Capability × cost × speed
Frontier
Capability = Modelyst cross-benchmark percentile · price & speed = Artificial Analysis medians · verified Jun 15, 2026 · methodology
League tables — best atAgentsCodeInstruction FollowingKnowledge & QALanguage & InstructionLong ContextMathReasoningVision & Multimodal
The efficient frontier — capability vs price
140 models · hover a point · click to openon the frontier — nothing is both cheaper and bettereverything else
Up and to the left wins: the gold staircase is the set of models nothing beats on both price and capability at once.
The race — capability over time
gold = set a new record on releaseThroughput value — output speed vs price
114 modelsOn the frontier · 13
cheapest → most capable| Model | Cap | $/1M | tok/s | Weights |
|---|---|---|---|---|
Qwen3.5 0.8B Alibaba | 19.3 | $0.02 | 20 | open |
Gemma 3n E4B Instruct Google | 32.2 | $0.025 | 40 | open |
Qwen3.5 2B Alibaba | 36.6 | $0.04 | 21 | open |
Qwen3.5 4B Alibaba | 65.3 | $0.06 | 23 | open |
gpt-oss-20b openai | 72.5 | $0.088 | 252 | open |
NVIDIA Nemotron 3 Nano 30B A3B NVIDIA | 73.1 | $0.096 | 85 | open |
Qwen3.5 9B Alibaba | 73.6 | $0.113 | 65 | open |
Gemma 4 12B Google | 73.8 | $0.15 | 161 | open |
DeepSeek V4 Flash DeepSeek | 91.4 | $0.175 | 114 | open |
MiniMax-M3 MiniMax | 94.6 | $0.525 | 59 | open |
DeepSeek V4 Pro DeepSeek | 95.0 | $0.544 | 89 | open |
Kimi K2.6 Kimi | 96.0 | $1.71 | 46 | open |
Qwen3.7 Max Alibaba | 96.1 | $3.75 | 199 | open |
Every point links to the model's page — scores with sources, latency, and the research behind it. Compare any of them head-to-head on Compare.
Workload cost calculator
Tokens per task × tasks per day → what each model actually costs to run, and how long a task takes.
Presets
| Model | Cap | $ / task | $ / day | time / task |
|---|---|---|---|---|
| Qwen3.5 0.8BAlibaba | 19 | $0 | $0.175 | 15s |
| Gemma 3n E4B InstructGoogle | 32 | $0.0001 | $0.26 | 8.0s |
| Qwen3.5 2BAlibaba | 37 | $0.0001 | $0.35 | 15s |
| Sarvam 30BSarvam | 39 | $0.0001 | $0.425 | 2.4s |
| LFM2 24B A2BLiquid AI | 29 | $0.0001 | $0.48 | 2.5s |
| Gemma 3 4B InstructGoogle | 30 | $0.0001 | $0.52 | — |
| Qwen3.5 4BAlibaba | 65 | $0.0001 | $0.525 | 14s |
| Nova MicroAmazon | 34 | $0.0001 | $0.56 | 1.6s |
| Llama 3.2 Instruct 1BMeta | 16 | $0.0001 | $0.575 | 4.0s |
| HyperNova 60B 2605Multiverse Computing | 70 | $0.0001 | $0.61 | 1.4s |
| NVIDIA-Nemotron-Nano-9B-v2nvidia | 49 | $0.0001 | $0.64 | 19s |
| Granite 4.1 8BIBM | 35 | $0.0001 | $0.65 | 2.9s |
Price arithmetic from live per-token prices (Artificial Analysis medians) — not a measured task benchmark. Time per task = latency + output tokens ÷ speed; ignores caching, rate limits and retries. For reasoning models, latency includes thinking time.