feat(model-sync): sync native models from modelschemas only, with no prices in model metadata

.changeset/model-meta-no-prices.md

@@ -0,0 +1,11 @@
+---
+'@tanstack/ai-anthropic': patch
+'@tanstack/ai-gemini': patch
+'@tanstack/ai-grok': patch
+'@tanstack/ai-groq': patch
+'@tanstack/ai-mistral': patch
+'@tanstack/ai-openai': patch
+'@tanstack/ai-openrouter': patch
+---
+
+Model metadata no longer carries prices. Nothing read them, and prices change faster than a package release. To price a call, read the live rate card from modelschemas: see the Model Prices guide.

.changeset/sync-modelschemas-models.md

@@ -0,0 +1,7 @@
+---
+'@tanstack/ai-byteplus': patch
+'@tanstack/ai-groq': patch
+'@tanstack/ai-mistral': patch
+---
+
+Add new BytePlus, Groq, and Mistral models from the modelschemas catalog.

.github/workflows/sync-models.yml

@@ -44,11 +44,15 @@ jobs:
         id: elevenlabs
         continue-on-error: true
         run: pnpm generate:models:elevenlabs
+        env:
+          MODELSCHEMAS_API_KEY: ${{ secrets.MODELSCHEMAS_API_KEY }}
 
       - name: Sync provider models
         id: provider
         continue-on-error: true
         run: pnpm exec tsx scripts/sync-provider-models.ts
+        env:
+          MODELSCHEMAS_API_KEY: ${{ secrets.MODELSCHEMAS_API_KEY }}
 
       - name: Format
         run: pnpm format
@@ -87,15 +91,15 @@ jobs:
 
           if [ -z "$EXISTING_PR" ] || [ "$EXISTING_PR" = "null" ]; then
             BODY=$(cat <<'PRBODY'
-          Automated daily sync of model metadata from OpenRouter, Vercel AI Gateway, Lovable AI Gateway, LLM Gateway, Neon AI Gateway, and modelschemas (ElevenLabs text-to-speech, music, sound effects, and voice design).
+          Automated daily sync of model metadata from OpenRouter, Vercel AI Gateway, Lovable AI Gateway, LLM Gateway, Neon AI Gateway, and modelschemas (native provider models, plus ElevenLabs text-to-speech, music, sound effects, and voice design).
 
           - Fetches the latest model list from OpenRouter (chat + `GET /api/v1/videos/models`)
           - Fetches the latest model list from Vercel AI Gateway
           - Fetches the latest model list from Lovable AI Gateway
           - Fetches the latest model list from LLM Gateway (`GET /v1/models`)
           - Fetches the latest model list from Neon AI Gateway
           - Converts to the internal adapter format
-          - Syncs provider-specific model metadata for affected packages
+          - Inserts new native-provider models from modelschemas (`@modelschemas/client`)
           - Inserts new ElevenLabs text-to-speech, music, sound-effect, and voice-design models from modelschemas
           - Creates a patch changeset for all changed packages
           PRBODY

CONTRIBUTING.md

@@ -55,18 +55,24 @@ For deeper architecture details (adapter system, isomorphic tools, framework int
 
 1. Fetches OpenRouter, Vercel AI Gateway, Lovable AI Gateway, and Neon AI Gateway catalogs.
 2. Regenerates `packages/ai-openrouter/src/model-meta.ts` and the Vercel Gateway and Neon model lists.
-3. Inserts **new** native-provider models into `packages/ai-openai`, `ai-anthropic`, `ai-gemini`, and `ai-grok`.
-4. Inserts **new** ElevenLabs text-to-speech, music, sound-effect, and voice-design ids from [modelschemas](https://modelschemas.com/v1/schemas/elevenlabs) into `packages/ai-elevenlabs/src/model-meta.ts`. Transcription ids stay hand-maintained.
+3. Inserts **new** native-provider models from [modelschemas](https://modelschemas.com) (`@modelschemas/client`) into `ai-openai`, `ai-anthropic`, `ai-gemini`, `ai-grok`, `ai-groq`, `ai-mistral`, and `ai-byteplus`. Ids, activity, limits, modalities, and capabilities come from each provider's own catalog, with no second source. A field the catalog leaves empty stays empty, and a row with no modalities is skipped (the log counts it). Model metadata has no prices: they change faster than a release, so apps read them live (`docs/advanced/model-prices.md`).
+4. Inserts **new** ElevenLabs text-to-speech, music, sound-effect, and voice-design ids from [modelschemas](https://modelschemas.com/v1/schemas/elevenlabs) into `packages/ai-elevenlabs/src/model-meta.ts` (`scripts/sync-elevenlabs-models.ts`, also through `@modelschemas/client`). Text-to-speech ids are catalog rows with `providerMetadata.canDoTextToSpeech`. Music, sound-effect, and voice-design ids are the `model_id` enums of their request schemas. Transcription ids stay hand-maintained.
 5. Writes a patch changeset for the packages that changed.
 
