agent-models

A curated AI model registry — per-model pricing, context window, output cap and capability flags — published as one static JSON:

https://models.opspresso.com/models.json

(https://opspresso.github.io/agent-models/models.json redirects there.)

Agent Studio is the original consumer it is shaped for: the text and image fields mirror its src/domain/llm/models.ts, and the catalog is the list that registry will load instead of carrying the numbers in code. Specialized models are identified by the embedding, rerank, or transcription capability; consumers that do not support them can skip those entries. Anything that can read JSON can use it the same way. A browsable view of the same file is at https://models.opspresso.com/ — docs/index.html, a static page that reads the catalog and mirrors the console's own Models page (https://studio.opspresso.com/models): the same catalog header, filter row (search, provider, capabilities, sort) and one card per route, drawn with the same tokens, faces and brand marks (docs/icons/brands/, MIT-licensed Lobe Icons). The studio's mark is re-hued green for this site's logo and favicon, so the two are recognisably siblings and never mistaken for each other.

The catalog

{
  "version": 1,
  "updatedAt": "2026-08-20T00:03:11.000Z",   // when the content last changed — not when it was last checked
  "source": "https://github.com/opspresso/agent-models",
  "providers": ["openai", "anthropic", "google", "xai", "bedrock", "openrouter"],
  "makers": { "openai": "OpenAI", "anthropic": "Anthropic", "zhipu": "Z.ai", ... },
  "models": [
    {
      "id": "openrouter/claude-opus-4.8",        // provider/family — the id a run names
      "provider": "openrouter",                  // the route
      "family": "claude-opus-4.8",               // the model
      "maker": "anthropic",                      // who made it, whatever the route
      "displayName": "Opus 4.8",
      "pricing": { "inputPer1M": 5, "outputPer1M": 25, "cachedInputPer1M": 0.5 },
      "capabilities": { "tools": true, "structuredOutput": false, "imageInput": true, "reasoning": true },
      "contextWindow": 1000000,
      "maxTokens": 128000,
      "wireId": "anthropic/claude-opus-4.8",     // only when the provider names it differently
      "hidden": true                             // only when retired: resolvable, not offered
    }
  ]
}

Model types

The catalog has five mutually exclusive model types. Type is derived from capability flags so existing text and image entries keep their published shape:

Type	Discriminator	Pricing contract	Limits
Text	neither type flag is set	positive inputPer1M and outputPer1M	positive contextWindow and maxTokens
Image	capabilities.imageGeneration: true	imageOutputPer1M or perImage	zero is allowed when the provider publishes no token limits
Embedding	capabilities.embedding: true	positive inputPer1M; outputPer1M: 0	positive contextWindow; maxTokens: 0
Rerank	capabilities.rerank: true	positive inputPer1M or perSearch; outputPer1M: 0	positive contextWindow; maxTokens: 0
Transcription	capabilities.transcription: true	positive input/output token rates or perAudioMinute	zero limits are allowed for duration-priced models

The four type flags are mutually exclusive, and an offering may not change its family's type. imageInput remains an independent capability: a text, embedding or rerank model can accept images without becoming an image-generation model.

Prices are USD per million tokens; image models add imageInputPer1M, imageOutputPer1M, perImage and perInputImage, and are priced by token rate or per image. Every rate is the provider's base, standard-tier rate — a long-context or priority tier is not expressed. Embedding models charge inputPer1M only, carry outputPer1M: 0, and use maxTokens: 0 because they produce vectors rather than generated tokens. Rerank models charge per million input tokens or per search unit. Transcription models charge per million input/output tokens or per audio minute, depending on the upstream provider's billing contract. pricing.discount, when present, is a promotional discount (a fraction) that the stated rates are already net of — OpenRouter publishes one per endpoint, and the catalog's price is the default endpoint's — so a reader can tell a promotion from a price and put the list rate back with inputPer1M / (1 - discount). Absent means no discount is known. reasoningWithTools: false marks a model whose provider rejects tools together with reasoning_effort on chat/completions.

hidden is for routes that have been retired: a stored configuration may still name one, and past usage is priced by looking it up, so it stays in the list and out of the picker. An entry is deleted only when nothing can ever have referred to it.

