smoltalk 0.3.1 → 0.4.0

package/dist/classes/message/AssistantMessage.d.ts

@@ -29,12 +29,14 @@ export declare const AssistantMessageJSONSchema: z.ZodObject<{
         inputTokens: z.ZodNumber;
         outputTokens: z.ZodNumber;
         cachedInputTokens: z.ZodOptional<z.ZodNumber>;
+        cacheCreationInputTokens: z.ZodOptional<z.ZodNumber>;
         totalTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>>;
     cost: z.ZodOptional<z.ZodObject<{
         inputCost: z.ZodNumber;
         outputCost: z.ZodNumber;
         cachedInputCost: z.ZodOptional<z.ZodNumber>;
+        cacheCreationInputCost: z.ZodOptional<z.ZodNumber>;
         totalCost: z.ZodNumber;
         currency: z.ZodString;
     }, z.core.$strip>>;

package/dist/clients/anthropic.js

@@ -8,6 +8,82 @@ import { SmolContentPolicyError, SmolContextWindowExceededError, } from "../smol
 import { BaseClient } from "./baseClient.js";
 import { Model } from "../model.js";
 const DEFAULT_MAX_TOKENS = 4096;
+/**
+ * Attach ephemeral cache_control breakpoints to (up to) three places:
+ *   1. the last tool definition
+ *   2. the last system block (promoting system from string to array form)
+ *   3. the last block of the last user message
+ *
+ * Anthropic enforces minimum prefix sizes; smaller prefixes silently no-op.
+ */
+function applyCacheBreakpoints(req) {
+    const cc = { type: "ephemeral" };
+    // Tools: mark the last tool.
+    let tools = req.tools;
+    if (tools && tools.length > 0) {
+        const lastIdx = tools.length - 1;
+        const marked = [];
+        for (let i = 0; i < tools.length; i++) {
+            if (i === lastIdx) {
+                marked.push({ ...tools[i], cache_control: cc });
+            }
+            else {
+                marked.push(tools[i]);
+            }
+        }
+        tools = marked;
+    }
+    // System: promote string to array form so the last block can be marked.
+    let system = req.system;
+    if (typeof system === "string" && system.length > 0) {
+        system = [{ type: "text", text: system, cache_control: cc }];
+    }
+    else if (Array.isArray(system) && system.length > 0) {
+        const lastIdx = system.length - 1;
+        const marked = [];
+        for (let i = 0; i < system.length; i++) {
+            if (i === lastIdx) {
+                marked.push({ ...system[i], cache_control: cc });
+            }
+            else {
+                marked.push(system[i]);
+            }
+        }
+        system = marked;
+    }
+    // Messages: mark the last block of the last user message.
+    let messages = req.messages;
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const m = messages[i];
+        if (m.role !== "user")
+            continue;
+        let blocks;
+        if (typeof m.content === "string") {
+            blocks = [{ type: "text", text: m.content }];
+        }
+        else {
+            blocks = [...m.content];
+        }
+        if (blocks.length === 0)
+            break;
+        blocks[blocks.length - 1] = {
+            ...blocks[blocks.length - 1],
+            cache_control: cc,
+        };
+        const rebuilt = [];
+        for (let j = 0; j < messages.length; j++) {
+            if (j === i) {
+                rebuilt.push({ ...m, content: blocks });
+            }
+            else {
+                rebuilt.push(messages[j]);
+            }
+        }
+        messages = rebuilt;
+        break;
+    }
+    return { system, messages, tools };
+}
 export class SmolAnthropic extends BaseClient {
     client;
     logger;
@@ -22,11 +98,22 @@ export class SmolAnthropic extends BaseClient {
         return this.model.getModel();
     }
     calculateUsageAndCost(usageData) {
+        const cacheRead = usageData.cache_read_input_tokens ?? 0;
+        const cacheCreation = usageData.cache_creation_input_tokens ?? 0;
         const usage = {
             inputTokens: usageData.input_tokens,
             outputTokens: usageData.output_tokens,
-            totalTokens: usageData.input_tokens + usageData.output_tokens,
+            totalTokens: usageData.input_tokens +
+                cacheRead +
+                cacheCreation +
+                usageData.output_tokens,
         };
+        if (cacheRead > 0) {
+            usage.cachedInputTokens = cacheRead;
+        }
+        if (cacheCreation > 0) {
+            usage.cacheCreationInputTokens = cacheCreation;
+        }
         const cost = this.model.calculateCost(usage) ?? undefined;
         return { usage, cost };
     }
@@ -81,7 +168,12 @@ export class SmolAnthropic extends BaseClient {
                     budget_tokens: reasoningBudgetMap[config.reasoningEffort],
                 }
                 : undefined;
-        return { system, messages: anthropicMessages, tools, thinking };
+        const cachingEnabled = config.caching?.enabled !== false;
+        const baseRequest = { system, messages: anthropicMessages, tools };
+        const finalRequest = cachingEnabled
+            ? applyCacheBreakpoints(baseRequest)
+            : baseRequest;
+        return { ...finalRequest, thinking };
     }
     rethrowAsSmolError(error) {
         if (error instanceof Anthropic.APIError) {
@@ -198,10 +290,15 @@ export class SmolAnthropic extends BaseClient {
         // Track thinking blocks by index: index -> { text, signature }
         const thinkingBlockMap = new Map();
         let inputTokens = 0;
+        let cacheReadTokens = 0;
+        let cacheCreationTokens = 0;
         let outputTokens = 0;
         for await (const event of stream) {
             if (event.type === "message_start") {
-                inputTokens = event.message.usage.input_tokens;
+                const u = event.message.usage;
+                inputTokens = u.input_tokens;
+                cacheReadTokens = u.cache_read_input_tokens ?? 0;
+                cacheCreationTokens = u.cache_creation_input_tokens ?? 0;
             }
             else if (event.type === "content_block_start") {
                 if (event.content_block.type === "tool_use") {
@@ -252,6 +349,15 @@ export class SmolAnthropic extends BaseClient {
             }
             else if (event.type === "message_delta") {
                 outputTokens = event.usage.output_tokens;
+                // Defensive: in practice Anthropic only sends cache fields on
+                // message_start, but read them here too so we don't miss an
+                // update if the SDK changes.
+                if (event.usage.cache_read_input_tokens != null) {
+                    cacheReadTokens = event.usage.cache_read_input_tokens;
+                }
+                if (event.usage.cache_creation_input_tokens != null) {
+                    cacheCreationTokens = event.usage.cache_creation_input_tokens;
+                }
             }
         }
         this.logger.debug("Streaming response completed from Anthropic");
@@ -269,8 +375,14 @@ export class SmolAnthropic extends BaseClient {
         const usage = {
             inputTokens,
             outputTokens,
-            totalTokens: inputTokens + outputTokens,
+            totalTokens: inputTokens + cacheReadTokens + cacheCreationTokens + outputTokens,
         };
+        if (cacheReadTokens > 0) {
+            usage.cachedInputTokens = cacheReadTokens;
+        }
+        if (cacheCreationTokens > 0) {
+            usage.cacheCreationInputTokens = cacheCreationTokens;
+        }
         const cost = this.model.calculateCost(usage) ?? undefined;
         yield {
             type: "done",

