by Z.ai: GLM 5.3 Flash
- Inception: Mercury 2 · $0.625 · 0.23 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-5 Nano · $0.313 · 0.22 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: GPT-4o (2024-11-20) · $8.13 · 0.06 — beaten by Google: Gemini 4 Argon
- DeepSeek: DeepSeek V3 · $0.747 · -0.01 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Mistral Medium 3 · $1.60 · -0.06 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 30B A3B · $0.405 · -0.08 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 12B · $0.125 · -0.11 — beaten by Qwen: Qwen3.7 Flash
- Qwen: Qwen3 VL 30B A3B Instruct · $0.487 · -0.25 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 4 Scout · $0.250 · -0.25 — beaten by Z.ai: GLM 5.3 Flash
- Mistral: Mistral Medium 3.1 · $1.60 · -0.30 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen-Plus · $0.650 · -0.33 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- OpenAI: GPT-4 Turbo · $25.00 · -0.36 — beaten by Google: Gemini 4 Argon
- OpenAI: GPT-4.1 Nano · $0.325 · -0.42 — beaten by Z.ai: GLM 5.3 Flash
- Mistral Large 2407 · $5.00 · -0.51 — beaten by Meta: Muse Spark 1.3
- Perplexity: Sonar Pro · $12.00 · -0.53 — beaten by Google: Gemini 4 Argon
- Qwen: Qwen3 Coder 30B A3B Instruct · $0.228 · -0.56 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.3 70B Instruct · $0.265 · -0.57 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 Coder Next · $0.630 · -0.59 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Devstral 2 2512 · $1.60 · -0.61 — beaten by Xiaomi: MiMo-V2.6-Pro
- OpenAI: GPT-4o-mini · $0.487 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Mistral: Mistral Small 3.1 24B · $0.504 · -0.63 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- IBM: Granite 4.2 8B · $0.203 · -0.78 — beaten by Z.ai: GLM 5.3 Flash
- Perplexity: Sonar · $1.00 · -0.83 — beaten by Xiaomi: MiMo-V2.6-Pro
- Qwen: Qwen3 VL 8B Instruct · $0.625 · -0.84 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Reka Flash 3 · $0.175 · -0.84 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 2 27B · $0.650 · -0.93 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 70B Instruct · $0.400 · -0.97 — beaten by Z.ai: GLM 5.3 Flash
- Google: Gemma 3 4B · $0.087 · -1.02 — beaten by OpenAI: gpt-oss-20b
- Cohere: Command R+ (08-2024) · $8.13 · -1.20 — beaten by Google: Gemini 4 Argon
- Microsoft: Phi 4 · $0.123 · -1.21 — beaten by Qwen: Qwen3.7 Flash
- Mistral: Mistral Small 3 · $0.072 · -1.23 — beaten by OpenAI: gpt-oss-20b
- Mistral: Saba · $0.500 · -1.48 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Nous: Hermes 3 70B Instruct · $0.700 · -1.55 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- WizardLM-2 8x22B · $0.620 · -1.67 — beaten by DeepSeek: DeepSeek V4.1 Flash 1017
- Meta: Llama 3.1 8B Instruct · $0.072 · -2.02 — beaten by OpenAI: gpt-oss-20b
- Mistral: Mixtral 8x22B Instruct · $5.00 · -2.02 — beaten by Meta: Muse Spark 1.3
- Mistral Large · $5.00 · -2.42 — beaten by Meta: Muse Spark 1.3
- OpenAI: GPT-3.5 Turbo · $1.25 · -2.54 — beaten by Xiaomi: MiMo-V2.6-Pro
- Meta: Llama 3.2 3B Instruct · $0.260 · -2.57 — beaten by Z.ai: GLM 5.3 Flash
- Meta: Llama 3.2 1B Instruct · $0.158 · -3.43 — beaten by Qwen: Qwen3.7 Flash
Best under 24 GB
gemma-4-31bNeeds 18.3 GB of memory · ability 2.75 from 22 boards. No smaller model scores higher.