+The sync stops with an error when modelschemas is down, when a catalog is empty, when no row has a `rawId`, `firstSeenAt`, or `deprecatedAt` key, or when an array or type map it must update is missing from a `model-meta.ts`.
+
+Reads against modelschemas work without a key (60 req/h per IP). Set `MODELSCHEMAS_API_KEY` on the workflow when you want the higher limit.
+
 Rules the generator follows:
 
 - Keep OpenRouter routing aliases (ids that start with `~`) in the OpenRouter catalog. Users can pass `chat({ model: '~anthropic/claude-haiku-latest' })`. The generated constant name maps `~` to `_`.
-- Do **not** copy those aliases into native provider files (`ai-openai`, `ai-anthropic`, `ai-gemini`, `ai-grok`). Those adapters only accept the provider's own ids.
-- For a new native-provider model, write id, modalities, and pricing. Infer features from OpenRouter `supported_parameters` when that field exists. Do **not** copy another model's tool list (`computer_use`, `google_search`, `x_search`, and similar).
-- Anthropic first-party ids use dashes (`claude-fable-5-1`). OpenRouter uses dots (`claude-fable-5.1`). The generator hyphenates Anthropic ids on insert. Do not copy the dotted OpenRouter slug into `ai-anthropic`.
+- Do **not** copy those aliases into native provider files (any of the eight packages above). Those adapters only accept the provider's own ids.
+- For a new native-provider model, write id, limits, and modalities. Infer features from the catalog capabilities: a map of flags where only `true` counts (an absent key is unknown), or a feature object, as BytePlus sends. Do **not** copy another model's tool list (`computer_use`, `google_search`, `x_search`, and similar). BytePlus inserts never claim `structured_outputs`: that gate is a live-probed list.
+- ElevenLabs music and voice-design ids are pinned to SDK unions. An id the ElevenLabs SDK does not know yet fails the type-check, so wait for the SDK. BytePlus video and image duration/size tables stay hand-curated.
+- Anthropic first-party ids use dashes (`claude-fable-5-1`). OpenRouter uses dots (`claude-fable-5.1`). The generator takes the modelschemas native id (already hyphenated) and does not copy the dotted OpenRouter slug into `ai-anthropic`. Dated Anthropic snapshots (`claude-haiku-4-5-20251001`) are skipped; the undated alias is the public id.
 - Write a row in the provider's `*ChatModelToolCapabilitiesByName` map for every new chat model, even when `supports.tools` is still `[]`. Missing that row makes `ResolveToolCapabilities` fall back to `readonly []`.
-- For new Anthropic models, infer the provider-options mix from the catalog: no sampling parameters → `AnthropicMaxTokensOptions`; `reasoning.mandatory` → `AnthropicAdaptiveOnlyThinkingOptions` plus `AnthropicOutputConfigOptions`.
+- For new Anthropic models, infer the provider-options mix from the catalog: no sampling parameters → `AnthropicMaxTokensOptions`; `reasoning.mandatory` → `AnthropicAdaptiveOnlyThinkingOptions` plus `AnthropicOutputConfigOptions`; reasoning params without sampling → adaptive-or-disabled thinking. Every new Claude model gets `AnthropicCacheControlOptions`.
+- `firstSeenAt` is when modelschemas first crawled an id, not its release date. The first sync of a new provider can need a one-off older cutoff in `scripts/.sync-models-last-run`.
 - Leave curated tools and flags on existing models alone. Edit those by hand after the sync PR opens.
 
