codex-island-model-catalog

Model prices for CodexIsland, published as a plain JSON file so a new model does not require an app release.

Endpoint: https://ericjypark.github.io/codex-island-model-catalog/v1/models.json

{
  "schemaVersion": 1,
  "generatedAt": "2026-07-26T04:16:19Z",
  "source": "litellm + overrides",
  "models": {
    "claude-opus-4-8": {
      "displayName": "Opus 4.8",
      "inputPerMillion": 5,
      "outputPerMillion": 25,
      "cacheCreationPerMillion": 6.25,
      "cacheReadPerMillion": 0.5
    }
  }
}

Prices are US dollars per million tokens. Keys are canonical model ids with any 8-digit date suffix stripped (claude-haiku-4-5-20251001 → claude-haiku-4-5), matching how the app normalizes the ids it reads out of local CLI session logs. Ids carrying a dashed date (gpt-5-2025-08-07) are published as-is, because the app does not strip those either — collapsing them would make the app look up an id this file never publishes and silently price it at $0.

schemaVersion is 1 and will stay 1. A breaking change to this shape ships at /v2/models.json so installed apps keep working.

How a price gets here

A bot refreshes this file every six hours. You do not need its source to check a value — both inputs are public and the rules are below.

1. Fetch LiteLLM's price table.
2. Keep entries whose mode is chat or responses, whose id contains no / (that drops azure/, bedrock/, vertex_ai/ re-listings of the same model), and whose id matches a pattern in config.json. The gpt-[0-9]* pattern covers numbered GPT generations, including GPT-6 and future generations, so a major-version launch needs no filter update.
3. Convert per-token prices to per-million by multiplying by 1,000,000, then round to 8 decimal places. The rounding exists so a price reads as 0.2 rather than 0.19999999999999998 — binary floating point leaves residue that would otherwise be published verbatim into a file meant to be read.
4. Where LiteLLM lists no cache-write rate, use the input rate: OpenAI bills cache writes at the standard input rate. Where it lists no cache-read rate, use 0.
5. Apply overrides.json on top, field by field.
6. Generate displayName from the id unless an override supplies one.

Correcting a price

Open a PR against overrides.json. An override beats whatever upstream says, and only the fields you list are replaced:

{ "gpt-5.6": { "cacheCreationPerMillion": 6.25 } }

An override can also introduce a model LiteLLM does not list at all — supply all four rates and a displayName in that case.

The entries there now are deliberate: OpenAI began billing cache writes at 1.25× input starting with 5.6, and LiteLLM lists no cache-read rate for the pro tier because those models have no prompt caching.

Why you can trust the numbers

The app's release cycle used to be a human review gate on every price change. This file removes that gate, so the bot carries its own:

• If the merged model count falls below half of what is already published, the whole run is rejected and nothing is committed.
• A negative rate, or any rate above $1000 per million, is rejected.
• A rate that moved by 10× or more from its published value is rejected.
• A rejected model keeps its previously published value rather than disappearing. A missing model prices to $0 in the app, which is the failure this whole arrangement exists to prevent.
• Models are never removed. If LiteLLM drops an entry, the published value stays until a human removes it deliberately.
• The file is committed only when a price actually changes, so generatedAt marks the last real change rather than the last time the bot ran.

Override values deliberately skip the absolute ceiling — that is what lets a genuinely expensive model exceed it — so review of the PR is the control on those. Negative values are still rejected.

Repository contents

Path	What it is
v1/models.json	Generated. The only file the app reads.
overrides.json	Hand-maintained. Beats LiteLLM. Open a PR here.
config.json	Which model id patterns are tracked.