Model prices are a table in the code: OpenAI only, and fixed between releases #38

Priority

extreme

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Today

A standard run's cost estimate comes from a table written in the code, run.py:86:

PRICES = {
    "gpt-5-mini": {"input": 0.25, "cached": 0.025, "output": 2.0},
    ...
    "text-embedding-ada-002": {"input": 0.1, "cached": 0.1, "output": 0.0},
}

It holds 13 OpenAI models and nothing else, so a run sent directly to another provider has no cost unless the reply carries one (OpenRouter's does). The numbers change only with a release. A model missing from it is priced by the first entry its name starts with (relay.py:383), so gpt-4.1-nano is priced as gpt-4.1. A user cannot add or correct a price.

Wanted

Prices are the user's data, not the code's, for anyone who reads a run's cost or wants to bound it.

• One price list in the user's configuration: per model, as its provider names it, the input, cached input, cache write and output prices per million tokens and the model's output limit.
• Filled on request from a public list keyed by the providers' own model names (LiteLLM's model_prices_and_context_window.json), from Settings and from the CLI, with its source and date kept. Never fetched during a run.
• A user's own price for a model wins over the list; --price MODEL=INPUT,OUTPUT[,CACHED] prices one run.
• A call's cost: the provider's own figure when its reply carries one, else the price of exactly that model name, else none.
• No price ships with Nullius: a new profile shows no estimate until its list is filled once.

Acceptance

• No price is written in the code: run.PRICES is gone and a new profile has no prices.
• nullius prices update fills the user's price list from the public list, with its source and date, and Settings has the same action; a run never fetches it.
• A user's own price for a model (nullius prices set, --price on a run) wins over the list, and can be removed.
• A call is priced by its exact model name or not at all: no prefix match.
• A run in docker or on a host is priced from the list of the machine that started it.
• A job's config.json records the prices it was given and where the list came from.
• The field guide, the Settings guide and the CLI pages say where an estimate's price comes from.

Dependencies and related

#14 (a spending cap) waits on this: a cap needs a price for every model a run calls.