Left out: 65 models
- Qwen: Qwen3.8 27B · 15.5 GB · 2.17 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.6 27B · 15.5 GB · 2.06 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.6 35B A3B · 20.1 GB · 1.86 — beaten by gemma-4-31b
- Qwen: Qwen3.5-27B · 15.5 GB · 1.39 — beaten by gemma-4-26b-a4b
- Google: Gemma 4 26B A4B · 14.4 GB · 1.32 — beaten by Qwen: Qwen3.5-9B
- Qwen: Qwen3 30B A3B Thinking 2507 · 17.1 GB · 1.05 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.5-35B-A3B · 20.1 GB · 1.05 — beaten by gemma-4-31b
- NVIDIA: Nemotron 3 Nano 30B A3B · 17.6 GB · 0.79 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 30B A3B Instruct 2507 · 17.1 GB · 0.73 — beaten by gemma-4-26b-a4b
- Z.ai: GLM 4.7 Flash · 17.4 GB · 0.49 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 14B · 8.3 GB · 0.34 — beaten by Qwen: Qwen3.5-9B
- Google: Gemma 3 27B · 15.3 GB · 0.29 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 32B · 18.3 GB · 0.26 — beaten by gemma-4-31b
- Qwen: Qwen3 VL 32B Instruct · 18.6 GB · 0.23 — beaten by gemma-4-31b
- nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4 · 10.0 GB · 0.18 — beaten by Qwen: Qwen3.5-9B
- qwq-32b · 18.3 GB · 0.12 — beaten by gemma-4-31b
- granite-4-2-30b · 16.4 GB · 0.03 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 30B A3B · 17.1 GB · -0.08 — beaten by gemma-4-26b-a4b
- Google: Gemma 3 12B · 6.8 GB · -0.11 — beaten by Qwen: Qwen3.5-9B
- olmo-3-1-32b-instruct · 18.0 GB · -0.16 — beaten by gemma-4-26b-a4b
- OpenAI: gpt-oss-20b · 11.7 GB · -0.18 — beaten by Qwen: Qwen3.5-9B
- Qwen: Qwen3 VL 30B A3B Instruct · 17.4 GB · -0.25 — beaten by gemma-4-26b-a4b
- olmo-3-32b-think · 18.0 GB · -0.38 — beaten by gemma-4-26b-a4b
- nvidia-nemotron-3-nano-30b-a3b-bf16 · 17.6 GB · -0.48 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 Coder 30B A3B Instruct · 17.1 GB · -0.56 — beaten by gemma-4-26b-a4b
- granite-4-1-8b · 4.9 GB · -0.57 — beaten by Qwen: Qwen3 8B
- olmo-3-1-32b-think · 18.0 GB · -0.59 — beaten by gemma-4-26b-a4b
- Mistral: Mistral Small 3.1 24B · 13.4 GB · -0.63 — beaten by Qwen: Qwen3.5-9B
- IBM: Granite 4.2 8B · 4.9 GB · -0.78 — beaten by Qwen: Qwen3 8B
- Qwen: Qwen3 VL 8B Instruct · 4.9 GB · -0.84 — beaten by Qwen: Qwen3 8B
- Reka Flash 3 · 11.7 GB · -0.84 — beaten by Qwen: Qwen3.5-9B
- Google: Gemma 2 27B · 15.2 GB · -0.93 — beaten by gemma-4-26b-a4b
- Google: Gemma 3 4B · 2.4 GB · -1.02 — beaten by granite-4-2-3b
- qwen2-5-coder-32b-instruct · 18.3 GB · -1.12 — beaten by gemma-4-31b
- Microsoft: Phi 4 · 8.2 GB · -1.21 — beaten by Qwen: Qwen3.5-9B
- Mistral: Mistral Small 3 · 13.2 GB · -1.23 — beaten by Qwen: Qwen3.5-9B
- gemma-2-9b-it · 5.2 GB · -1.40 — beaten by Qwen: Qwen3 8B
- olmo-2-0325-32b-instruct · 18.0 GB · -1.43 — beaten by gemma-4-26b-a4b
- llama-3-1-tulu-3-8b · 4.5 GB · -1.81 — beaten by granite-4-2-3b
- granite-3-1-8b-instruct · 4.6 GB · -1.91 — beaten by granite-4-2-3b
- yi-1-5-34b-chat · 19.2 GB · -1.97 — beaten by gemma-4-31b
- Meta: Llama 3.1 8B Instruct · 4.5 GB · -2.02 — beaten by granite-4-2-3b
- qwen1-5-32b-chat · 18.2 GB · -2.11 — beaten by gemma-4-26b-a4b
- phi-3-medium-4k-instruct · 7.8 GB · -2.17 — beaten by Qwen: Qwen3.5-9B
- gemma-2-2b-it · 1.5 GB · -2.27 — beaten by granite-3-1-2b-instruct
- qwen1-5-14b-chat · 7.9 GB · -2.30 — beaten by Qwen: Qwen3.5-9B
- granite-3-0-8b-instruct · 4.6 GB · -2.30 — beaten by granite-4-2-3b
- qwq-32b-preview · 18.3 GB · -2.36 — beaten by gemma-4-31b
- gemma-1-1-7b-it · 4.8 GB · -2.40 — beaten by Qwen: Qwen3 8B
- yi-34b-chat · 19.2 GB · -2.43 — beaten by gemma-4-31b
- phi-3-small-8k-instruct · 4.1 GB · -2.48 — beaten by granite-4-2-3b
- Meta: Llama 3.2 3B Instruct · 1.8 GB · -2.57 — beaten by granite-3-1-2b-instruct
- granite-3-0-2b-instruct · 1.5 GB · -2.66 — beaten by granite-3-1-2b-instruct
- qwen1-5-7b-chat · 4.3 GB · -2.90 — beaten by granite-4-2-3b
- qwen-14b-chat · 7.9 GB · -2.91 — beaten by Qwen: Qwen3.5-9B
- mistral-7b-instruct-v0-2 · 4.0 GB · -2.95 — beaten by granite-4-2-3b
- gemma-7b-it · 4.8 GB · -3.08 — beaten by Qwen: Qwen3 8B
- phi-3-mini-4k-instruct · 2.1 GB · -3.09 — beaten by granite-4-2-3b
- zephyr-7b-alpha · 4.0 GB · -3.15 — beaten by granite-4-2-3b
- zephyr-7b-beta · 4.0 GB · -3.17 — beaten by granite-4-2-3b
- codellama-34b-instruct · 18.9 GB · -3.20 — beaten by gemma-4-31b
- phi-3-mini-128k-instruct · 2.1 GB · -3.21 — beaten by granite-4-2-3b
- qwen1-5-4b-chat · 2.2 GB · -3.56 — beaten by granite-4-2-3b
- gemma-2b-it · 1.4 GB · -3.56 — beaten by gemma-1-1-2b-it
- olmo-7b-instruct · 3.8 GB · -3.76 — beaten by granite-4-2-3b
Best under 80 GB
gemma-4-31bThe same pick as "Best under 24 GB": nothing needing more than 18.3 GB and up to 80.0 GB scores higher.
