LLM pricing and speed comparison
22 models have both a published API price and a composite index. The cheapest is gpt-oss-120b at $0.060 per million tokens (blended). The fastest measured is Reflection Beam 1.0 at about 371 tokens per second. 7 models sit on the value line, meaning each beats every cheaper model on the index. Open-weight models can also run on a rented GPU at a flat hourly rate, which this page converts to a per-token cost for one user.
Price vs quality
Each dot is a model with an API price and a composite index. The dashed line joins the best value at every budget. Blended price = (3 × input + output) ÷ 4.
- OpenLLM Buddy catalog
- Open weights
- Closed API
Every model, price and speed
API figures are published prices and measured speeds, linked on each model page. Rows marked ≈ are estimates.
| Model | Index | API in / out per 1M | API speed | One-GPU speed | OpenLLM Buddy | Self-hosted per 1M out |
|---|---|---|---|---|---|---|
| gpt-oss-120b | 57.3 | $0.030 / $0.150 | 212 tok/s | — | — | — |
| Gemma 4 31B | 79.2 | $0.100 / $0.340 | 35 tok/s | ≈ 25–34 tok/s · RTX 4090 4-bit | $1.10/hr | ≈ $10.20 |
| Claude Haiku 5.5 | 72.4 | $0.100 / $0.500 | 236 tok/s | — | — | — |
| GLM 5.3 Flash | — | $0.150 / $0.500 | 75 tok/s | — | $25.42/hr | — |
| Gemma 4 26B A4B | 72.9 | $0.150 / $0.600 | 113 tok/s | ≈ 133–160 tok/s · RTX 4090 4-bit | see model | — |
| MiniMax M3.1 Flash | — | $0.150 / $0.600 | 136 tok/s | — | — | — |
| Reflection Beam 1.0 | 87.0 | $0.250 / $1.00 | 371 tok/s | — | — | — |
| DeepSeek V4.1 Flash | 79.8 | $0.300 / $1.20 | 217 tok/s | — | $25.42/hr | — |
| MiMo V2.6 Pro | — | $0.435 / $0.870 | 47 tok/s | — | $50.81/hr | — |
| Nemotron 3 Ultra | 74.4 | $0.500 / $2.20 | 161 tok/s | — | — | — |
| Qwen3.8 27B | 83.2 | $0.425 / $2.55 | 41 tok/s | ≈ 52–69 tok/s · H100 80GB FP8 | $4.82/hr | ≈ $22.07 |
| Qwen3.6 27B | 85.8 | $0.450 / $2.70 | 57 tok/s | ≈ 28–39 tok/s · RTX 4090 4-bit | see model | — |
| Mistral Large 4 | — | $0.700 / $2.10 | 112 tok/s | — | — | — |
| Tencent Hy4 Preview | 89.9 | $0.834 / $2.50 | 52 tok/s | — | — | — |
| Step 5 Preview | 78.7 | $1.00 / $2.70 | 146 tok/s | — | — | — |
| Gemini 3.8 Flash | 70.8 | $0.750 / $3.75 | 135 tok/s | — | — | — |
| DeepSeek V4 Pro | 91.0 | $1.32 / $3.96 | 94 tok/s | — | — | — |
| Muse Spark Open | 88.0 | $1.25 / $4.25 | 196 tok/s | — | — | — |
| GLM 5.3 | 82.7 | $1.40 / $4.40 | 88 tok/s | — | — | — |
| Grok 4.7 | — | $2.00 / $6.00 | 84 tok/s | — | — | — |
| Qwen3.8 2.4T-A95B | — | $2.00 / $6.00 | 41 tok/s | — | — | — |
| Kimi K3.1 Pro | 94.4 | $2.00 / $10.00 | 31 tok/s | — | — | — |
| Claude Sonnet 5.5 | 91.9 | $2.00 / $10.00 | 138 tok/s | — | — | — |
| GPT-6.1 Sol | 86.5 | $2.00 / $10.00 | 91 tok/s | — | — | — |
| Gemini 4 Argon | — | $2.00 / $10.00 | 83 tok/s | — | — | — |
| Kimi K3 | 90.6 | $3.00 / $15.00 | 66 tok/s | — | $38.33/hr | — |
| Claude Opus 5.5 | 98.4 | $4.00 / $20.00 | 96 tok/s | — | — | — |
| GPT-6 Astra | 91.4 | $10.00 / $50.00 | 74 tok/s | — | — | — |
| Claude Fable 5.1 | 91.2 | $10.00 / $50.00 | 71 tok/s | — | — | — |
How speed is estimated
Estimate for one user on one GPU: active weights read per token ÷ memory bandwidth at 50–70% efficiency, plus 3 ms overhead per token. Real speed varies with engine, context length and quantization; serving several users at once raises total throughput.
API vs self-hosting
APIs bill every token, so cost grows with use. A rented GPU costs the same per hour however much you send it, so the per-token figure above is the worst case: one user, one request at a time. Batch requests or share the GPU across a team and the cost per token drops quickly, and prompts never leave your server.
Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works