The same model costs up to 4.66x more depending on provider — full pricing matrix published
I pulled every model x provider listing from the Inference Providers router (GET /v1/models) and published the full matrix:
https://huggingface.co/datasets/hardik90/hf-inference-pricing-matrix
The headline numbers (snapshot 2026-08-01):
- 14 providers, 107 models, 295 model x provider listings
- 44 models are served by multiple priced providers — for those, routing is a pure cost lever (same weights, same outputs)
- Worst spreads: Qwen3-235B-A22B-Instruct-2507 4.66x, openai/gpt-oss-120b 4.41x (10 providers), gemma-4-31B-it 3.92x
- deepinfra is the cheapest provider for most listed models; latency/throughput leaders vary per model
Every row includes: input/output price per 1M tokens, context length, first-token latency, throughput, tool support, structured-output support. Two files: provider-pricing.csv (per provider) and model-summary.csv (cheapest/fastest per model, price spread), both also in Parquet.
Why it matters: the router passes provider prices through with no markup, so for any multi-provider model, your choice of provider changes your inference bill by up to 4.4x with zero quality difference. The router supports :cheapest / :fastest / provider-pinned suffixes — the matrix tells you which to pin.
I'll refresh this snapshot monthly (pricing changes frequently). Feedback welcome — and if your team runs serious volume, I do written cost audits (per-model provider selection + endpoint right-sizing + deployable routing policy); open a discussion or message my profile.
— Hardik