Left out: 94 models
- Qwen: Qwen3.8 27B · 15.5 GB · 2.17 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.6 27B · 15.5 GB · 2.06 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.6 35B A3B · 20.1 GB · 1.86 — beaten by gemma-4-31b
- Qwen: Qwen3 Next 80B A3B Thinking · 45.4 GB · 1.78 — beaten by gemma-4-31b
- Z.ai: GLM 4.6V · 60.2 GB · 1.62 — beaten by gemma-4-31b
- Qwen: Qwen3.5-122B-A10B · 69.9 GB · 1.58 — beaten by gemma-4-31b
- nvidia-nemotron-3-5-super-vl-120b-a12b-nvfp4 · 36.8 GB · 1.57 — beaten by gemma-4-31b
- Qwen: Qwen3.5-27B · 15.5 GB · 1.39 — beaten by gemma-4-26b-a4b
- Google: Gemma 4 26B A4B · 14.4 GB · 1.32 — beaten by Qwen: Qwen3.5-9B
- Qwen: Qwen3 30B A3B Thinking 2507 · 17.1 GB · 1.05 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.5-35B-A3B · 20.1 GB · 1.05 — beaten by gemma-4-31b
- Qwen: Qwen3 Next 80B A3B Instruct · 45.4 GB · 0.95 — beaten by gemma-4-31b
- NVIDIA: Nemotron 3 Nano 30B A3B · 17.6 GB · 0.79 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 30B A3B Instruct 2507 · 17.1 GB · 0.73 — beaten by gemma-4-26b-a4b
- NVIDIA: Nemotron 3 Super · 69.1 GB · 0.65 — beaten by gemma-4-31b
- Z.ai: GLM 4.5 Air · 61.7 GB · 0.63 — beaten by gemma-4-31b
- Cohere: Command A · 62.1 GB · 0.60 — beaten by gemma-4-31b
- Z.ai: GLM 4.7 Flash · 17.4 GB · 0.49 — beaten by gemma-4-26b-a4b
- OpenAI: gpt-oss-120b · 65.3 GB · 0.35 — beaten by gemma-4-31b
- Qwen: Qwen3 14B · 8.3 GB · 0.34 — beaten by Qwen: Qwen3.5-9B
- Google: Gemma 3 27B · 15.3 GB · 0.29 — beaten by gemma-4-26b-a4b
- Z.ai: GLM 4.5V · 60.2 GB · 0.26 — beaten by gemma-4-31b
- Qwen: Qwen3 32B · 18.3 GB · 0.26 — beaten by gemma-4-31b
- Qwen: Qwen3 VL 32B Instruct · 18.6 GB · 0.23 — beaten by gemma-4-31b
- nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4 · 10.0 GB · 0.18 — beaten by Qwen: Qwen3.5-9B
- qwq-32b · 18.3 GB · 0.12 — beaten by gemma-4-31b
- granite-4-2-30b · 16.4 GB · 0.03 — beaten by gemma-4-26b-a4b
- ling-flash-2-0 · 57.5 GB · 0.02 — beaten by gemma-4-31b
- Qwen: Qwen3 30B A3B · 17.1 GB · -0.08 — beaten by gemma-4-26b-a4b
- Google: Gemma 3 12B · 6.8 GB · -0.11 — beaten by Qwen: Qwen3.5-9B
- ring-flash-2-0 · 57.5 GB · -0.11 — beaten by gemma-4-31b
- olmo-3-1-32b-instruct · 18.0 GB · -0.16 — beaten by gemma-4-26b-a4b
- OpenAI: gpt-oss-20b · 11.7 GB · -0.18 — beaten by Qwen: Qwen3.5-9B
- Qwen: Qwen3 VL 30B A3B Instruct · 17.4 GB · -0.25 — beaten by gemma-4-26b-a4b
- Meta: Llama 4 Scout · 60.7 GB · -0.25 — beaten by gemma-4-31b
- DeepSeek: R1 Distill Llama 70B · 39.4 GB · -0.27 — beaten by gemma-4-31b
- olmo-3-32b-think · 18.0 GB · -0.38 — beaten by gemma-4-26b-a4b
- nvidia-nemotron-3-nano-30b-a3b-bf16 · 17.6 GB · -0.48 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 Coder 30B A3B Instruct · 17.1 GB · -0.56 — beaten by gemma-4-26b-a4b
- Meta: Llama 3.3 70B Instruct · 39.4 GB · -0.57 — beaten by gemma-4-31b
- granite-4-1-8b · 4.9 GB · -0.57 — beaten by Qwen: Qwen3 8B
