Providers & pricing

llmprof prices every call from input/output dollars per token. It covers 1000+ model ids across the providers people actually use, is fully overridable, and needs no network call: the data is bundled.

Where the prices come from

Prices resolve from three tiers, highest precedence first:

1. Your overrides - anything you set via LLMPROF_PRICING (below).
2. Curated families - a small hand-maintained table that is authoritative for the headline and newest models (the latest Claude, GPT, and Gemini), so a brand new release is priced correctly even before the snapshot catches up.
3. Bundled snapshot - a vendored copy of LiteLLM’s pricing database (MIT, Copyright (c) 2023 Berri AI), the ecosystem’s provider-sourced table. It gives broad coverage for everything else: OpenAI, Anthropic, Gemini, DeepSeek, xAI Grok, Cohere, and hosted open weights (Llama, Qwen, Mistral / Mixtral, Gemma, GLM, Phi, and more).

The snapshot ships in the package and loads with no network call, so pricing works fully offline. Refresh it by re-running the generator and committing the diff:

python scripts/pricing/sync_pricing.py --date YYYY-MM-DD

Within each tier, model ids match by exact id first, then the longest substring key, so gpt-4o-mini is priced differently from gpt-4o, and a provider-prefixed id like meta-llama/Llama-3.1-405B-Instruct still resolves.

Unknown models degrade gracefully

If a model id matches nothing in the table, the call still shows its token breakdown - only the dollar figure is omitted. Accuracy of the displayed cost matters more than coverage, so prices that could not be verified are left out rather than guessed.

Hosted open-weight prices are representative

The same open model (say Llama 3.3 70B) costs different amounts on DeepInfra vs Fireworks vs Cerebras vs the first-party API. The curated and snapshot values are representative of one provider; pin exact numbers for yours with an override.

Overriding prices

Point LLMPROF_PRICING at a JSON file of {"model": [input_per_1k, output_per_1k]}:

{
  "my-finetuned-gpt-4o": [0.003, 0.012],
  "llama-3.3-70b": [0.00023, 0.0004]
}

LLMPROF_PRICING=./prices.json llmprof up

Overrides are the highest-precedence tier, so they win over both the curated and bundled tables. You only need to list what you want to change or add.