package/dist/clients/google.js

@@ -31,12 +31,15 @@ export class SmolGoogle extends BaseClient {
         let usage;
         let cost;
         if (usageMetadata) {
+            const cached = usageMetadata.cachedContentTokenCount ?? 0;
             usage = {
-                inputTokens: usageMetadata.promptTokenCount || 0,
+                inputTokens: Math.max(0, (usageMetadata.promptTokenCount || 0) - cached),
                 outputTokens: usageMetadata.candidatesTokenCount || 0,
-                cachedInputTokens: usageMetadata.cachedContentTokenCount,
                 totalTokens: usageMetadata.totalTokenCount,
             };
+            if (cached > 0) {
+                usage.cachedInputTokens = cached;
+            }
             const calculatedCost = this.model.calculateCost(usage);
             if (calculatedCost) {
                 cost = calculatedCost;

package/dist/clients/openai.js

@@ -30,12 +30,15 @@ export class SmolOpenAi extends BaseClient {
         let usage;
         let cost;
         if (usageData) {
+            const cached = usageData.prompt_tokens_details?.cached_tokens ?? 0;
             usage = {
-                inputTokens: usageData.prompt_tokens || 0,
+                inputTokens: Math.max(0, (usageData.prompt_tokens || 0) - cached),
                 outputTokens: usageData.completion_tokens || 0,
-                cachedInputTokens: usageData.prompt_tokens_details?.cached_tokens,
                 totalTokens: usageData.total_tokens,
             };
+            if (cached > 0) {
+                usage.cachedInputTokens = cached;
+            }
             const calculatedCost = this.model.calculateCost(usage);
             if (calculatedCost) {
                 cost = calculatedCost;

package/dist/clients/openaiResponses.js

@@ -89,12 +89,15 @@ export class SmolOpenAiResponses extends BaseClient {
         let usage;
         let cost;
         if (usageData) {
+            const cached = usageData.input_tokens_details?.cached_tokens ?? 0;
             usage = {
-                inputTokens: usageData.input_tokens || 0,
+                inputTokens: Math.max(0, (usageData.input_tokens || 0) - cached),
                 outputTokens: usageData.output_tokens || 0,
-                cachedInputTokens: usageData.input_tokens_details?.cached_tokens,
                 totalTokens: usageData.total_tokens,
             };
+            if (cached > 0) {
+                usage.cachedInputTokens = cached;
+            }
             const calculatedCost = this.model.calculateCost(usage);
             if (calculatedCost) {
                 cost = calculatedCost;

