Pay per token, nothing else.
Every rate below is per million tokens. Hosted inference does not require an hourly GPU rental or reserved capacity.
- Billed per 1M tokens
- Cached input rates
- Streaming supported
- Usage-based billing
Rates per 1M tokens
One rate card for every request: streaming or not, tools or not. Cached input is the rate for prompt prefixes served from cache.
| Model | Context | Input | Cached input | Output |
|---|---|---|---|---|
| GPT-OSS 120Bopenrelay/gpt-oss-120b | 128K | $0.15 | $0.015 | $0.60 |
| GPT-OSS 20Bopenrelay/gpt-oss-20b | 128K | $0.05 | $0.005 | $0.20 |
| Gemma 4 31B NVFP4 32Kopenrelay/gemma-4-31b-nvfp4-32k | 32K | $0.12 | n/a | $0.35 |
| Gemma 4 31Bopenrelay/gemma-4-31b | 32K | $0.99 | n/a | $1.49 |
| GLM 5.2openrelay/glm-5.2 | 1M | $1.82 | $0.338 | $5.72 |
| DeepSeek-OCR 2openrelay/deepseek-ocr-2 | 8K | $0.039 | n/a | $0.039 |
| DeepSeek V3.1 Terminusopenrelay/deepseek-v3.1-terminus | 64K | $0.27 | n/a | $1.00 |
| Embeddings | Max input | Input | Notes |
|---|---|---|---|
| BGE-M3openrelay/bge-m3 | 8K | $0.013 | Input-only metering via /v1/embeddings |
All prices in USD per 1M tokens. You pay only for the tokens you use, metered per request. More models are on the way; browse the inference catalog or request a model.
The models behind the rates
GPT-OSS 120B
OpenAI
- Context
- 128K
- Parameters
- 120B
- Input / 1M
- $0.15
- Output / 1M
- $0.60
OpenAI's 120B open-weight model, with reasoning, tool calling, and structured output support.
GPT-OSS 20B
OpenAI
- Context
- 128K
- Parameters
- 20B
- Input / 1M
- $0.05
- Output / 1M
- $0.20
The smaller GPT-OSS variant, with tool calling and structured output support at lower token rates.
Gemma 4 31B NVFP4 32K
- Context
- 32K
- Parameters
- 31B
- Input / 1M
- $0.12
- Output / 1M
- $0.35
An NVFP4-quantized Gemma 4 31B deployment with a 32K-token context window and lower token rates.
Gemma 4 31B
- Context
- 32K
- Parameters
- 31B
- Input / 1M
- $0.99
- Output / 1M
- $1.49
Google's dense 31B model with text and image input, reasoning, tool calling, and a 32K-token context window.
GLM 5.2
Zhipu
- Context
- 1M
- Parameters
- 744B MoE
- Input / 1M
- $1.82
- Output / 1M
- $5.72
Zhipu's GLM model for reasoning, coding, and tool use in English and Chinese, with a 1M-token context window.
DeepSeek-OCR 2
DeepSeek
- Context
- 8K
- Parameters
- -
- Input / 1M
- $0.039
- Output / 1M
- $0.039
DeepSeek's second-generation OCR model. Reads document images (scans, receipts, screenshots, tables) and returns structured markdown that preserves headings, tables, and layout.
DeepSeek V3.1 Terminus
DeepSeek
- Context
- 64K
- Parameters
- 685B MoE
- Input / 1M
- $0.27
- Output / 1M
- $1.00
DeepSeek's V3.1 Terminus release: a sparse mixture-of-experts model with a hybrid thinking mode, tuned for multi-step agentic work and consistent tool calling.
How billing works
Metered per request, itemized per token
Cached input
Prompt prefixes served from cache bill at the cached input rate: 10x cheaper than fresh input on GPT-OSS models. It applies automatically, with no configuration. How to structure prompts so they hit the cache.
Request features
Supported request features do not add a separate fee. Each request is metered on input and output tokens at the model's rates.
Batch at 50% off
Large offline jobs can run on the Batch Inference API at half the per-token rate, with results delivered within 24 hours.
Send your first request today.
Grab an API key, point your OpenAI SDK at OpenRelay, and pay only for what you run.