 Do not rebase or hand-edit `automated/sync-models`. The next scheduled run force-pushes that branch from `main`. Merge generator fixes to `main` first, then let the workflow rebuild the sync PR.

docs/advanced/model-prices.md

@@ -0,0 +1,155 @@
+---
+title: Model Prices
+id: model-prices
+description: "Price each model call from its token usage, with live rate cards from modelschemas."
+keywords:
+  - tanstack ai
+  - model prices
+  - token cost
+  - modelschemas
+  - rate card
+  - onUsage
+---
+
+Get a model's rate card from [modelschemas](https://modelschemas.com) with `@modelschemas/client`. Check it with `rateCardSchema`, then price each call with `evaluateCard` from `@modelschemas/rate-card`. Prices change often, so read them live. Do not copy them into your code.
+
+<!-- ::start:tabs variant="package-manager" mode="install" -->
+
+react: @modelschemas/client @modelschemas/rate-card zod
+vue: @modelschemas/client @modelschemas/rate-card zod
+solid: @modelschemas/client @modelschemas/rate-card zod
+svelte: @modelschemas/client @modelschemas/rate-card zod
+preact: @modelschemas/client @modelschemas/rate-card zod
+angular: @modelschemas/client @modelschemas/rate-card zod
+octane: @modelschemas/client @modelschemas/rate-card zod
+vanilla: @modelschemas/client @modelschemas/rate-card zod
+
+<!-- ::end:tabs -->
+
+## Price a chat call
+
+Price each call in the [`onUsage`](./middleware-advanced#onusage) hook of a middleware, on the server:
+
+```typescript
+import { chat } from "@tanstack/ai";
+import type { ChatMiddleware, TokenUsage } from "@tanstack/ai";
+import { anthropicText } from "@tanstack/ai-anthropic";
+import { createModelschemasClient, getModel } from "@modelschemas/client";
+import { RateCardError, evaluateCard } from "@modelschemas/rate-card/evaluator";
+import type { RateCard } from "@modelschemas/rate-card/evaluator";
+import { rateCardSchema } from "@modelschemas/rate-card/schema";
+
+const client = createModelschemasClient({ baseUrl: "https://modelschemas.com" });
+
+// Keep each rate card for one hour.
+const cards = new Map<string, { at: number; card: Promise<RateCard | undefined> }>();
+
+function rateCard(provider: string, modelId: string) {
+  const key = `${provider}/${modelId}`;
+  const hit = cards.get(key);
+  if (hit && Date.now() - hit.at < 60 * 60 * 1000) return hit.card;
+  // A model with no rate card, or only a price summary, fails the check.
+  const card = getModel({ client, path: { provider, modelId } }).then(({ data }) => {
+    const parsed = rateCardSchema.safeParse(data?.pricing);
+    return parsed.success ? parsed.data : undefined;
+  });
+  cards.set(key, { at: Date.now(), card });
+  return card;
+}
+
+// A rate card counts each kind of input token once. `promptTokens` holds them all.
+function cardUsage(usage: TokenUsage) {
+  const cached = usage.promptTokensDetails?.cachedTokens ?? 0;
+  const written = usage.promptTokensDetails?.cacheWriteTokens ?? 0;
+  const written1h = usage.promptTokensDetails?.cacheWrite1hTokens ?? 0;
+  const cacheCounts = {
+    cache_read_tokens: cached,
+    cache_write_tokens: written - written1h,
+    cache_write_1h_tokens: written1h,
+  };
+  return {
+    input_tokens: usage.promptTokens - cached - written,
+    output_tokens: usage.completionTokens,
+    // A card without a cache price rejects a cache count, even a 0.
+    ...Object.fromEntries(Object.entries(cacheCounts).filter(([, count]) => count > 0)),
+  };
+}
+
+const costLog: ChatMiddleware = {
+  name: "cost-log",
+  onUsage: async (ctx, usage) => {
+    const card = await rateCard("anthropic", ctx.model);
+    if (!card) return console.log(`${ctx.model}: no price on modelschemas`);
+    try {
+      const { amount, currency } = evaluateCard(card, { usage: cardUsage(usage) });
+      const price = new Intl.NumberFormat("en", { style: "currency", currency, maximumFractionDigits: 6 });
+      console.log(`${ctx.model}: ${price.format(amount)}`);
+    } catch (error) {
+      if (!(error instanceof RateCardError)) throw error;
+      console.log(`${ctx.model}: price not known (${error.message})`);
+    }
+  },
+};
+
+const stream = chat({
+  adapter: anthropicText("claude-sonnet-5-5"),
+  messages: [{ role: "user", content: "Hello!" }],
+  threadId: "thread-1",
+  middleware: [costLog],
+});
+```
+
+The log shows the cost of each model call, for example `claude-sonnet-5-5: $0.013`. A card uses long-context tiers when the input is long enough.
+
+## When the price is not known
+
+The rate card does not guess:
+
+- `rateCardSchema.safeParse` fails when modelschemas has no rate card for the model.
+- `evaluateCard` throws `RateCardError` when the call used tokens that the card has no price for.
+- `currency` is the card's own currency, for example `USD` or `CNY`. Do not add amounts in two currencies.
+
+modelschemas allows 60 calls an hour without a key. With the one-hour cache above, that is one call for each model. For a higher limit, pass `apiKey` to `createModelschemasClient`.
+
+## Price an image before you make it
+
+A media card prices the request, so you can show the cost before you call the model:
+
+```typescript
+import { createModelschemasClient, getModel } from "@modelschemas/client";
+import { evaluateCard } from "@modelschemas/rate-card/evaluator";
+import { rateCardSchema } from "@modelschemas/rate-card/schema";
+
+const client = createModelschemasClient({ baseUrl: "https://modelschemas.com" });
+
+const { data } = await getModel({
+  client,
+  path: { provider: "fal", modelId: "alibaba/qwen-image-3/edit" },
+});
+const parsed = rateCardSchema.safeParse(data?.pricing);
+if (parsed.success) {
+  const { amount, currency } = evaluateCard(parsed.data, {
+    request: { num_images: 2 },
+  });
+  console.log(`Two images cost ${amount} ${currency}`);
+}
+```
+
+## Provider ids
+
+Pass the modelschemas provider id, not the package name:
+
+| Package | modelschemas provider |
+| --- | --- |
+| `@tanstack/ai-anthropic` | `anthropic` |
+| `@tanstack/ai-openai` | `openai`, or `azure` for Azure OpenAI |
+| `@tanstack/ai-gemini` | `gemini` |
+| `@tanstack/ai-vertex` | `google-vertex` |
+| `@tanstack/ai-grok` | `grok` |
+| `@tanstack/ai-groq` | `groq` |
+| `@tanstack/ai-mistral` | `mistral` |
+| `@tanstack/ai-bedrock` | `amazon-bedrock` |
+| `@tanstack/ai-openrouter` | `openrouter` |
+| `@tanstack/ai-vercel-gateway` | `vercel` |
+| `@tanstack/ai-cloudflare` | `cloudflare-workers-ai` |
+| `@tanstack/ai-fal` | `fal` |

