Methodology
One model, one unit, the provider’s own page behind every number, and a date on every row.
The reference model: gpt-oss-120b
Providers price dozens of models differently, so a single “from” price compares nothing. We price every provider on one model: gpt-oss-120b, OpenAI’s open-weight model released in August 2025. It is the open model the most providers sell per token: 21 of the 28 live providers here publish a per-token price for it.
Llama 3.3 70B Instruct is the secondary reference. Fewer providers list it each month, so its columns stay filled only where a provider still sells it.
Units
- Token prices are US dollars per 1 million tokens, input and output listed separately, for the standard serverless tier. Batch, priority, cached-input and committed-use rates are on each provider’s page and in the pricing detail on its profile.
- Where a provider bills in its own unit (Cloudflare’s Neurons), we use the USD-per-token equivalent the provider publishes.
- Hyperscaler prices are for the region named in the source: AWS US East, Azure East US 2 (Global deployment).
- Dedicated prices are USD per H100 GPU-hour, on demand. Per-second and per-minute rates are converted to hours, and the profile shows the original rate.
- The cheapest-API ranking adds a blended price, (3 × input + output) ÷ 4, the same 3:1 mix Artificial Analysis uses. It is our arithmetic on the listed prices.
Speed
The speed column is the output speed the provider itself publishes for gpt-oss-120b, labelled vendor-stated and linked to the page that states it. Each speed is the provider’s own figure, so you can check it at the source in one click. When an independent per-provider benchmark for gpt-oss-120b publishes, it gets its own column with its own source.
Sources and dates
Every value comes from the provider’s pricing page, documentation, trust center or price API. Each profile lists those pages with the date we read them, and each number links straight to its page. Every filled cell is a published figure, and an empty cell marks a figure the provider has yet to publish. Compliance (SOC 2, HIPAA, zero data retention) counts when the provider states it on its own pages.
The monthly re-check
Inference prices move: Cerebras raised gpt-oss-120b from $0.25 / $0.69 at launch to $0.35 / $0.75. We re-verify every row each month and on any reported change, and each row shows its verified date. The newest check is 9 Nov 2026.
Retired APIs
When a provider shuts its public per-token API, it leaves the main table but keeps its page, marked with its status and the cheapest live alternatives. Today that covers Lambda (Inference API sunset on 25 September 2025) and Anyscale (no public per-token endpoint).
Corrections and sponsors
Providers and readers can send a correction with a public source to hello@aiinferenceproviders.com. We update the row and its date. The featured slot is paid and labelled; every row’s data and sort position come from the sources alone.
Field definitions
22 fields- Status
- Whether the provider's public inference API is live.
- Hardware
- Chips that serve requests: GPUs, the vendor's own custom silicon, or both.
- Hardware detail
- Published chip types (e.g. LPU, wafer-scale engine, RDU, NVIDIA H100/B200).
- OpenAI API
- Accepts OpenAI Chat Completions-style requests, per vendor docs.
- Serverless
- Pay-per-use API billed per token or per second, with no GPU to manage.
- Dedicated
- Single-tenant deployments or reserved capacity.
- Fine-tuning
- Hosted fine-tuning or LoRA adapters offered.
- Models (published) models
- Model count as published or counted from the official model list on the verified date.
- Flagship models
- Notable models served, from the official list.
- Serves gpt-oss-120b
- Offers the reference model gpt-oss-120b pay-per-token.
- gpt-oss-120b in $/1M tok
- Standard serverless input price for the reference model gpt-oss-120b, per 1M tokens.
- gpt-oss-120b out $/1M tok
- Standard serverless output price for the reference model gpt-oss-120b, per 1M tokens.
- Llama 3.3 70B in $/1M tok
- Secondary reference: Llama 3.3 70B Instruct input price per 1M tokens, where still listed.
- Llama 3.3 70B out $/1M tok
- Secondary reference: Llama 3.3 70B Instruct output price per 1M tokens, where still listed.
- Vendor speed tok/s
- Vendor-stated output speed on gpt-oss-120b, linked to the vendor’s own page.
- Speed source
- Who published the speed figure and when.
- Dedicated H100 $/GPU-hr
- On-demand price of one H100 for dedicated deployments.
- Regions
- Published serving or data regions.
- SOC 2
- SOC 2 status as stated by the vendor.
- HIPAA
- Vendor states HIPAA support or offers a BAA.
- ZDR
- Zero data retention: the vendor states prompts and outputs are not stored, by default or as a setting.
- Free tier
- Free credits or free rate-limited access, as published.