- olmo-3-1-32b-think · 18.0 GB · -0.59 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 Coder Next · 44.5 GB · -0.59 — beaten by gemma-4-31b
- Mistral: Devstral 2 2512 · 69.9 GB · -0.61 — beaten by gemma-4-31b
- Mistral: Mistral Small 3.1 24B · 13.4 GB · -0.63 — beaten by Qwen: Qwen3.5-9B
- qwen2-5-72b-instruct · 40.6 GB · -0.67 — beaten by gemma-4-31b
- IBM: Granite 4.2 8B · 4.9 GB · -0.78 — beaten by Qwen: Qwen3 8B
- Qwen: Qwen3 VL 8B Instruct · 4.9 GB · -0.84 — beaten by Qwen: Qwen3 8B
- Reka Flash 3 · 11.7 GB · -0.84 — beaten by Qwen: Qwen3.5-9B
- Google: Gemma 2 27B · 15.2 GB · -0.93 — beaten by gemma-4-26b-a4b
- Meta: Llama 3.1 70B Instruct · 39.4 GB · -0.97 — beaten by gemma-4-31b
- llama-3-1-tulu-3-70b · 39.4 GB · -1.01 — beaten by gemma-4-31b
- Google: Gemma 3 4B · 2.4 GB · -1.02 — beaten by granite-4-2-3b
- llama-3-1-nemotron-51b-instruct · 28.8 GB · -1.04 — beaten by gemma-4-31b
- qwen2-5-coder-32b-instruct · 18.3 GB · -1.12 — beaten by gemma-4-31b
- Microsoft: Phi 4 · 8.2 GB · -1.21 — beaten by Qwen: Qwen3.5-9B
- Mistral: Mistral Small 3 · 13.2 GB · -1.23 — beaten by Qwen: Qwen3.5-9B
- qwen2-72b-instruct · 40.6 GB · -1.34 — beaten by gemma-4-31b
- gemma-2-9b-it · 5.2 GB · -1.40 — beaten by Qwen: Qwen3 8B
- olmo-2-0325-32b-instruct · 18.0 GB · -1.43 — beaten by gemma-4-26b-a4b
- Nous: Hermes 3 70B Instruct · 39.4 GB · -1.55 — beaten by gemma-4-31b
- mixtral-8x22b-instruct-v0-1 · 78.6 GB · -1.69 — beaten by gemma-4-31b
- qwen1-5-110b-chat · 62.1 GB · -1.70 — beaten by gemma-4-31b
- qwen1-5-72b-chat · 40.4 GB · -1.75 — beaten by gemma-4-31b
- llama-3-1-tulu-3-8b · 4.5 GB · -1.81 — beaten by granite-4-2-3b
- granite-3-1-8b-instruct · 4.6 GB · -1.91 — beaten by granite-4-2-3b
- yi-1-5-34b-chat · 19.2 GB · -1.97 — beaten by gemma-4-31b
- Meta: Llama 3.1 8B Instruct · 4.5 GB · -2.02 — beaten by granite-4-2-3b
- Mistral: Mixtral 8x22B Instruct · 78.6 GB · -2.02 — beaten by gemma-4-31b
- qwen1-5-32b-chat · 18.2 GB · -2.11 — beaten by gemma-4-26b-a4b
- phi-3-medium-4k-instruct · 7.8 GB · -2.17 — beaten by Qwen: Qwen3.5-9B
- mixtral-8x7b-instruct-v0-1 · 26.1 GB · -2.20 — beaten by gemma-4-31b
- gemma-2-2b-it · 1.5 GB · -2.27 — beaten by granite-3-1-2b-instruct
- qwen1-5-14b-chat · 7.9 GB · -2.30 — beaten by Qwen: Qwen3.5-9B
- granite-3-0-8b-instruct · 4.6 GB · -2.30 — beaten by granite-4-2-3b
- qwq-32b-preview · 18.3 GB · -2.36 — beaten by gemma-4-31b
- gemma-1-1-7b-it · 4.8 GB · -2.40 — beaten by Qwen: Qwen3 8B
- yi-34b-chat · 19.2 GB · -2.43 — beaten by gemma-4-31b
- phi-3-small-8k-instruct · 4.1 GB · -2.48 — beaten by granite-4-2-3b
- Meta: Llama 3.2 3B Instruct · 1.8 GB · -2.57 — beaten by granite-3-1-2b-instruct
- granite-3-0-2b-instruct · 1.5 GB · -2.66 — beaten by granite-3-1-2b-instruct
- qwen1-5-7b-chat · 4.3 GB · -2.90 — beaten by granite-4-2-3b
- qwen-14b-chat · 7.9 GB · -2.91 — beaten by Qwen: Qwen3.5-9B
- mistral-7b-instruct-v0-2 · 4.0 GB · -2.95 — beaten by granite-4-2-3b
- gemma-7b-it · 4.8 GB · -3.08 — beaten by Qwen: Qwen3 8B
- phi-3-mini-4k-instruct · 2.1 GB · -3.09 — beaten by granite-4-2-3b