docs/advanced/prompt-caching.md

@@ -76,6 +76,8 @@ const stream = chat({
 
 From the second call of a conversation, most of the input comes from the cache.
 
+To turn this usage into a price, see [Model Prices](./model-prices).
+
 ## How it works
 
 - The cache key is the first of these that you set:

docs/community-adapters/guide.md

@@ -342,7 +342,7 @@ const EFFORT_NONE_LOW_TO_XHIGH = {
 
 const GPT5_2 = {
   name: 'gpt-5.2',
-  // ...supports, pricing
+  // ...supports
   reasoning: EFFORT_NONE_LOW_TO_XHIGH,
 } as const
 

docs/config.json

@@ -1650,7 +1650,12 @@
           "label": "Prompt Caching",
           "to": "advanced/prompt-caching",
           "addedAt": "2026-10-09",
-          "updatedAt": "2026-10-22"
+          "updatedAt": "2026-10-26"
+        },
+        {
+          "label": "Model Prices",
+          "to": "advanced/model-prices",
+          "addedAt": "2026-10-26"
         },
         {
           "label": "Mid-Conversation Changes",

kiira.config.ts

@@ -58,6 +58,8 @@ export default defineConfig({
     '@cloudflare/workers-types': '^4.20260317.1',
     '@decartai/tanstack-ai-adapter': '^0.0.2',
     '@mynthio/tanstack-ai-adapter': '^0.0.14',
+    // Live model prices (docs/advanced/model-prices.md)
+    '@modelschemas/rate-card': '^0.2.1',
     '@cencori/ai-sdk': '^0.4.0',
     '@soniox/tanstack-ai-adapter': '^0.1.2',
     '@upstash/agentkit-tanstack-ai': '^0.1.1',

package.json

@@ -72,6 +72,7 @@
     "@changesets/changelog-github": "^0.7.0",
     "@changesets/cli": "^2.30.0",
     "@faker-js/faker": "^10.1.0",
+    "@modelschemas/client": "^0.2.1",
     "@stylistic/eslint-plugin": "^5.10.0",
     "@tanstack/ai": "workspace:*",
     "@tanstack/ai-grok-build": "workspace:*",