Pricing table

Costs are computed from an embedded, filtered snapshot of LiteLLM’s pricing table (internal/pricing/table.json.gz), refreshed in the background at startup and cached, so pricing works instantly and offline on first run. Each request is priced at its own model and its own long-context tier, so a /model switch or one very long prompt never reprices the rest of a session.

Every figure is an API-price estimate. It ignores subscription plans (Claude Pro/Max, ChatGPT/Codex) entirely, so a session running inside a plan’s included usage still shows a nonzero cost. See honest labelling.

The embedded table

The binary carries a filtered snapshot of LiteLLM’s model_prices_and_context_window.json. wattop reads it at startup, so no network call is on the startup path. It is filtered to the models seen in use (claude-opus-5, claude-fable-5-1, claude-sonnet-5, claude-haiku-4-5, gpt-5.6-terra, gpt-5-codex) plus their current and recent sibling releases, so a version bump on either agent still resolves without a refresh.

The current embedded snapshot:

Field	Value
Upstream commit	56a61cf016a4fd44de520f5fcd54d6263f915699
Fetched	2026-09-06T01:29:41Z
SHA-256 of the upstream JSON	f68d88c12610ea31ab355a1293fde55aeed6fa78a1f4b182c67be47d80b1d202
Upstream entries at fetch time	3818
Keys kept	66

Background refresh

At startup, wattop checks its cache file:

$XDG_CACHE_HOME/wattop/pricing.json

When XDG_CACHE_HOME is unset, that is ~/.cache/wattop/pricing.json.

• If the cache is less than 24 hours old, wattop uses it and makes no request.
• Otherwise it makes one GET of the full upstream table from raw.githubusercontent.com in the background (10-second timeout), writes the result to the cache, and switches to it. The full table resolves strictly more models than the embedded snapshot, never fewer.
• If the request fails, the failure is not fatal and the table embedded in the binary stays in use.

So there is at most one pricing request per 24 hours, shared across every wattop you run. --demo never refreshes; it uses the embedded table only.

wattop doctor prints the table in use: its source URL, SHA-256, entry count, age and status. If the table is more than 30 days old, because refreshes keep failing, the pricing source is reported as degraded in doctor and on the dashboard.

Unknown models

A model name is matched against the table exactly first, then with a bracketed suffix stripped (claude-opus-5[1m] to claude-opus-5), then with a trailing -YYYYMMDD date stripped, then in its anthropic./openai.-prefixed forms, then by longest prefix, so it resolves to the closest sibling model. If nothing matches, the cost renders $—, never $0.00, and the model is listed in the footer’s unpriced-model list (unpriced_models in --json). A ~ before a cost means it leaves out usage on an unpriced model.

Verifying the checksum

U=https://raw.githubusercontent.com/BerriAI/litellm
U=$U/main/model_prices_and_context_window.json
curl -fsSL "$U" | shasum -a 256

Compare the first field of the output against the SHA-256 above. A mismatch alone is not a problem: it just means upstream has moved since the embedded snapshot was generated.

Updating the embedded table

In a source checkout, make pricing pulls the latest upstream table and regenerates internal/pricing/table.json.gz and docs/pricing-update.md with the commit and checksum that produced it.

To add a model upstream has not published yet, add its bare key (matching LiteLLM’s key exactly, e.g. claude-opus-6) to the KEYS heredoc in scripts/pricing.sh, then re-run make pricing. If upstream genuinely has no entry yet, make pricing logs it as missing and drops it; until upstream publishes a real entry, the model resolves through the fallbacks above or renders $—.