- zephyr-7b-alpha · 4.0 GB · -3.15 — beaten by granite-4-2-3b
- zephyr-7b-beta · 4.0 GB · -3.17 — beaten by granite-4-2-3b
- codellama-34b-instruct · 18.9 GB · -3.20 — beaten by gemma-4-31b
- phi-3-mini-128k-instruct · 2.1 GB · -3.21 — beaten by granite-4-2-3b
- codellama-70b-instruct · 38.5 GB · -3.26 — beaten by gemma-4-31b
- qwen1-5-4b-chat · 2.2 GB · -3.56 — beaten by granite-4-2-3b
- gemma-2b-it · 1.4 GB · -3.56 — beaten by gemma-1-1-2b-it
- olmo-7b-instruct · 3.8 GB · -3.76 — beaten by granite-4-2-3b
Best under 2 × 141 GB (two H200s)
Z.ai: GLM 5.3 FlashNeeds 180 GB of memory · ability 3.58 from 27 boards. No smaller model scores higher.
Left out: 126 models
- MiniMax: MiniMax M2.1 · 128 GB · 2.70 — beaten by gemma-4-31b
- Xiaomi: MiMo-V2.6-Flash · 174 GB · 2.67 — beaten by DeepSeek: DeepSeek V4 Flash 0731
- Qwen: Qwen3.5 397B A17B · 225 GB · 2.44 — beaten by Z.ai: GLM 5.3 Flash
- MiniMax: MiniMax M3 · 239 GB · 2.28 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.7 · 200 GB · 2.20 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3.8 27B · 15.5 GB · 2.17 — beaten by gemma-4-26b-a4b
- StepFun: Step 3.7 Flash · 113 GB · 2.16 — beaten by gemma-4-31b
- DeepSeek: DeepSeek V4 Flash 0731 · 170 GB · 2.08 — beaten by DeepSeek: DeepSeek V4 Flash 0731
- Xiaomi: MiMo-V2.5 · 174 GB · 2.06 — beaten by DeepSeek: DeepSeek V4 Flash 0731
- Qwen: Qwen3.6 27B · 15.5 GB · 2.06 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.6 35B A3B · 20.1 GB · 1.86 — beaten by gemma-4-31b
- Qwen: Qwen3 Next 80B A3B Thinking · 45.4 GB · 1.78 — beaten by gemma-4-31b
- nvidia-nemotron-3-ultra-550b-a55b-nvfp4 · 169 GB · 1.70 — beaten by Tencent: Hy3
- qwen3-235b-a22b-instruct-2507 · 131 GB · 1.68 — beaten by gemma-4-31b
- Z.ai: GLM 4.6V · 60.2 GB · 1.62 — beaten by gemma-4-31b
- Qwen: Qwen3.5-122B-A10B · 69.9 GB · 1.58 — beaten by gemma-4-31b
- Tencent: Hy3 preview · 167 GB · 1.58 — beaten by Tencent: Hy3
- nvidia-nemotron-3-5-super-vl-120b-a12b-nvfp4 · 36.8 GB · 1.57 — beaten by gemma-4-31b
- MiniMax: MiniMax M2.7 · 128 GB · 1.56 — beaten by gemma-4-31b
- Qwen: Qwen3 VL 235B A22B Instruct · 132 GB · 1.41 — beaten by gemma-4-31b
- Qwen: Qwen3.5-27B · 15.5 GB · 1.39 — beaten by gemma-4-26b-a4b
- Z.ai: GLM 4.5 · 200 GB · 1.36 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 VL 235B A22B Thinking · 132 GB · 1.36 — beaten by gemma-4-31b
- Thinking Machines: Inkling Small · 149 GB · 1.33 — beaten by gemma-4-31b
- Google: Gemma 4 26B A4B · 14.4 GB · 1.32 — beaten by Qwen: Qwen3.5-9B
- Qwen: Qwen3 235B A22B Thinking 2507 · 131 GB · 1.24 — beaten by gemma-4-31b
- mimo-v2-flash · 173 GB · 1.10 — beaten by DeepSeek: DeepSeek V4 Flash 0731
- Google: Gemma 4 31B · 17.5 GB · 1.07 — beaten by gemma-4-26b-a4b
- StepFun: Step 3.5 Flash · 111 GB · 1.06 — beaten by gemma-4-31b
- Qwen: Qwen3 30B A3B Thinking 2507 · 17.1 GB · 1.05 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3.5-35B-A3B · 20.1 GB · 1.05 — beaten by gemma-4-31b
