model-pricing

A Paseo plugin that adds a Model pricing panel to the sidebar: one table of what every model on the providers you follow costs, so choosing which model to point an agent at does not mean opening five vendor pricing pages and doing arithmetic.

Each row carries the price per million tokens in and out, the context window, how much the model can write in one go, whether it can reason, call tools, return structured output and take a temperature, and a relative-cost figure that measures it against the cheapest model on screen. Sort by any of it, search for a model by name, and switch providers on and off.

Prices come from two places, neither of which needs an API key. Anthropic, OpenAI and Fireworks AI publish no pricing API at all — their own model endpoints want a key and return no prices — so those three, together with Ollama Cloud, are read from models.dev, a community catalog of model specifications. OpenRouter publishes its own, so it is read directly from OpenRouter. The daemon does the fetching; nothing is sent anywhere.

What you need

• Paseo 0.9.0 or newer, with plugins enabled.
• Network access from the daemon machine to models.dev and openrouter.ai.

No accounts, keys or logins. The plugin only reads public price lists.

Install

paseo plugin install npm:@gpambrozio/paseo-model-pricing

That is the shortest route on Paseo 0.9 or newer, which installs plugins straight from npm; add @<version> to pin one. On 0.8, install the last release that supported it from this repository instead:

paseo plugin add gpambrozio/paseo-plugins --path model-pricing

Pin a repository install with --ref model-pricing/v<version>. To hack on it, clone the repository and paseo plugin install "$PWD" from this folder after npm install and npm run typecheck.

Settings

Under Settings › Plugins › Model pricing, or the gear in the panel header.

Providers. One switch each for Anthropic, OpenAI, Fireworks AI, Ollama Cloud and OpenRouter. A provider switched off is not shown and not fetched, so turning off the ones you do not use makes the panel lighter as well as shorter. All five are on to begin with, which is a little over five hundred models — OpenRouter alone is four hundred of them.

Input / output blend. The relative-cost column needs one number per model, and a model's input and output prices are usually different, often by five times. This setting says how much of that blend is input. The default, 80/20, suits an agent that reads a large repository and writes a small patch; a chat that drafts prose is closer to 25/75. The ranking genuinely reorders between them, which is why it is a choice rather than a constant. The prices themselves are always shown in full, so only the last column changes.

Only models that can call tools. On by default, because a model that cannot call a tool cannot run an agent. Models whose provider does not say either way are always shown rather than guessed at.

Hiding a provider for a moment

The coloured dots under the heading are buttons. Tap one and that provider's models drop out of the table and its dot goes hollow; tap it again and they come back. This is instant and does not touch the network — it is for narrowing the table while you compare two providers, not for saying which ones you care about. Use the settings switches for that: a provider switched off there is not fetched at all. Hiding is forgotten when you restart the app.

Reading the table

Prices are US dollars per million tokens, written input first. An em dash means the provider did not publish that fact, not that the answer is no — a third of the models in the catalog simply do not state whether they support structured output.

The relative column divides each model's blended price by the cheapest one currently on screen, so the cheapest always reads 1.0× and the rest say how many times more they cost. It is computed over what you can see: switching a provider off or typing in the search box moves the baseline, because a multiple of a price that is not on screen would not mean anything. Models that are genuinely free say Free rather than 0.0×.

Press a row — or a card, on a narrow window — and the provider's own page for that model opens in your browser, where the description, the licence and the vendor's own price live.

On a phone or a narrow window the ten columns become one card per model — the same information, laid out to be read rather than scanned across.

Where prices are kept

Fetched prices are cached on the daemon under $PASEO_HOME/plugins/model-pricing/, one small file per upstream, and reused for twelve hours. Refresh ignores that and asks again. The refresh is cheap: models.dev supports conditional requests, so an unchanged catalog costs a single round trip rather than a 4.7 MB download.

If one upstream cannot be reached, the panel keeps showing the last prices it has for that source with a note saying so and how old they are, and the other source's rows are unaffected.

Limitations

• The prices are only as current as their source. models.dev is community-maintained, so a price change at Anthropic or OpenAI reaches this table when somebody updates the catalog, not when the vendor announces it. Each provider's own pricing page remains the authority; this table is for comparing, not for billing.
• Only per-token prices are shown. Cached input, batch discounts, image and audio pricing, and minimum charges are not in the table, so a model's real cost can be well below what the input column implies if you use prompt caching heavily.
• A model that appears twice is not a bug. With OpenRouter on, a model sold both directly and through the gateway is two rows with two prices, which is usually the comparison you wanted.
• A model's link is worked out from its name, because no catalog publishes one. It is right for every model in the table today, but a provider that retires a page or files a new model somewhere unexpected can leave a row pointing at a page that no longer exists.
• "Ollama" here means Ollama Cloud, the hosted models with real per-token prices — not the models pulled onto your own machine, which cost nothing to run and are not listed.