How the registry is written

What a model is and who serves it are two lists, resolved into the catalog at build time.

models/
  providers.json              the routes a model id may be prefixed with, in catalog order
  makers.json                 maker id → { displayName, openrouterVendor? }
  families/<maker>.json       { "<family>": { displayName, pricing, capabilities, contextWindow, maxTokens, note? } }
  offerings/<provider>.json   route overrides plus hidden/presence/ranking lifecycle bookkeeping
  removals.json               exact published ids proposed for permanent removal through a PR

A family states the model once — price, window, what it can do. An offering says a provider serves it, under which wire name, and what the route changes: a router's own rate, a gateway that cannot do structured output, a retirement. Overrides are shallow merges over the family, so an offering names only what differs. The file an entry sits in is its maker or provider; neither is repeated inside the entry.

Three things follow from the split and are enforced by validateRegistry (and CI):

• the same model reached two ways has one name, one maker, one window and one kind — an offering may not change contextWindow or any model-type discriminator;
• a wireId exists only to say something the id does not: a router route must carry a vendor-qualified one (anthropic/claude-opus-4.8), and a dotted Anthropic id must carry the hyphenated name Anthropic actually serves (claude-opus-4-8);
• a text model is priced on both sides, an image model by token rate or per image, an embedding model on input, a rerank model by input or search, and a transcription model by tokens or audio minute; a cached rate never exceeds the uncached one, and the output cap fits in the window.

A note holds provenance a reader of the file needs — where an odd number came from, why a retired route is priced the way it is. It is for people; the catalog does not carry it.

Routes

• Bedrock offerings go through the OpenAI-compatible bedrock-mantle endpoint; wire ids are the endpoint's own (openai.gpt-oss-120b). structuredOutput is off for the route, not the model: mantle rejects the parameter whatever is behind it. Claude is absent on purpose — Bedrock serves it on the Anthropic Messages API only.
• OpenRouter ids are vendor/model, so every offering carries a wire id. A route restates a price only while OpenRouter's catalog differs from the family's rate.
• Anthropic hyphenates what this registry dots, hence the wire ids on claude-*-4.x.

Adding or changing a model

1. Add the family to models/families/<maker>.json (a new maker, including its optional OpenRouter vendor slug, goes in makers.json first) and an offering to each models/offerings/<provider>.json that serves it.
2. pnpm format — rewrites the files in canonical key order and validates them.
3. pnpm build — regenerates docs/models.json. CI fails if the committed catalog does not match its sources, so commit both.

Published ids first become hidden tombstones. After an automatically hidden route remains hidden for 30 days, the workflow creates or updates one draft deletion PR. That PR accumulates an exact { id, reason, requestedAt } entry in models/removals.json for each route, requests its reviewer when first opened (removals.json's code owner) and is not auto-merged. Removals are checked twice: the update workflow refuses to commit one at all, and CI rejects, on pull requests, a removal that no entry requests. A push straight to main skips only that removal check — the type check, tests and build:check still run.

Numbers come from the provider, not from memory: OpenRouter's /api/v1/models, /api/v1/embeddings/models, modality-filtered model catalogs and public model pages, xAI's /v1/language-models (prices in 1e-10 USD per token — 12500 is $1.25/M), Anthropic's /v1/models (max_input_tokens, max_tokens), the AWS Pricing API for Bedrock (mind the unit — 1K tokens and 1M tokens rows are mixed), and the pricing pages for OpenAI and Google, which publish no API for it.

Keeping it current

.github/workflows/update.yml runs every day at 22:00 UTC (and on demand), validates whatever moved and merges routine updates through a pull request — which republishes the Pages site (main:/docs, served at models.opspresso.com) — and tells people about it. Three phases, every source independent of the others:

1. fetch — each source reads its catalog; one failing is reported and the rest go on;
2. discover — what the catalogs list that the registry does not: new families and new routes (below), plus OpenRouter's ranking lifecycle;
3. apply — numbers, discounts and catalog presence for every route the registry now has, the ones just added included.

Only fetch failures are isolated that way: an error thrown while any source discovers or applies stops the run from writing models/ at all.

Source	Needs	May change	May add
OpenRouter /api/v1/models, /images/models, /embeddings/models, modality-filtered catalogs, /models/{id}/endpoints, public model pages and weekly rankings feeds	nothing	a router-only family's price, discount and window, plus generated output cap where applicable; an OpenRouter offering's price override, discount included (set while the router's rate differs from the family's, dropped when they agree)	eligible text, image, embedding, rerank and transcription families; OpenRouter routes to existing families
xAI /v1/language-models, /v1/image-generation-models	XAI_API_KEY	the text families' token prices (matched by id or alias; image models stay hand-kept)	xai/ routes
Anthropic /v1/models	ANTHROPIC_API_KEY	the families' contextWindow and maxTokens (no price is published)	anthropic/ routes
OpenAI /v1/models	OPENAI_API_KEY	no number — presence only	openai/ routes
Google /v1beta/models	GOOGLE_API_KEY	the families' contextWindow and generated maxTokens (no price is published; a cap that exceeds the family's window is noted, not applied)	google/ text and embedding routes

A key that is not set skips its source and says so. Bedrock has no source here: its routes are added and retired by hand, with numbers from the AWS Pricing API.