- MiniMax: MiniMax M2.5 · 128 GB · 0.97 — beaten by gemma-4-31b
- Qwen: Qwen3 Next 80B A3B Instruct · 45.4 GB · 0.95 — beaten by gemma-4-31b
- qwen3-coder-480b-a35b-instruct · 268 GB · 0.83 — beaten by Z.ai: GLM 5.3 Flash
- NVIDIA: Nemotron 3 Nano 30B A3B · 17.6 GB · 0.79 — beaten by gemma-4-26b-a4b
- trinity-large-preview · 223 GB · 0.76 — beaten by Z.ai: GLM 5.3 Flash
- Qwen: Qwen3 30B A3B Instruct 2507 · 17.1 GB · 0.73 — beaten by gemma-4-26b-a4b
- NVIDIA: Nemotron 3 Super · 69.1 GB · 0.65 — beaten by gemma-4-31b
- Z.ai: GLM 4.5 Air · 61.7 GB · 0.63 — beaten by gemma-4-31b
- Cohere: Command A · 62.1 GB · 0.60 — beaten by gemma-4-31b
- qwen3-235b-a22b · 131 GB · 0.58 — beaten by gemma-4-31b
- Arcee AI: Trinity Large Thinking · 223 GB · 0.57 — beaten by Z.ai: GLM 5.3 Flash
- Z.ai: GLM 4.7 Flash · 17.4 GB · 0.49 — beaten by gemma-4-26b-a4b
- MiniMax: MiniMax M2 · 128 GB · 0.41 — beaten by gemma-4-31b
- Meta: Llama 4 Maverick · 224 GB · 0.40 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: gpt-oss-120b · 65.3 GB · 0.35 — beaten by gemma-4-31b
- Qwen: Qwen3 14B · 8.3 GB · 0.34 — beaten by Qwen: Qwen3.5-9B
- llama-3-1-nemotron-ultra-253b-v1 · 142 GB · 0.30 — beaten by gemma-4-31b
- Google: Gemma 3 27B · 15.3 GB · 0.29 — beaten by gemma-4-26b-a4b
- Z.ai: GLM 4.5V · 60.2 GB · 0.26 — beaten by gemma-4-31b
- Qwen: Qwen3 32B · 18.3 GB · 0.26 — beaten by gemma-4-31b
- Qwen: Qwen3 VL 32B Instruct · 18.6 GB · 0.23 — beaten by gemma-4-31b
- nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4 · 10.0 GB · 0.18 — beaten by Qwen: Qwen3.5-9B
- qwq-32b · 18.3 GB · 0.12 — beaten by gemma-4-31b
- granite-4-2-30b · 16.4 GB · 0.03 — beaten by gemma-4-26b-a4b
- ling-flash-2-0 · 57.5 GB · 0.02 — beaten by gemma-4-31b
- Qwen: Qwen3 30B A3B · 17.1 GB · -0.08 — beaten by gemma-4-26b-a4b
- Google: Gemma 3 12B · 6.8 GB · -0.11 — beaten by Qwen: Qwen3.5-9B
- ring-flash-2-0 · 57.5 GB · -0.11 — beaten by gemma-4-31b
- olmo-3-1-32b-instruct · 18.0 GB · -0.16 — beaten by gemma-4-26b-a4b
- llama-3-1-405b-instruct-fp8 · 227 GB · -0.18 — beaten by Z.ai: GLM 5.3 Flash
- OpenAI: gpt-oss-20b · 11.7 GB · -0.18 — beaten by Qwen: Qwen3.5-9B
- deepseek-v2-5-1210 · 132 GB · -0.20 — beaten by gemma-4-31b
- Qwen: Qwen3 VL 30B A3B Instruct · 17.4 GB · -0.25 — beaten by gemma-4-26b-a4b
- Meta: Llama 4 Scout · 60.7 GB · -0.25 — beaten by gemma-4-31b
- DeepSeek: R1 Distill Llama 70B · 39.4 GB · -0.27 — beaten by gemma-4-31b
- olmo-3-32b-think · 18.0 GB · -0.38 — beaten by gemma-4-26b-a4b
- nvidia-nemotron-3-nano-30b-a3b-bf16 · 17.6 GB · -0.48 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 Coder 30B A3B Instruct · 17.1 GB · -0.56 — beaten by gemma-4-26b-a4b
- Meta: Llama 3.3 70B Instruct · 39.4 GB · -0.57 — beaten by gemma-4-31b
- granite-4-1-8b · 4.9 GB · -0.57 — beaten by Qwen: Qwen3 8B
- olmo-3-1-32b-think · 18.0 GB · -0.59 — beaten by gemma-4-26b-a4b
- Qwen: Qwen3 Coder Next · 44.5 GB · -0.59 — beaten by gemma-4-31b
- Mistral: Devstral 2 2512 · 69.9 GB · -0.61 — beaten by gemma-4-31b
