smoltalk 0.8.3 → 0.8.4

package/README.md

@@ -335,6 +335,78 @@ a `hostedTools` catalog (`getHostedTools()`); the published file is kept current
 by a daily CI job that translates [models.dev](https://models.dev) into
 smoltalk's shape.
 
+## Custom models & pricing
+
+If you use a model that isn't in smoltalk's baked-in catalog (a self-hosted
+model, a brand-new release, an OpenAI-compatible endpoint), smoltalk has no
+pricing for it and the `cost` field is simply omitted from the result — nothing
+errors, you just get `usage` without `cost`. Teach it the price and cost
+tracking starts working.
+
+**One model — `registerTextModel` (recommended).** Register once at startup:
+
+```ts
+import { registerTextModel, textSync } from "smoltalk";
+
+registerTextModel({
+  modelName: "my-model",
+  provider: "openai-compat", // must match the provider you call with (see below)
+  inputTokenCost: 0.5, // USD per 1M input tokens
+  outputTokenCost: 1.5, // USD per 1M output tokens
+  cachedInputTokenCost: 0.05, // optional
+  cacheCreationInputTokenCost: 0.625, // optional
+  maxInputTokens: 128000, // required by the type, even if you only want pricing
+  maxOutputTokens: 8192,
+});
+
+const res = await textSync({
+  model: "my-model",
+  provider: "openai-compat",
+  baseUrl: { openAiCompat: "https://my-endpoint/v1" },
+  messages,
+});
+// res.value.cost is now populated from the rates above.
+```
+
+**Per-call only — `config.modelData`.** When you can't register globally (e.g.
+per-tenant rates), pass a minimal blob for a single call. It layers over the
+baseline exactly like a refresh blob:
+
+```ts
+import { textSync, type ModelDataBlob } from "smoltalk";
+
+const modelData: ModelDataBlob = {
+  schemaVersion: 1,
+  generatedAt: new Date().toISOString(),
+  hostedTools: [],
+  models: [
+    {
+      type: "text",
+      modelName: "my-model",
+      provider: "openai-compat",
+      maxInputTokens: 128000,
+      maxOutputTokens: 8192,
+      inputTokenCost: 0.5,
+      outputTokenCost: 1.5,
+    },
+  ],
+};
+
+await textSync({ model: "my-model", provider: "openai-compat", messages, modelData });
+```
+
+**The provider must match.** The registry is keyed by `provider:modelName`, so
+the `provider` you register (or put in the blob) has to equal the `provider` you
+pass at call time. Registering `my-model` under `"openai-compat"` but calling it
+with `provider: "openrouter"` looks up a different key, finds no price, and
+silently drops the `cost` field. When in doubt, register under the same provider
+string you call with.
+
+Overriding an entry that *is* in the catalog works the same way — merges are
+field-by-field (see "Refreshing model data" above), so registering just
+`inputTokenCost` / `outputTokenCost` for a known model updates only those fields
+and leaves its limits and capabilities intact.
+
 ## Hosted tools catalog
 
 Each cloud provider offers server-side "hosted" tools (web search, code

package/dist/models.d.ts

@@ -1228,6 +1228,35 @@ export declare const textModels: readonly [{
     readonly costUnit: "characters";
     readonly disabled: true;
     readonly provider: "google";
+}, {
+    readonly type: "text";
+    readonly modelName: "claude-fable-5";
+    readonly description: "Anthropic's most capable widely released model, for the most demanding reasoning and long-horizon agentic work. Thinking is always on (cannot be disabled); the raw chain of thought is never returned. Requires 30-day data retention (not available under zero data retention). 1M context window, 128K max output.";
+    readonly maxInputTokens: 1000000;
+    readonly maxOutputTokens: 128000;
+    readonly inputTokenCost: 10;
+    readonly cachedInputTokenCost: 1;
+    readonly cacheCreationInputTokenCost: 12.5;
+    readonly outputTokenCost: 50;
+    readonly reasoning: {
+        readonly thinkingStyle: "adaptive";
+        readonly levels: readonly ["low", "medium", "high", "xhigh", "max"];
+        readonly defaultLevel: "high";
+        readonly canDisable: false;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly modalities: {
+        readonly input: readonly ["text", "image", "pdf"];
+        readonly output: readonly ["text"];
+    };
+    readonly knowledge: "2026-01";
+    readonly releaseDate: "2026-06-09";
+    readonly lastUpdated: "2026-06-09";
+    readonly family: "claude-fable";
+    readonly openWeights: false;
+    readonly temperatureSupported: false;
+    readonly provider: "anthropic";
 }, {
     readonly type: "text";
     readonly modelName: "claude-opus-4-8";
@@ -1310,6 +1339,35 @@ export declare const textModels: readonly [{
     readonly openWeights: false;
     readonly temperatureSupported: true;
     readonly provider: "anthropic";
+}, {
+    readonly type: "text";
+    readonly modelName: "claude-sonnet-5";
+    readonly description: "The best combination of speed and intelligence in the Sonnet tier, with near-Opus quality on coding and agentic work. Adaptive thinking on by default; supports the full low/medium/high/xhigh/max effort range. New tokenizer (~30% more tokens for the same text vs Sonnet 4.6). Standard pricing $3/$15; introductory $2/$10 per MTok through 2026-08-31. 1M context window, 128K max output.";
+    readonly maxInputTokens: 1000000;
+    readonly maxOutputTokens: 128000;
+    readonly inputTokenCost: 3;
+    readonly cachedInputTokenCost: 0.3;
+    readonly cacheCreationInputTokenCost: 3.75;
+    readonly outputTokenCost: 15;
+    readonly reasoning: {
+        readonly thinkingStyle: "adaptive";
+        readonly levels: readonly ["low", "medium", "high", "xhigh", "max"];
+        readonly defaultLevel: "high";
+        readonly canDisable: true;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly modalities: {
+        readonly input: readonly ["text", "image", "pdf"];
+        readonly output: readonly ["text"];
+    };
+    readonly knowledge: "2026-01";
+    readonly releaseDate: "2026-06-30";
+    readonly lastUpdated: "2026-06-30";
+    readonly family: "claude-sonnet";
+    readonly openWeights: false;
+    readonly temperatureSupported: false;
+    readonly provider: "anthropic";
 }, {
     readonly type: "text";
     readonly modelName: "claude-sonnet-4-6";