Additions

A new family is created from OpenRouter's catalog when a listing is filed under a maker's openrouterVendor in makers.json; is a priced text model (not a :suffix variant — :free, :nitro, :thinking and the rest — or a -20250929/-0813 dated snapshot); and states a context window plus a usable max output in the aggregate model or at least one endpoint. It must also be either from OpenAI, Anthropic, Google or xAI and first listed within the last 30 days, or present in the weekly usage leaderboard's first 20 open-weight or first 20 closed-weight rows. Ranked models are eligible regardless of listing age. The ranking is the same one shown at openrouter.ai/rankings: prompt and completion tokens are totalled by variant, models with a Hugging Face id are open weight, and the rest are closed weight. A ranked serving variant qualifies its standard model family.

The text family gets OpenRouter's numbers and flags (tools, structured_outputs, image input, reasoning), a note saying when and where it came from, and an OpenRouter route. The aggregate max output is preferred when it is a positive integer that fits the window; otherwise the largest valid endpoint cap is used because OpenRouter routes a max_tokens request only to endpoints that support it. An eligible text model from a maker this registry does not know and an eligible listing without any output cap go to the needs a look list instead. Unranked non-major text models are ignored. If the public text rankings feed fails, changes shape or cannot fill both weekly Top 20s, it is discarded whole — the run still updates existing routes and major-maker additions but adds no text family from another maker. Older unranked listings are the backlog, which is a person's — the window keeps a first run from importing a major vendor's whole history.

An image family is created when it is among the first 20 ranked models the image catalog can resolve, ordered by image-producing request count as on openrouter.ai/rankings/image. The image catalog supplies its name and modalities; /models/{id}/endpoints supplies the normalized image-token price, context window and output cap. Explicit zero limits are retained for image-only models when OpenRouter publishes them. A previously unknown maker is added from the ranked model's vendor slug and display-name prefix. Image ranking failure or missing endpoint price/limits fails closed for image additions without blocking text updates.

An embedding family is created when it is among the first 20 ranked models the embeddings catalog can resolve, ordered by embedding request count as on openrouter.ai/rankings/embeddings. The catalog supplies its name, input-token price, input modalities and context window. Its outputPer1M and maxTokens are zero because the endpoint returns vectors, not generated tokens. As with images, a previously unknown maker is adopted from the vendor slug and display-name prefix, and an incomplete ranking feed fails closed without blocking other modalities.

A rerank family is created when it is among the first 20 ranked models the rerank catalog can resolve. The catalog supplies its name, modalities and context window. The public model page supplies the native billing SKU because OpenRouter's Models API currently reports zero token rates for rerank models: either inputPer1M or perSearch is recorded. Its outputPer1M and maxTokens are zero because the endpoint returns relevance scores.

A transcription family is created when it is among the first 20 ranked models the transcription catalog can resolve. Duration-priced models use perAudioMinute and zero token limits; token-priced models retain their input/output token rates and published limits. For rerank and transcription, a complete ranking may contain fewer than 20 rows when the corresponding catalog itself is smaller than 20.

A new route is added when a catalog serves a family the registry already has: OpenRouter under <vendor>/<family>, a vendor under the family id (Anthropic's hyphenated spelling becomes the wireId). OpenRouter applies the major-maker or leaderboard eligibility gate to routes too, though without the 30-day listing window — a major maker's older listing can gain a route while founding no family. Two further guards: a family every route of which is hidden is never routed again — it was retired on purpose — and an OpenRouter listing is routed only when its context window equals the family's, the one cheap identity check there is (qwen/qwen3-235b-a22b is the original model; this registry's qwen3-235b-a22b is the Instruct 2507). A text route narrows tools/structuredOutput when the router lacks them. Image, embedding, rerank and transcription routes are added only while their model is in the corresponding weekly Top 20. The first vendor route to a router-only family puts the family at the list price (the router's discount moves to the router's offering).