- Mistral: Mistral Small 3.1 24B · 13.4 GB · -0.63 — beaten by Qwen: Qwen3.5-9B
- deepseek-v2-5 · 132 GB · -0.65 — beaten by gemma-4-31b
- qwen2-5-72b-instruct · 40.6 GB · -0.67 — beaten by gemma-4-31b
- IBM: Granite 4.2 8B · 4.9 GB · -0.78 — beaten by Qwen: Qwen3 8B
- Qwen: Qwen3 VL 8B Instruct · 4.9 GB · -0.84 — beaten by Qwen: Qwen3 8B
- Reka Flash 3 · 11.7 GB · -0.84 — beaten by Qwen: Qwen3.5-9B
- Google: Gemma 2 27B · 15.2 GB · -0.93 — beaten by gemma-4-26b-a4b
- Meta: Llama 3.1 70B Instruct · 39.4 GB · -0.97 — beaten by gemma-4-31b
- llama-3-1-tulu-3-70b · 39.4 GB · -1.01 — beaten by gemma-4-31b
- Google: Gemma 3 4B · 2.4 GB · -1.02 — beaten by granite-4-2-3b
- llama-3-1-nemotron-51b-instruct · 28.8 GB · -1.04 — beaten by gemma-4-31b
- qwen2-5-coder-32b-instruct · 18.3 GB · -1.12 — beaten by gemma-4-31b
- Microsoft: Phi 4 · 8.2 GB · -1.21 — beaten by Qwen: Qwen3.5-9B
- Mistral: Mistral Small 3 · 13.2 GB · -1.23 — beaten by Qwen: Qwen3.5-9B
- qwen2-72b-instruct · 40.6 GB · -1.34 — beaten by gemma-4-31b
- gemma-2-9b-it · 5.2 GB · -1.40 — beaten by Qwen: Qwen3 8B
- olmo-2-0325-32b-instruct · 18.0 GB · -1.43 — beaten by gemma-4-26b-a4b
- Nous: Hermes 3 70B Instruct · 39.4 GB · -1.55 — beaten by gemma-4-31b
- mixtral-8x22b-instruct-v0-1 · 78.6 GB · -1.69 — beaten by gemma-4-31b
- qwen1-5-110b-chat · 62.1 GB · -1.70 — beaten by gemma-4-31b
- qwen1-5-72b-chat · 40.4 GB · -1.75 — beaten by gemma-4-31b
- llama-3-1-tulu-3-8b · 4.5 GB · -1.81 — beaten by granite-4-2-3b
- granite-3-1-8b-instruct · 4.6 GB · -1.91 — beaten by granite-4-2-3b
- yi-1-5-34b-chat · 19.2 GB · -1.97 — beaten by gemma-4-31b
- Meta: Llama 3.1 8B Instruct · 4.5 GB · -2.02 — beaten by granite-4-2-3b
- Mistral: Mixtral 8x22B Instruct · 78.6 GB · -2.02 — beaten by gemma-4-31b
- qwen1-5-32b-chat · 18.2 GB · -2.11 — beaten by gemma-4-26b-a4b
- phi-3-medium-4k-instruct · 7.8 GB · -2.17 — beaten by Qwen: Qwen3.5-9B
- mixtral-8x7b-instruct-v0-1 · 26.1 GB · -2.20 — beaten by gemma-4-31b
- gemma-2-2b-it · 1.5 GB · -2.27 — beaten by granite-3-1-2b-instruct
- qwen1-5-14b-chat · 7.9 GB · -2.30 — beaten by Qwen: Qwen3.5-9B
- granite-3-0-8b-instruct · 4.6 GB · -2.30 — beaten by granite-4-2-3b
- qwq-32b-preview · 18.3 GB · -2.36 — beaten by gemma-4-31b
- gemma-1-1-7b-it · 4.8 GB · -2.40 — beaten by Qwen: Qwen3 8B
- snowflake-arctic-instruct · 267 GB · -2.41 — beaten by Z.ai: GLM 5.3 Flash
- yi-34b-chat · 19.2 GB · -2.43 — beaten by gemma-4-31b
- phi-3-small-8k-instruct · 4.1 GB · -2.48 — beaten by granite-4-2-3b
- Meta: Llama 3.2 3B Instruct · 1.8 GB · -2.57 — beaten by granite-3-1-2b-instruct
- granite-3-0-2b-instruct · 1.5 GB · -2.66 — beaten by granite-3-1-2b-instruct
- qwen1-5-7b-chat · 4.3 GB · -2.90 — beaten by granite-4-2-3b
- qwen-14b-chat · 7.9 GB · -2.91 — beaten by Qwen: Qwen3.5-9B
- mistral-7b-instruct-v0-2 · 4.0 GB · -2.95 — beaten by granite-4-2-3b
- gemma-7b-it · 4.8 GB · -3.08 — beaten by Qwen: Qwen3 8B
- phi-3-mini-4k-instruct · 2.1 GB · -3.09 — beaten by granite-4-2-3b