package/dist/model.d.ts

@@ -11,10 +11,12 @@ export declare class Model {
         inputTokens: number;
         outputTokens: number;
         cachedInputTokens?: number;
+        cacheCreationInputTokens?: number;
     }): {
         inputCost: number;
         outputCost: number;
         cachedInputCost?: number;
+        cacheCreationInputCost?: number;
         totalCost: number;
         currency: string;
     } | null;

package/dist/model.js

@@ -26,16 +26,49 @@ export class Model {
         if (!model || !isTextModel(model)) {
             return null;
         }
+        const cachedTokens = usage.cachedInputTokens ?? 0;
+        const cacheCreationTokens = usage.cacheCreationInputTokens ?? 0;
+        // Disjoint buckets. If a discount price isn't defined for this model,
+        // the tokens were still billed by the provider — charge them at the
+        // full input rate so totalCost stays honest.
+        const cachedRate = model.cachedInputTokenCost ?? model.inputTokenCost ?? 0;
+        const cacheCreationRate = model.cacheCreationInputTokenCost ?? model.inputTokenCost ?? 0;
         const inputCost = round((usage.inputTokens * (model.inputTokenCost || 0)) / 1_000_000, 6);
         const outputCost = round((usage.outputTokens * (model.outputTokenCost || 0)) / 1_000_000, 6);
-        const cachedInputCost = usage.cachedInputTokens && model.cachedInputTokenCost
-            ? round((usage.cachedInputTokens * model.cachedInputTokenCost) / 1_000_000, 6)
-            : undefined;
-        const totalCost = round(inputCost + outputCost + (cachedInputCost || 0), 6);
+        // Only expose cachedInputCost / cacheCreationInputCost when the model
+        // actually has a distinct discount price. Otherwise, fold those dollars
+        // into inputCost so the user isn't misled by a $0 cached field.
+        let cachedInputCost;
+        let cacheCreationInputCost;
+        let foldedInputDollars = 0;
+        if (cachedTokens > 0) {
+            const dollars = (cachedTokens * cachedRate) / 1_000_000;
+            if (model.cachedInputTokenCost != null) {
+                cachedInputCost = round(dollars, 6);
+            }
+            else {
+                foldedInputDollars += dollars;
+            }
+        }
+        if (cacheCreationTokens > 0) {
+            const dollars = (cacheCreationTokens * cacheCreationRate) / 1_000_000;
+            if (model.cacheCreationInputTokenCost != null) {
+                cacheCreationInputCost = round(dollars, 6);
+            }
+            else {
+                foldedInputDollars += dollars;
+            }
+        }
+        const finalInputCost = round(inputCost + foldedInputDollars, 6);
+        const totalCost = round(finalInputCost +
+            outputCost +
+            (cachedInputCost || 0) +
+            (cacheCreationInputCost || 0), 6);
         return {
-            inputCost,
+            inputCost: finalInputCost,
             outputCost,
             cachedInputCost,
+            cacheCreationInputCost,
             totalCost,
             currency: "USD",
         };

