AI inference providers, priced on one model
The cheapest direct API for gpt-oss-120b is CoreWeave at $0.03 input and $0.17 output per 1M tokens. OpenRouter routes it from $0.03 / $0.15, plus a 5.5% credit fee. Cerebras is the fastest at about 3,000 tokens/s, 6× Groq’s 500 (both vendor-stated). Prices as of .
Low: OpenRouter $0.15 High: Parasail, Cerebras, Cloudflare Workers AI $0.75
Output speed, gpt-oss-120b tokens/s, vendor-stated
| Hardware | OpenAI API | Serverless | Dedicated | SOC 2 | ZDR | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OpenRouter | $0.151 | $0.031 | — | Yes | $0.101 | $0.321 | — | Yes | — | — | Yes | 2026-11-09 |
| CoreWeave | $0.171 | $0.031 | GPU | Yes | $0.711 | $0.711 | — | Yes | Yes | — | — | 2026-11-09 |
| DeepInfra | $0.171 | $0.0371 | GPU | Yes | $0.102 | $0.322 | — | Yes | Yes | Yes | Yes | 2026-11-09 |
| Hugging Face Inference Providers | $0.171 | $0.0371 | — | Yes | $0.1351 | $0.401 | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| Crusoe | $0.251 | $0.051 | GPU | — | — | — | — | Yes | Yes | Type 2 | — | 2026-11-09 |
| Novita AI | $0.251 | $0.051 | GPU | Yes | $0.1351 | $0.401 | — | Yes | Yes | Yes | — | 2026-11-09 |
| Google Vertex AI | $0.361 | $0.091 | Mixed | Yes | $0.721 | $0.721 | — | Yes | Yes | Yes | — | 2026-11-09 |
| SiliconFlow | $0.451 | $0.051 | GPU | — | — | — | — | Yes | Yes | — | — | 2026-11-09 |
| Baseten | $0.502 | $0.102 | GPU | Yes | — | — | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| SambaNova | $0.591 | $0.221 | Custom silicon | Yes | $0.601 | $1.201 | — | Yes | — | Type 2 | — | 2026-11-09 |
| Amazon Bedrock | $0.602 | $0.152 | — | Yes | $0.722 | $0.722 | — | Yes | Yes | Yes | — | 2026-11-09 |
| Azure AI Foundry | $0.602 | $0.152 | — | Yes | $0.716 | $0.716 | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| Featherless AI | $0.602 | $0.152 | GPU | Yes | — | — | — | Yes | Yes | — | Yes | 2026-11-09 |
| Fireworks AI | $0.601 | $0.151 | GPU | Yes | — | — | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| Groq | $0.602 | $0.152 | Custom silicon | Yes | — | — | 5001 | Yes | — | Yes | Yes | 2026-11-09 |
| Nebius Token Factory | $0.601 | $0.151 | GPU | Yes | — | — | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| Together AI | $0.601 | $0.151 | GPU | Yes | — | — | — | Yes | Yes | Type 2 | Yes | 2026-11-09 |
| Replicate | $0.722 | $0.182 | GPU | — | — | — | — | Yes | — | — | — | 2026-11-09 |
| Cerebras | $0.751 | $0.351 | Custom silicon | Yes | — | — | 3,0001 | Yes | Yes | — | Yes | 2026-11-09 |
| Cloudflare Workers AI | $0.751 | $0.351 | GPU | Yes | $0.2931 | $2.2531 | — | Yes | — | — | Yes | 2026-11-09 |
| Parasail | $0.751 | $0.101 | GPU | Yes | $0.221 | $0.501 | — | Yes | Yes | Type 2 | — | 2026-11-09 |
| Alibaba Cloud Model Studio | — | — | — | Yes | — | — | — | Yes | — | — | — | 2026-11-09 |
| FriendliAI | — | — | GPU | Yes | — | — | — | Yes | Yes | Type 2 | — | 2026-11-09 |
| Hyperbolic | — | — | GPU | Yes | — | — | — | Yes | Yes | — | — | 2026-11-09 |
| Mistral AI | — | — | — | — | — | — | — | Yes | — | — | — | 2026-11-09 |
| Modal | — | — | GPU | — | — | — | — | Yes | Yes | Type 2 | — | 2026-11-09 |
| Nscale | — | — | — | — | — | — | — | Yes | Yes | Type 2 | — | 2026-11-09 |
| Scaleway Generative APIs | — | — | — | — | — | — | — | Yes | — | — | — | 2026-11-09 |
Prices are USD per 1M tokens for the standard serverless tier. Speed is the figure each provider publishes for its own API. Sorted by gpt-oss-120b output price; click any header to re-sort. Retired APIs, kept for reference: Anyscale No public API, Lambda Discontinued.
Cheapest LLM API for gpt-oss-120b
Full price rankingHow the table is built
Full methodologyOne reference model. Every provider is priced on gpt-oss-120b, the open-weight model the most providers sell per token, so the price column compares like with like. Llama 3.3 70B sits beside it for teams still running Llama.
Primary sources. Each value comes from the provider’s own pricing page, docs or trust center. The small marker next to a number opens that page; the tooltip shows the date we read it. An empty cell means the provider does not publish the figure.
Re-checked monthly. Inference prices move fast: Cerebras raised gpt-oss-120b since launch. Every row carries its verified date, newest 9 Nov 2026. Spot a changed price? Email hello@aiinferenceproviders.com with the link.
What developers ask before picking an inference API
What is the cheapest API for gpt-oss-120b?
CoreWeave lists $0.03 input / $0.17 output per 1M tokens. Through OpenRouter, the cheapest routed hosts charge $0.03 / $0.15, plus the 5.5% credit-purchase fee. Artificial Analysis agrees: it ranks CoreWeave cheapest on blended price at $0.044. See the full ranking.
Which inference provider is fastest?
Custom silicon leads on published speed. Cerebras quotes about 3,000 tokens/s on gpt-oss-120b (Wafer-scale engine (CS-3 systems)). Groq quotes about 500 tokens/s on gpt-oss-120b (LPU). Both are vendor figures, linked to the page that states them. The table adds an independent speed column the day a third-party benchmark for gpt-oss-120b goes live.
Can I switch providers without rewriting code?
Usually yes. 21 of the 28 live providers document an OpenAI-compatible API: Alibaba Cloud Model Studio, Amazon Bedrock, Azure AI Foundry, Baseten, Cerebras, Cloudflare Workers AI, CoreWeave, DeepInfra, Featherless AI, Fireworks AI, FriendliAI, Google Vertex AI, Groq, Hugging Face Inference Providers, Hyperbolic, Nebius Token Factory, Novita AI, OpenRouter, Parasail, SambaNova and Together AI. In most cases you change the base URL, the API key and the model name, and your client code stays as it is.
Do inference providers store my prompts?
Fireworks applies zero data retention by default for open models. DeepInfra never writes inputs to disk. Groq keeps nothing by default, may hold data up to 30 days for abuse checks and offers a ZDR toggle. Together and Nebius make ZDR an opt-in setting (Together, Nebius). Filter the table by ZDR to see every provider that states it.