Retirement

Every source also watches presence: whether each live route is still in its provider's catalog. A route that is not gets missingSince and an observation count; retries on the same UTC date count once. The day it is back, the bookkeeping goes. After 7 successful absent catalog observations the route is set hidden: true, missingSince is dropped and a sentence is appended to its note saying when and why. The lifecycle update itself never deletes; it preserves the tombstone until the separate draft deletion PR is reviewed.

OpenRouter uses separate admission and retention thresholds so models around the cutoff do not churn. Text enters through the weekly open- or closed-weight Top 20 and remains eligible through the corresponding Top 50; image, embedding, rerank and transcription models enter through their weekly Top 20 and remain eligible through the corresponding Top 30. A listing is exempt from ranking retirement for its first 90 days. Major-maker text routes with a live vendor route are also exempt, while an OpenRouter-only OpenAI, Anthropic, Google or xAI family follows the same retention policy as other OpenRouter-only families.

After 30 successful ranking observations outside the retention threshold, the route is hidden with a structured hiddenAt date. Re-entry during the tombstone period restores it automatically. A failed ranking read or a feed that cannot supply the complete Top 50/30 retention set does not advance this lifecycle. Routes hidden manually remain manual and are never restored by automation. An automatically hidden route becomes a permanent-removal candidate only after its 30-day tombstone period; candidates accumulate in the existing draft removal PR until it is reviewed.

A source read that fails is not observed at all — neither lifecycle advances — and an empty or malformed catalog is treated as a failed read, not as everything retired (only OpenRouter's text catalog can also detect a truncated listing). Bedrock has no presence source, so its routes are retired by hand: set hidden: true and say why in note. OpenRouter's announced expiration_date is reported while it is at most a year out — a date already past included.

Before anything is written, a safety gate quarantines destructive bulk changes: removals, provider-wide disappearance and mass catalog hiding. Ranking-qualified retirement, valid provider metadata changes and additions are applied automatically. The report is still produced for review, but models/ remains unchanged when the gate trips. Network reads refuse redirects, time out after 30 seconds and cap JSON responses at 10 MiB. After reviewing the complete diff, a person may apply that exact anomaly set by rerunning with the digest-specific --approve-anomaly=<digest> printed in the report — from CI by dispatching Update models with the digest in the approve_anomaly input. The approval is the digest alone, so a set that has since changed is quarantined again with a new one rather than waved through.

Notifications

Two channels with two roles, both optional and run even when a later build or test fails (scripts/notify.ts):

• Slack (SLACK_WEBHOOK_URL) carries events: a message per run that changed something — a price, a route added, a family added, a model retired — or that failed: a source that could not be read, a quarantined anomaly set, a patch that would leave the registry invalid. The run is linked. A quiet day posts nothing.
• A rolling GitHub issue (the workflow's own token) carries state: "Model registry: needs a look", rewritten by every run with everything that needs a person — an unknown maker, a window a router disagrees on, an absence being counted down, a source that could not be read — and closed the day the list is empty. Act on an item and it disappears the next day.

The job summary holds the full report on every run, including one that refused to write: a table of what changed and the needs a look list.

Run the update locally with pnpm update-models (--dry-run to only report); the keys are read from the same environment variable names. --reset-openrouter first preserves every OpenRouter route as a hidden reset tombstone, then restores what the retention sets (text Top 50, specialized types Top 30) and major-maker policies still allow and adds what the admission policies admit — with the 30-day listing window switched off, so a major maker's whole history is eligible. The reset is not written unless all rankings are complete and every image Top 20 endpoint has usable standard-tier price and limits. Free variants and otherwise unpriced specialized models remain ineligible without making the ranking feed incomplete.

Commands

pnpm install
pnpm format          # canonical key order + validation of models/
pnpm build           # models/ → docs/models.json
pnpm build:check     # exit 1 if docs/models.json is stale (CI)
pnpm check-removals -- HEAD^ --pull-request # verify PR removals against models/removals.json
pnpm propose-removals # turn update-report.json candidates into a draft-PR deletion patch
pnpm update-models   # pull the live sources into models/ (--dry-run to report only)
pnpm update-models --reset-openrouter # rebuild only OpenRouter from all eligible models
pnpm notify          # deliver update-report.json to Slack / the issue
pnpm typecheck
pnpm test

Node 24+, pnpm 11. The scripts are TypeScript run by Node directly; nothing is installed to run them — typescript and @types/node are dev dependencies for the type check alone.