package/dist/models.d.ts

@@ -16,6 +16,7 @@ export type BaseModel = {
     description?: string;
     inputTokenCost?: number;
     cachedInputTokenCost?: number;
+    cacheCreationInputTokenCost?: number;
     outputTokenCost?: number;
     disabled?: boolean;
     costUnit?: "tokens" | "characters" | "minutes";
@@ -466,10 +467,27 @@ export declare const textModels: readonly [{
     readonly outputTokenCost: 12;
     readonly disabled: true;
     readonly provider: "google";
+}, {
+    readonly type: "text";
+    readonly modelName: "gemini-3.5-flash";
+    readonly description: "Latest Gemini 3.5 Flash model (GA May 2026). Outperforms Gemini 3.1 Pro on coding and agentic suites at 4x the speed. 1M context window, 64K output. Context caching: $0.15/1M read.";
+    readonly maxInputTokens: 1048576;
+    readonly maxOutputTokens: 65536;
+    readonly inputTokenCost: 1.5;
+    readonly cachedInputTokenCost: 0.15;
+    readonly outputTokenCost: 9;
+    readonly reasoning: {
+        readonly levels: readonly ["minimal", "low", "medium", "high"];
+        readonly defaultLevel: "high";
+        readonly canDisable: false;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly provider: "google";
 }, {
     readonly type: "text";
     readonly modelName: "gemini-3-flash-preview";
-    readonly description: "Latest Gemini 3 flash model with 1M context window and 64K output. Outperforms 2.5 Pro while being 3x faster. Optimized for agentic workflows and coding. Includes context caching for 90% cost reductions.";
+    readonly description: "Gemini 3 Flash preview. Superseded by gemini-3.5-flash. 1M context window and 64K output. Optimized for agentic workflows and coding.";
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.5;
@@ -485,8 +503,8 @@ export declare const textModels: readonly [{
     readonly provider: "google";
 }, {
     readonly type: "text";
-    readonly modelName: "gemini-3.1-flash-lite-preview";
-    readonly description: "Most cost-effective Gemini 3.1 model with thinking support and 1M context window. 2.5x faster TTFA and 45% faster output than 2.5 Flash. Released March 2026.";
+    readonly modelName: "gemini-3.1-flash-lite";
+    readonly description: "Most cost-effective Gemini 3.1 model (GA). Thinking support, 1M context window, 64K output. 2.5x faster TTFA and 45% faster output than 2.5 Flash.";
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.25;
@@ -500,6 +518,16 @@ export declare const textModels: readonly [{
         readonly outputsSignatures: true;
     };
     readonly provider: "google";
+}, {
+    readonly type: "text";
+    readonly modelName: "gemini-3.1-flash-lite-preview";
+    readonly description: "DEPRECATED: Preview version, discontinued July 9, 2026. Use gemini-3.1-flash-lite instead.";
+    readonly maxInputTokens: 1048576;
+    readonly maxOutputTokens: 65536;
+    readonly inputTokenCost: 0.25;
+    readonly outputTokenCost: 1.5;
+    readonly disabled: true;
+    readonly provider: "google";
 }, {
     readonly type: "text";
     readonly modelName: "gemini-2.5-pro";
@@ -507,6 +535,7 @@ export declare const textModels: readonly [{
     readonly maxInputTokens: 2097152;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 1.25;
+    readonly cachedInputTokenCost: 0.31;
     readonly outputTokenCost: 10;
     readonly outputTokensPerSecond: 145;
     readonly reasoning: {
@@ -522,6 +551,7 @@ export declare const textModels: readonly [{
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.3;
+    readonly cachedInputTokenCost: 0.075;
     readonly outputTokenCost: 2.5;
     readonly outputTokensPerSecond: 245;
     readonly reasoning: {
@@ -537,6 +567,7 @@ export declare const textModels: readonly [{
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.1;
+    readonly cachedInputTokenCost: 0.025;
     readonly outputTokenCost: 0.4;
     readonly outputTokensPerSecond: 400;
     readonly reasoning: {
@@ -611,14 +642,30 @@ export declare const textModels: readonly [{
     readonly costUnit: "characters";
     readonly disabled: true;
     readonly provider: "google";
+}, {
+    readonly type: "text";
+    readonly modelName: "claude-opus-4-8";
+    readonly description: "The most capable Claude model for complex reasoning and agentic coding. Same per-token pricing as Opus 4.7 with improved tool-use efficiency (~290 tokens for tool-use system prompt vs 675 on 4.7). 1M context window, 128K max output.";
+    readonly maxInputTokens: 1000000;
+    readonly maxOutputTokens: 128000;
+    readonly inputTokenCost: 5;
+    readonly cachedInputTokenCost: 0.5;
+    readonly outputTokenCost: 25;
+    readonly reasoning: {
+        readonly canDisable: false;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly provider: "anthropic";
 }, {