package/dist/models.js

@@ -1206,6 +1206,36 @@ export const textModels = [
         disabled: true,
         provider: "google",
     },
+    {
+        type: "text",
+        modelName: "claude-fable-5",
+        description: "Anthropic's most capable widely released model, for the most demanding reasoning and long-horizon agentic work. Thinking is always on (cannot be disabled); the raw chain of thought is never returned. Requires 30-day data retention (not available under zero data retention). 1M context window, 128K max output.",
+        maxInputTokens: 1000000,
+        maxOutputTokens: 128000,
+        inputTokenCost: 10,
+        cachedInputTokenCost: 1,
+        cacheCreationInputTokenCost: 12.5,
+        outputTokenCost: 50,
+        reasoning: {
+            thinkingStyle: "adaptive",
+            levels: ["low", "medium", "high", "xhigh", "max"],
+            defaultLevel: "high",
+            canDisable: false,
+            outputsThinking: true,
+            outputsSignatures: true,
+        },
+        modalities: {
+            input: ["text", "image", "pdf"],
+            output: ["text"],
+        },
+        knowledge: "2026-01",
+        releaseDate: "2026-06-09",
+        lastUpdated: "2026-06-09",
+        family: "claude-fable",
+        openWeights: false,
+        temperatureSupported: false,
+        provider: "anthropic",
+    },
     {
         type: "text",
         modelName: "claude-opus-4-8",
@@ -1291,6 +1321,36 @@ export const textModels = [
         temperatureSupported: true,
         provider: "anthropic",
     },
+    {
+        type: "text",
+        modelName: "claude-sonnet-5",
+        description: "The best combination of speed and intelligence in the Sonnet tier, with near-Opus quality on coding and agentic work. Adaptive thinking on by default; supports the full low/medium/high/xhigh/max effort range. New tokenizer (~30% more tokens for the same text vs Sonnet 4.6). Standard pricing $3/$15; introductory $2/$10 per MTok through 2026-08-31. 1M context window, 128K max output.",
+        maxInputTokens: 1000000,
+        maxOutputTokens: 128000,
+        inputTokenCost: 3,
+        cachedInputTokenCost: 0.3,
+        cacheCreationInputTokenCost: 3.75,
+        outputTokenCost: 15,
+        reasoning: {
+            thinkingStyle: "adaptive",
+            levels: ["low", "medium", "high", "xhigh", "max"],
+            defaultLevel: "high",
+            canDisable: true,
+            outputsThinking: true,
+            outputsSignatures: true,
+        },
+        modalities: {
+            input: ["text", "image", "pdf"],
+            output: ["text"],
+        },
+        knowledge: "2026-01",
+        releaseDate: "2026-06-30",
+        lastUpdated: "2026-06-30",
+        family: "claude-sonnet",
+        openWeights: false,
+        temperatureSupported: false,
+        provider: "anthropic",
+    },
     {
         type: "text",
         modelName: "claude-sonnet-4-6",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "smoltalk",
-  "version": "0.8.3",
+  "version": "0.8.4",
   "description": "A common interface for LLM APIs",
   "homepage": "https://github.com/egonSchiele/smoltalk",
   "files": [