PromptSpend for VS Code

What the model on this line costs, and when that price was last confirmed.

A linter for model choice, not a calculator in a sidebar. Write a model id and the price appears beside it, carrying the date it was verified and the vendor page it came from.

response = client.messages.create(
    model="claude-sonnet-5",   # $3 / $15 per M · max out $0.061
    max_tokens=4096,
)

What it does

Prices where the decision is made. Every model id in your code is annotated with its input and output rates. Hover for the full rate card: cached-input and cache-write rates, the context window, the output ceiling, long-context tiers, batch discounts, and a link to the vendor page the number was read from.

The output ceiling in money. A max_tokens near a model id becomes max out $0.061 — what the response alone costs if it runs to the cap, every call. A rate is abstract; that number is not. It says max out every time, because the input side is unknown at that point in the file and is often the larger half.

Impossible requests are named, not priced. A cap above what the model can emit shows ⚠ above model max and no figure. That request gets rejected or truncated — it does not cost more — so quoting a price for it would be quoting money nobody will ever be billed.

Model-choice diagnostics. Deprecated, legacy, disputed and stale models are flagged in the Problems panel. Optionally (promptspend.diagnostics: "all") so is a cheaper sibling that clears the same capability estimate. Everything is Information severity, never a warning: none of these are mistakes, and a squiggle that calls a defensible engineering decision wrong is how a linter gets uninstalled.

Cost a prompt. Select some text, right-click → Estimate selection. It prices the selection against the models already named in that file, cheapest first, saying for each whether the token count was measured exactly or estimated from a calibrated ratio.

See the whole repository. The LLM models in use view in the Explorer lists every model the codebase calls, most expensive first, with the files and lines that call it.

Where the prices come from

The catalog is fetched from promptspend.com/data/pricing.json, the artifact a GitHub Action re-verifies every morning against vendor pricing pages, a community catalog and an independent cross-check.

There is deliberately no bundled fallback. An extension sitting in the Marketplace for three months with prices baked into it would be the purest possible version of the problem this project exists to prevent, so if the catalog cannot be reached the extension says so and shows nothing rather than showing a number it cannot stand behind. Cached prices are used for up to a day when a refresh fails; past that they are withheld.

The status bar shows the date the catalog was generated, always. That is what makes the six-hour refresh interval honest rather than lax — you can see how current the data is without taking anyone's word for it.

Privacy

Nothing is sent anywhere. The extension makes exactly one network request — a GET for the public pricing catalog — and reads your files only in memory to find model ids. No telemetry, no account, no prompt text leaves your machine. Token counting runs locally, using the model's own tokenizer where one can be run and a clearly labelled ratio where it cannot.

Settings

Commands

Things it deliberately does not do

• Guess. A rate the provider does not publish is absent, never averaged or defaulted. A model with no capability estimate is never offered as a cheaper alternative, because "might be worse" is not an answer to "cheaper and as good".
• Price a dated snapshot id as its undated cousin. claude-sonnet-5-20260101 is not in the catalog, so it gets no annotation rather than an invented one.
• Recommend. The capability estimate behind the cheaper-model suggestion is illustrative, not a benchmark, so the wording is always worth testing rather than use this instead.

Related

• promptspend.com — the estimator, the value map and the pipeline's provenance
• @promptspend/mcp — the same catalog for coding agents
• promptspend.dev — a keyless JSON API over the catalog

Licence

MIT — see LICENSE. Source at github.com/AndrewAvery7/promptspend, in vscode/.