     readonly type: "text";
     readonly modelName: "claude-opus-4-7";
-    readonly description: "The most capable Claude model for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.";
+    readonly description: "Claude Opus 4.7 for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.";
     readonly maxInputTokens: 1000000;
     readonly maxOutputTokens: 128000;
     readonly inputTokenCost: 5;
     readonly cachedInputTokenCost: 0.5;
+    readonly cacheCreationInputTokenCost: 6.25;
     readonly outputTokenCost: 25;
     readonly outputTokensPerSecond: 72;
     readonly reasoning: {
@@ -635,6 +682,7 @@ export declare const textModels: readonly [{
     readonly maxOutputTokens: 128000;
     readonly inputTokenCost: 5;
     readonly cachedInputTokenCost: 0.5;
+    readonly cacheCreationInputTokenCost: 6.25;
     readonly outputTokenCost: 25;
     readonly outputTokensPerSecond: 53;
     readonly reasoning: {
@@ -651,6 +699,7 @@ export declare const textModels: readonly [{
     readonly maxOutputTokens: 64000;
     readonly inputTokenCost: 3;
     readonly cachedInputTokenCost: 0.3;
+    readonly cacheCreationInputTokenCost: 3.75;
     readonly outputTokenCost: 15;
     readonly outputTokensPerSecond: 52;
     readonly reasoning: {
@@ -667,6 +716,7 @@ export declare const textModels: readonly [{
     readonly maxOutputTokens: 64000;
     readonly inputTokenCost: 1;
     readonly cachedInputTokenCost: 0.1;
+    readonly cacheCreationInputTokenCost: 1.25;
     readonly outputTokenCost: 5;
     readonly outputTokensPerSecond: 97;
     readonly reasoning: {
@@ -734,7 +784,14 @@ export declare const imageModels: readonly [{
     readonly type: "image";
     readonly modelName: "gemini-3.1-flash-image-preview";
     readonly provider: "google";
-    readonly description: "Fast image generation with Gemini 3.1 Flash. Supports resolutions from 512px to 4096px. ~$0.067/image at 1K resolution.";
+    readonly description: "DEPRECATED: Preview version. Use gemini-3.1-flash-image instead.";
+    readonly costPerImage: 0.067;
+    readonly disabled: true;
+}, {
+    readonly type: "image";
+    readonly modelName: "gemini-3.1-flash-image";
+    readonly provider: "google";
+    readonly description: "Fast image generation with Gemini 3.1 Flash (GA). Supports resolutions from 512px to 4096px. ~$0.045/image at 512px, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K.";
     readonly costPerImage: 0.067;
 }];
 export declare const embeddingsModels: EmbeddingsModel[];
@@ -1156,10 +1213,27 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly outputTokenCost: 12;
     readonly disabled: true;
     readonly provider: "google";
+} | {
+    readonly type: "text";
+    readonly modelName: "gemini-3.5-flash";
+    readonly description: "Latest Gemini 3.5 Flash model (GA May 2026). Outperforms Gemini 3.1 Pro on coding and agentic suites at 4x the speed. 1M context window, 64K output. Context caching: $0.15/1M read.";
+    readonly maxInputTokens: 1048576;
+    readonly maxOutputTokens: 65536;
+    readonly inputTokenCost: 1.5;
+    readonly cachedInputTokenCost: 0.15;
+    readonly outputTokenCost: 9;
+    readonly reasoning: {
+        readonly levels: readonly ["minimal", "low", "medium", "high"];
+        readonly defaultLevel: "high";
+        readonly canDisable: false;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly provider: "google";
 } | {
     readonly type: "text";
     readonly modelName: "gemini-3-flash-preview";
-    readonly description: "Latest Gemini 3 flash model with 1M context window and 64K output. Outperforms 2.5 Pro while being 3x faster. Optimized for agentic workflows and coding. Includes context caching for 90% cost reductions.";
+    readonly description: "Gemini 3 Flash preview. Superseded by gemini-3.5-flash. 1M context window and 64K output. Optimized for agentic workflows and coding.";
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.5;
@@ -1175,8 +1249,8 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly provider: "google";
 } | {
     readonly type: "text";
-    readonly modelName: "gemini-3.1-flash-lite-preview";
-    readonly description: "Most cost-effective Gemini 3.1 model with thinking support and 1M context window. 2.5x faster TTFA and 45% faster output than 2.5 Flash. Released March 2026.";
+    readonly modelName: "gemini-3.1-flash-lite";
+    readonly description: "Most cost-effective Gemini 3.1 model (GA). Thinking support, 1M context window, 64K output. 2.5x faster TTFA and 45% faster output than 2.5 Flash.";
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.25;
@@ -1190,6 +1264,16 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
         readonly outputsSignatures: true;
     };
     readonly provider: "google";
+} | {
