Skip to content

pricing.md

Latest commit

 

History

History
90 lines (66 loc) · 4.05 KB

pricing.md

File metadata and controls

90 lines (66 loc) · 4.05 KB
Raw

Pricing

CodingAgentRunner.Pricing provides a pure cost API and a static price snapshot with history. A caller can also supply its own listings.

It ships in the core CodingAgentRunner package — reference it the same way you reference CliThinkingLevels.

The model

  • A ModelListing is one model: its canonical id, any aliases that resolve to it, an optional vendor, and a price history.
  • A ModelPrice is one entry in that history: InputPerMTok, OutputPerMTok, optional CacheReadPerMTok / CacheWritePerMTok, a Currency, a ValidFrom UTC instant (inclusive), and a Source / Note / Unconfirmed flag.
  • Prices are kept, not overwritten. A price change adds a new entry with a later ValidFrom. The cost of a run is computed with the entry that was valid at the run's timestamp, so a re-priced run in the past still costs what it cost then.

The cost API

ModelPriceCatalog.Default is the seeded catalog. It uses the TokenEconomy catalogue snapshot dated 2026-10-04, marked for TokenEconomy 0.3.7. Everything on it is pure and deterministic.

using CodingAgentRunner.Pricing;

var catalog = ModelPriceCatalog.Default;

// What was this model's price at a given time?
PriceResolution p = catalog.ResolvePrice("claude-opus-4-8", DateTime.UtcNow);

// Cost of a run's usage, priced at the run's UTC timestamp.
CostBreakdown cost = catalog.ComputeCost(
    "claude-opus-4-8",
    new TokenUsage(Input: 120_000, Output: 8_000, CacheRead: 40_000),
    runStartedUtc);

if (cost.HasPrice)
    Console.WriteLine($"{cost.Total} {cost.Currency}" + (cost.Unconfirmed ? " (unconfirmed)" : ""));

// List endpoint: every model and its history.
foreach (var listing in catalog.Listings) { /* … */ }

You can also build a catalog from your own ModelListing set: new ModelPriceCatalog(listings).

Unknown and unpriced models are explicit — never a silent $0

The status is part of the return type, so a missing price can't be mistaken for a free run:

Situation PriceStatus CostBreakdown.Total
Price found for the timestamp Resolved the computed total
Model id not in the catalog UnknownModel null
Model known, but no price valid at/before the timestamp NoPriceForDate null

ComputeCost returns Total == null for both non-Resolved cases. There is no logging to consult — the outcome is in the value.

Lookups

Model ids and aliases resolve case- and dot/dash-insensitively, so claude-opus-4.8, claude-opus-4-8, dated snapshots like claude-haiku-4-5-20251001, and gpt-5-6 / gpt-5.6 all resolve to their listing.

Seed data and confidence

The pinned source is pricing/model-prices.json, copied from TokenEconomy's src/TokenEconomy/catalog/model-prices.json. Its source version, date, path, and SHA-256 are in pricing/catalog-source.json. The generated ModelPriceSeed.cs carries all supported listing fields and dated price history, including the source text and confidence flag. The pricing test compares every generated listing and rate with the pinned JSON and checks its version and hash.

To refresh the seed, copy a newer TokenEconomy catalogue into pricing/model-prices.json, update pricing/catalog-source.json from its snapshot index, run python3 scripts/generate-model-price-seed.py, then run python3 scripts/generate-model-price-seed.py --check and the package tests. The runner keeps this snapshot in its package and has no TokenEconomy runtime dependency.

The seed represents the catalogue's standard API tariff. For example, GPT-6 Sol and GPT-6.1 Sol have higher long-context rates that this API cannot select from the token counts alone. Use a host-supplied catalog when the applicable tariff differs.

When an entry omits cache rates, cache-read and cache-write tokens are billed at the input rate rather than dropped — a documented approximation used only for entries that lack their own cache figures (the seeded Anthropic entries all carry them).