Provider comparison
Open-weight inference provider pricing comparison
Published 2026-07-11 · 11 providers · 2092 chat models
Open-weight inference providers like Groq, Fireworks AI, DeepInfra, and Together AI offer low-cost access to models like Llama, Qwen, and DeepSeek. With near-1:1 output/input ratios and per-token prices up to 10× cheaper than frontier APIs, they are an attractive option for production workloads. This guide compares pricing across 11 providers using current catalog data.
Provider pricing snapshot
The table below shows the cheapest chat model from each provider, plus average pricing across all their chat models. Open-weight providers tend to charge near-1:1 output/input ratios, unlike frontier providers where output costs 3-5× more per token.
| Provider | Cheapest model | Input / 1M | Output / 1M | Ratio | Avg input | Avg output | Models |
|---|---|---|---|---|---|---|---|
| Fireworks AI | starcoder2-3b | $0.1000 / 1M tokens | $0.1000 / 1M tokens | 1.0× | $0.5053 / 1M tokens | $0.7093 / 1M tokens | 243 |
| DeepInfra | Llama-3.2-3B-Instruct | $0.0200 / 1M tokens | $0.0200 / 1M tokens | 1.0× | $0.5983 / 1M tokens | $2.5647 / 1M tokens | 67 |
| Novita AI | paddleocr-vl | $0.0200 / 1M tokens | $0.0200 / 1M tokens | 1.0× | $0.3230 / 1M tokens | $1.0021 / 1M tokens | 80 |
| Together AI | together-ai-up-to-4b | $0.1000 / 1M tokens | $0.1000 / 1M tokens | 1.0× | $0.7466 / 1M tokens | $1.7756 / 1M tokens | 32 |
| Groq | gemma-7b-it | $0.0500 / 1M tokens | $0.0800 / 1M tokens | 1.6× | $0.2536 / 1M tokens | $0.6255 / 1M tokens | 11 |
| Cerebras | llama3.1-8b | $0.1000 / 1M tokens | $0.1000 / 1M tokens | 1.0× | $0.9714 / 1M tokens | $1.2786 / 1M tokens | 7 |
| Nebius AI | Qwen2.5-Coder-7B | $0.0100 / 1M tokens | $0.0300 / 1M tokens | 3.0× | $0.2644 / 1M tokens | $0.8004 / 1M tokens | 27 |
| Nscale | Qwen2.5-Coder-7B-Instruct | $0.0100 / 1M tokens | $0.0300 / 1M tokens | 3.0× | $0.1493 / 1M tokens | $0.1779 / 1M tokens | 14 |
| Lambda | llama3.2-3b-instruct | $0.0150 / 1M tokens | $0.0250 / 1M tokens | 1.7× | $0.1943 / 1M tokens | $0.3085 / 1M tokens | 20 |
| Hyperbolic | DeepSeek-V3 | $0.2000 / 1M tokens | $0.2000 / 1M tokens | 1.0× | $0.4081 / 1M tokens | $0.5094 / 1M tokens | 16 |
| Anyscale | Mixtral-8x7B-Instruct-v0.1 | $0.1500 / 1M tokens | $0.1500 / 1M tokens | 1.0× | $0.5042 / 1M tokens | $0.5042 / 1M tokens | 12 |
Cheapest models per workload
Different providers specialize in different model sizes. Here are the cheapest options for common model size ranges:
| Size range | Cheapest provider | Model | Combined price / 1M |
|---|
Key findings
- 10 of 11 providers offer at least one model with a near-1:1 output/input ratio — meaning output tokens cost about the same as input tokens.
- The cheapest entry point across all providers is DeepInfra, with Llama-3.2-3B-Instruct at $0.0400 / 1M tokens/1M combined.
- Average output/input ratio: 2.3× across open-weight providers vs 3.64× across all providers (see the output vs input pricing guide).
- Small models (≤8B) are available across nearly all providers, while large models (≥100B) are concentrated on Fireworks AI, Together AI, and DeepInfra.
Compare the cheapest models side by side
Open the compare page with all 11 cheapest models to see context windows, max tokens, and feature support alongside pricing.