+    readonly type: "text";
+    readonly modelName: "gemini-3.1-flash-lite-preview";
+    readonly description: "DEPRECATED: Preview version, discontinued July 9, 2026. Use gemini-3.1-flash-lite instead.";
+    readonly maxInputTokens: 1048576;
+    readonly maxOutputTokens: 65536;
+    readonly inputTokenCost: 0.25;
+    readonly outputTokenCost: 1.5;
+    readonly disabled: true;
+    readonly provider: "google";
 } | {
     readonly type: "text";
     readonly modelName: "gemini-2.5-pro";
@@ -1197,6 +1281,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxInputTokens: 2097152;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 1.25;
+    readonly cachedInputTokenCost: 0.31;
     readonly outputTokenCost: 10;
     readonly outputTokensPerSecond: 145;
     readonly reasoning: {
@@ -1212,6 +1297,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.3;
+    readonly cachedInputTokenCost: 0.075;
     readonly outputTokenCost: 2.5;
     readonly outputTokensPerSecond: 245;
     readonly reasoning: {
@@ -1227,6 +1313,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxInputTokens: 1048576;
     readonly maxOutputTokens: 65536;
     readonly inputTokenCost: 0.1;
+    readonly cachedInputTokenCost: 0.025;
     readonly outputTokenCost: 0.4;
     readonly outputTokensPerSecond: 400;
     readonly reasoning: {
@@ -1301,14 +1388,30 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly costUnit: "characters";
     readonly disabled: true;
     readonly provider: "google";
+} | {
+    readonly type: "text";
+    readonly modelName: "claude-opus-4-8";
+    readonly description: "The most capable Claude model for complex reasoning and agentic coding. Same per-token pricing as Opus 4.7 with improved tool-use efficiency (~290 tokens for tool-use system prompt vs 675 on 4.7). 1M context window, 128K max output.";
+    readonly maxInputTokens: 1000000;
+    readonly maxOutputTokens: 128000;
+    readonly inputTokenCost: 5;
+    readonly cachedInputTokenCost: 0.5;
+    readonly outputTokenCost: 25;
+    readonly reasoning: {
+        readonly canDisable: false;
+        readonly outputsThinking: true;
+        readonly outputsSignatures: true;
+    };
+    readonly provider: "anthropic";
 } | {
     readonly type: "text";
     readonly modelName: "claude-opus-4-7";
-    readonly description: "The most capable Claude model for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.";
+    readonly description: "Claude Opus 4.7 for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.";
     readonly maxInputTokens: 1000000;
     readonly maxOutputTokens: 128000;
     readonly inputTokenCost: 5;
     readonly cachedInputTokenCost: 0.5;
+    readonly cacheCreationInputTokenCost: 6.25;
     readonly outputTokenCost: 25;
     readonly outputTokensPerSecond: 72;
     readonly reasoning: {
@@ -1325,6 +1428,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxOutputTokens: 128000;
     readonly inputTokenCost: 5;
     readonly cachedInputTokenCost: 0.5;
+    readonly cacheCreationInputTokenCost: 6.25;
     readonly outputTokenCost: 25;
     readonly outputTokensPerSecond: 53;
     readonly reasoning: {
@@ -1341,6 +1445,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxOutputTokens: 64000;
     readonly inputTokenCost: 3;
     readonly cachedInputTokenCost: 0.3;
+    readonly cacheCreationInputTokenCost: 3.75;
     readonly outputTokenCost: 15;
     readonly outputTokensPerSecond: 52;
     readonly reasoning: {
@@ -1357,6 +1462,7 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly maxOutputTokens: 64000;
     readonly inputTokenCost: 1;
     readonly cachedInputTokenCost: 0.1;
+    readonly cacheCreationInputTokenCost: 1.25;
     readonly outputTokenCost: 5;
     readonly outputTokensPerSecond: 97;
     readonly reasoning: {
@@ -1423,7 +1529,14 @@ export declare function getModel(modelName: ModelName): TextModel | EmbeddingsMo
     readonly type: "image";
     readonly modelName: "gemini-3.1-flash-image-preview";
     readonly provider: "google";
-    readonly description: "Fast image generation with Gemini 3.1 Flash. Supports resolutions from 512px to 4096px. ~$0.067/image at 1K resolution.";
+    readonly description: "DEPRECATED: Preview version. Use gemini-3.1-flash-image instead.";
+    readonly costPerImage: 0.067;
+    readonly disabled: true;
+} | {
+    readonly type: "image";
+    readonly modelName: "gemini-3.1-flash-image";
+    readonly provider: "google";
+    readonly description: "Fast image generation with Gemini 3.1 Flash (GA). Supports resolutions from 512px to 4096px. ~$0.045/image at 512px, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K.";
     readonly costPerImage: 0.067;
 } | undefined;
 export declare function isImageModel(model: ModelType): model is ImageModel;