- zephyr-7b-alpha · 4.0 GB · -3.15 — beaten by granite-4-2-3b
- zephyr-7b-beta · 4.0 GB · -3.17 — beaten by granite-4-2-3b
- codellama-34b-instruct · 18.9 GB · -3.20 — beaten by gemma-4-31b
- phi-3-mini-128k-instruct · 2.1 GB · -3.21 — beaten by granite-4-2-3b
- codellama-70b-instruct · 38.5 GB · -3.26 — beaten by gemma-4-31b
- qwen1-5-4b-chat · 2.2 GB · -3.56 — beaten by granite-4-2-3b
- gemma-2b-it · 1.4 GB · -3.56 — beaten by gemma-1-1-2b-it
- olmo-7b-instruct · 3.8 GB · -3.76 — beaten by granite-4-2-3b
Cost frontier — models you call
Blended price per million tokens, assuming 3 output tokens for every input token. That mix is an assumption, not a measurement: a summarisation workload inverts it and would reorder the cheap end. Prices come from the OpenRouter catalogue, so they are one vendor's, and prices we could not re-confirm within four days are withheld rather than quoted.
| Model | USD / 1M tokens | vs cheapest | Ability | Boards |
|---|---|---|---|---|
| OpenAI: gpt-oss-20b | $0.072 | 1× | -0.18 | 26 |
| Qwen: Qwen3.7 Flash | $0.105 | 1× | 1.86 | 2 |
| Z.ai: GLM 5.3 Flash | $0.162 | 2× | 3.58 | 27 |
| DeepSeek: DeepSeek V4.1 Flash 1017 | $0.487 | 7× | 3.66 | 28 |
| Xiaomi: MiMo-V2.6-Pro | $0.761 | 11× | 3.73 | 19 |
| DeepSeek: DeepSeek V4 Pro 0813 | $1.65 | 23× | 3.75 | 4 |
| Z.ai: GLM 5.3 | $1.89 | 26× | 3.81 | 26 |
| Google: Gemini 3.7 Flash | $3.00 | 42× | 4.30 | 28 |
| Google: Gemini 3.8 Flash | $3.30 | 46× | 4.39 | 29 |
| Meta: Muse Spark 1.3 | $3.50 | 49× | 4.44 | 26 |
| Google: Gemini 4 Argon | $8.00 | 111× | 5.06 | 22 |
11 on the frontier; 194 models are left out because something else is both more capable and costs no more.
Hardware frontier — models you host
Memory needed to load the weights at INT4, plus a 1.2× runtime allowance. A floor, not a recommendation: it excludes long-context KV cache and multi-GPU overhead, both of which push the real figure up. Open-weight models only, and only those publishing an exact parameter count.
| Model | GPU memory | Ability | Boards |
|---|---|---|---|
| Meta: Llama 3.2 1B Instruct | 0.7 GB | -3.43 | 19 |
| gemma-1-1-2b-it | 1.4 GB | -3.24 | 17 |
| granite-3-1-2b-instruct | 1.4 GB | -2.25 | 16 |
| granite-4-2-3b | 2.0 GB | -0.89 | 15 |
| Qwen: Qwen3 8B | 4.6 GB | -0.14 | 6 |
| Qwen: Qwen3.5-9B | 5.4 GB | 1.40 | 5 |
| gemma-4-26b-a4b | 14.8 GB | 2.30 | 21 |
| gemma-4-31b | 18.3 GB | 2.75 | 22 |
| Tencent: Hy3 | 167 GB | 2.79 | 23 |
| DeepSeek: DeepSeek V4 Flash 0731 | 170 GB | 3.02 | 5 |
| Z.ai: GLM 5.3 Flash | 180 GB | 3.58 | 27 |
| Z.ai: GLM 5.3 | 421 GB | 3.81 | 26 |
| MoonshotAI: Kimi K3 | 1553 GB | 4.24 | 28 |
13 on the frontier; 160 models are left out because something else is both more capable and no larger.
Coverage
205 models could be placed on the cost frontier and 173 on the hardware frontier. Left out: 10 with a price too old to confirm, 24 priced models measured on too few boards to place, and 136 open-weight models publishing no parameter count.