Pricing and cost¶

PromptKit reports what a call cost, using token counts from the provider and a vendored table of model rates.

completion = run_prompt(prompt, inputs, engine)

completion.usage.prompt_tokens  # from the provider
completion.usage.estimated  # False when the provider reported them
engine.cost_of(completion)  # None when the model's rates are unknown

promptkit cost greet.yaml --model claude-opus-4-5
promptkit cost --models

Where the rates come from¶

promptkit/data/pricing.json is a snapshot vendored into the package, generated from LiteLLM's public model-pricing dataset (MIT licensed). The source and its licence are recorded in the file's _meta. There is no network call at import and no runtime dependency on LiteLLM.

Its _meta.updated field is the snapshot date, surfaced by promptkit cost and available as promptkit.pricing.PRICING_UPDATED. A scheduled workflow refreshes it and opens a pull request when rates move, so staleness is visible rather than silent.

make pricing-refresh    # update the snapshot
make pricing-check      # fail if it is out of date

Two entries are curated by hand because upstream dropped them while the models are still in use: claude-3-5-sonnet and claude-3-5-haiku. They are listed in _meta.curated.

How a model name is resolved¶

Lookup is deliberately conservative. A name resolves in exactly three ways:

Kind	When	Example
`exact`	The name is a key in the table	`gpt-4o-mini`
`snapshot`	The name is a dated or `-latest` variant of a key	`claude-3-5-sonnet-20241022`
`override`	You registered it yourself	`register_pricing("my-model", ...)`

Anything else returns None, and PromptKit reports no cost rather than a wrong one.

from promptkit.pricing import lookup

found = lookup("claude-3-5-sonnet-20241022")
found.matched  # 'claude-3-5-sonnet'
found.kind  # 'snapshot'
found.exact  # False — inferred from a dated variant
found.describe()  # "claude-3-5-sonnet (dated snapshot of ...)"

Why there is no prefix matching

An earlier version resolved any unknown name to its longest matching prefix. That looked helpful and was dangerous: claude-opus-4-5 resolved to claude-opus-4 and reported a cost three times too high, and o3-pro was priced as o3.

A cost report that is confidently wrong is worse than one that admits it does not know. Only dated snapshots — which genuinely share a price with their base model — are inferred.

A name that two providers price differently is excluded from the snapshot entirely, for the same reason.

Adding a model¶

from promptkit.pricing import register_pricing

register_pricing("my-finetune", input_per_million=1.5, output_per_million=6.0)

Overrides take precedence over the vendored table and apply to dated variants too. Use this for fine-tunes, private deployments, negotiated rates, or a model newer than the snapshot.

Exact token counts¶

Install promptkit-core[tokens] and counts come from tiktoken instead of a heuristic. promptkit cost labels which it used.

Honesty rules¶

Two things PromptKit will not do:

Report a cost it cannot justify. Unknown model, no number.
Pass an estimate off as a measurement. Usage.estimated is True whenever counts were derived rather than reported, and RequestCompleted.cost_exact is False when the rates came from a snapshot match.

Estimates from promptkit cost are, by nature, estimates: the input count is real but output length is an assumption you control with --output-tokens.