package/dist/models.js

@@ -459,10 +459,28 @@ export const textModels = [
         disabled: true,
         provider: "google",
     },
+    {
+        type: "text",
+        modelName: "gemini-3.5-flash",
+        description: "Latest Gemini 3.5 Flash model (GA May 2026). Outperforms Gemini 3.1 Pro on coding and agentic suites at 4x the speed. 1M context window, 64K output. Context caching: $0.15/1M read.",
+        maxInputTokens: 1_048_576,
+        maxOutputTokens: 65536,
+        inputTokenCost: 1.5,
+        cachedInputTokenCost: 0.15,
+        outputTokenCost: 9.0,
+        reasoning: {
+            levels: ["minimal", "low", "medium", "high"],
+            defaultLevel: "high",
+            canDisable: false,
+            outputsThinking: true,
+            outputsSignatures: true,
+        },
+        provider: "google",
+    },
     {
         type: "text",
         modelName: "gemini-3-flash-preview",
-        description: "Latest Gemini 3 flash model with 1M context window and 64K output. Outperforms 2.5 Pro while being 3x faster. Optimized for agentic workflows and coding. Includes context caching for 90% cost reductions.",
+        description: "Gemini 3 Flash preview. Superseded by gemini-3.5-flash. 1M context window and 64K output. Optimized for agentic workflows and coding.",
         maxInputTokens: 1_048_576,
         maxOutputTokens: 65536,
         inputTokenCost: 0.5,
@@ -479,8 +497,8 @@ export const textModels = [
     },
     {
         type: "text",
-        modelName: "gemini-3.1-flash-lite-preview",
-        description: "Most cost-effective Gemini 3.1 model with thinking support and 1M context window. 2.5x faster TTFA and 45% faster output than 2.5 Flash. Released March 2026.",
+        modelName: "gemini-3.1-flash-lite",
+        description: "Most cost-effective Gemini 3.1 model (GA). Thinking support, 1M context window, 64K output. 2.5x faster TTFA and 45% faster output than 2.5 Flash.",
         maxInputTokens: 1_048_576,
         maxOutputTokens: 65536,
         inputTokenCost: 0.25,
@@ -495,6 +513,17 @@ export const textModels = [
         },
         provider: "google",
     },
+    {
+        type: "text",
+        modelName: "gemini-3.1-flash-lite-preview",
+        description: "DEPRECATED: Preview version, discontinued July 9, 2026. Use gemini-3.1-flash-lite instead.",
+        maxInputTokens: 1_048_576,
+        maxOutputTokens: 65536,
+        inputTokenCost: 0.25,
+        outputTokenCost: 1.5,
+        disabled: true,
+        provider: "google",
+    },
     {
         type: "text",
         modelName: "gemini-2.5-pro",
@@ -502,6 +531,7 @@ export const textModels = [
         maxInputTokens: 2_097_152,
         maxOutputTokens: 65536,
         inputTokenCost: 1.25,
+        cachedInputTokenCost: 0.31,
         outputTokenCost: 10.0,
         outputTokensPerSecond: 145,
         reasoning: {
@@ -518,6 +548,7 @@ export const textModels = [
         maxInputTokens: 1_048_576,
         maxOutputTokens: 65536,
         inputTokenCost: 0.3,
+        cachedInputTokenCost: 0.075,
         outputTokenCost: 2.5,
         outputTokensPerSecond: 245,
         reasoning: {
@@ -534,6 +565,7 @@ export const textModels = [
         maxInputTokens: 1_048_576,
         maxOutputTokens: 65536,
         inputTokenCost: 0.1,
+        cachedInputTokenCost: 0.025,
         outputTokenCost: 0.4,
         outputTokensPerSecond: 400,
         reasoning: {
@@ -615,14 +647,31 @@ export const textModels = [
         disabled: true,
         provider: "google",
     },
+    {
+        type: "text",
+        modelName: "claude-opus-4-8",
+        description: "The most capable Claude model for complex reasoning and agentic coding. Same per-token pricing as Opus 4.7 with improved tool-use efficiency (~290 tokens for tool-use system prompt vs 675 on 4.7). 1M context window, 128K max output.",
+        maxInputTokens: 1_000_000,
+        maxOutputTokens: 128_000,
+        inputTokenCost: 5,
+        cachedInputTokenCost: 0.5,
+        outputTokenCost: 25,
+        reasoning: {
+            canDisable: false,
+            outputsThinking: true,
+            outputsSignatures: true,
+        },
+        provider: "anthropic",
+    },
     {
         type: "text",
         modelName: "claude-opus-4-7",
-        description: "The most capable Claude model for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.",
+        description: "Claude Opus 4.7 for complex reasoning and agentic coding. Features Adaptive Thinking that auto-tunes reasoning compute per request. 1M context window, 128K max output. Knowledge cutoff: January 2026.",
         maxInputTokens: 1_000_000,
         maxOutputTokens: 128_000,
         inputTokenCost: 5,
         cachedInputTokenCost: 0.5,
+        cacheCreationInputTokenCost: 6.25,
         outputTokenCost: 25,
         outputTokensPerSecond: 72,
         reasoning: {
@@ -640,6 +689,7 @@ export const textModels = [
         maxOutputTokens: 128_000,
         inputTokenCost: 5,
         cachedInputTokenCost: 0.5,
+        cacheCreationInputTokenCost: 6.25,
         outputTokenCost: 25,
         outputTokensPerSecond: 53,
         reasoning: {
@@ -657,6 +707,7 @@ export const textModels = [
         maxOutputTokens: 64_000,
         inputTokenCost: 3,
         cachedInputTokenCost: 0.3,
+        cacheCreationInputTokenCost: 3.75,
         outputTokenCost: 15,
         outputTokensPerSecond: 52,
         reasoning: {
@@ -674,6 +725,7 @@ export const textModels = [
         maxOutputTokens: 64_000,
         inputTokenCost: 1,
         cachedInputTokenCost: 0.1,
+        cacheCreationInputTokenCost: 1.25,
         outputTokenCost: 5,
         outputTokensPerSecond: 97,
         reasoning: {
@@ -751,7 +803,15 @@ export const imageModels = [
         type: "image",
         modelName: "gemini-3.1-flash-image-preview",
         provider: "google",
-        description: "Fast image generation with Gemini 3.1 Flash. Supports resolutions from 512px to 4096px. ~$0.067/image at 1K resolution.",
+        description: "DEPRECATED: Preview version. Use gemini-3.1-flash-image instead.",
+        costPerImage: 0.067,
+        disabled: true,
+    },
+    {
+        type: "image",
+        modelName: "gemini-3.1-flash-image",
+        provider: "google",
+        description: "Fast image generation with Gemini 3.1 Flash (GA). Supports resolutions from 512px to 4096px. ~$0.045/image at 512px, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K.",
         costPerImage: 0.067,
     },
 ];

package/dist/types/costEstimate.d.ts

@@ -3,6 +3,7 @@ export type CostEstimate = {
     inputCost: number;
     outputCost: number;
     cachedInputCost?: number;
+    cacheCreationInputCost?: number;
     totalCost: number;
     currency: string;
 };
@@ -10,6 +11,7 @@ export declare const CostEstimateSchema: z.ZodObject<{
     inputCost: z.ZodNumber;
     outputCost: z.ZodNumber;
     cachedInputCost: z.ZodOptional<z.ZodNumber>;
+    cacheCreationInputCost: z.ZodOptional<z.ZodNumber>;
     totalCost: z.ZodNumber;
     currency: z.ZodString;
 }, z.core.$strip>;

package/dist/types/costEstimate.js

@@ -3,6 +3,7 @@ export const CostEstimateSchema = z.object({
     inputCost: z.number(),
     outputCost: z.number(),
     cachedInputCost: z.number().optional(),
+    cacheCreationInputCost: z.number().optional(),
     totalCost: z.number(),
     currency: z.string(),
 });
@@ -24,6 +25,7 @@ export function addCosts(_a, _b) {
         inputCost: a.inputCost + b.inputCost,
         outputCost: a.outputCost + b.outputCost,
         cachedInputCost: (a.cachedInputCost || 0) + (b.cachedInputCost || 0),
+        cacheCreationInputCost: (a.cacheCreationInputCost || 0) + (b.cacheCreationInputCost || 0),
         totalCost: a.totalCost + b.totalCost,
         currency: a.currency,
     };

package/dist/types/tokenUsage.d.ts

@@ -3,12 +3,14 @@ export type TokenUsage = {
     inputTokens: number;
     outputTokens: number;
     cachedInputTokens?: number;
+    cacheCreationInputTokens?: number;
     totalTokens?: number;
 };
 export declare const TokenUsageSchema: z.ZodObject<{
     inputTokens: z.ZodNumber;
     outputTokens: z.ZodNumber;
     cachedInputTokens: z.ZodOptional<z.ZodNumber>;
+    cacheCreationInputTokens: z.ZodOptional<z.ZodNumber>;
     totalTokens: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strip>;
 export declare function addTokenUsage(_a?: TokenUsage, _b?: TokenUsage): TokenUsage;

package/dist/types/tokenUsage.js

@@ -3,6 +3,7 @@ export const TokenUsageSchema = z.object({
     inputTokens: z.number(),
     outputTokens: z.number(),
     cachedInputTokens: z.number().optional(),
+    cacheCreationInputTokens: z.number().optional(),
     totalTokens: z.number().optional(),
 });
 export function addTokenUsage(_a, _b) {
@@ -20,6 +21,7 @@ export function addTokenUsage(_a, _b) {
         inputTokens: a.inputTokens + b.inputTokens,
         outputTokens: a.outputTokens + b.outputTokens,
         cachedInputTokens: (a.cachedInputTokens || 0) + (b.cachedInputTokens || 0),
+        cacheCreationInputTokens: (a.cacheCreationInputTokens || 0) + (b.cacheCreationInputTokens || 0),
         totalTokens: (a.totalTokens || 0) + (b.totalTokens || 0),
     };
 }

package/dist/types.d.ts

@@ -69,6 +69,10 @@ export type SmolConfig = {
         enabled: boolean;
         budgetTokens?: number;
     };
+    /** Prompt caching. Currently used by Anthropic; OpenAI/Google cache automatically. Defaults to enabled. */
+    caching?: {
+        enabled?: boolean;
+    };
     /** Provider-agnostic reasoning effort level. */
     reasoningEffort?: "low" | "medium" | "high";
     responseFormatOptions?: Partial<{

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "smoltalk",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "A common interface for LLM APIs",
   "homepage": "https://github.com/egonSchiele/smoltalk",
   "files": [