🤖 feat: add support for Claude Sonnet 5

docs/config/models.mdx

@@ -15,7 +15,7 @@ Mux ships with curated models kept up to date with the frontier. Use any custom
 | ---------------------- | ----------------------------- | ------------------------------------------------------------ | ------- |
 | Fable 5                | anthropic:claude-fable-5      | `fable`                                                      |         |
 | Opus 4.8               | anthropic:claude-opus-4-8     | `opus`                                                       | ✓       |
-| Sonnet 4.6             | anthropic:claude-sonnet-4-6   | `sonnet`                                                     |         |
+| Sonnet 5               | anthropic:claude-sonnet-5     | `sonnet`                                                     |         |
 | Haiku 4.5              | anthropic:claude-haiku-4-5    | `haiku`                                                      |         |
 | GPT-5.5                | openai:gpt-5.5                | `gpt`, `gpt-5.5`                                             |         |
 | GPT-5.5 Pro            | openai:gpt-5.5-pro            | `gpt-pro`, `gpt-5.5-pro`                                     |         |

src/browser/utils/slashCommands/parser.test.ts

@@ -62,7 +62,7 @@ describe("commandParser", () => {
     });
 
     it("should parse /model with full provider:model format", () => {
-      expectModelSet("/model anthropic:claude-sonnet-4-6", KNOWN_MODELS.SONNET.id);
+      expectModelSet("/model anthropic:claude-sonnet-5", KNOWN_MODELS.SONNET.id);
     });
 
     it("should parse /compact -m with alias", () => {

src/common/constants/knownModels.ts

@@ -49,12 +49,18 @@ const MODEL_DEFINITIONS = {
     // (Opus 4.6/4.7 also reused 4.5 — tokenization is unchanged across the 4.x line).
     tokenizerOverride: "anthropic/claude-opus-4.5",
   },
+  // Claude Sonnet 5 - released June 30, 2026. The most agentic Sonnet yet (native 1M context,
+  // 128K max output, adaptive thinking + effort including native xhigh). Standard pricing matches
+  // Sonnet 4.6 ($3/M in, $15/M out); introductory $2/$10 applies through Aug 31, 2026. API id
+  // `claude-sonnet-5`. The bare `sonnet` alias tracks the latest Sonnet tier.
   SONNET: {
     provider: "anthropic",
-    providerModelId: "claude-sonnet-4-6",
+    providerModelId: "claude-sonnet-5",
     aliases: ["sonnet"],
     warm: true,
-    // Sonnet 4.6 tokenizer not yet available upstream; reuse 4.5 for approximate counting
+    // Sonnet 5 ships an updated tokenizer (same kind of change introduced with Opus 4.7) that
+    // isn't published upstream yet; reuse Sonnet 4.5 for approximate counting. Real usage can run
+    // ~1.0-1.35x higher than this estimate depending on content type.
     tokenizerOverride: "anthropic/claude-sonnet-4.5",
   },
   HAIKU: {

src/common/types/thinking.ts

@@ -27,9 +27,10 @@ export const THINKING_DISPLAY_LABELS: Record<ThinkingLevel, string> = {
 /**
  * Display label for thinking levels, with provider-aware xhigh labeling.
  *
- * Models with a native "xhigh" effort level (OpenAI, Anthropic Opus 4.7+) show
- * "XHIGH" for the xhigh ThinkingLevel; on those providers xhigh and max are
- * distinct. On Opus 4.6 (where xhigh maps to max effort), it shows "MAX".
+ * Models with a native "xhigh" effort level (OpenAI, Anthropic Opus 4.7+,
+ * Anthropic Sonnet 5+) show "XHIGH" for the xhigh ThinkingLevel; on those
+ * providers xhigh and max are distinct. On Opus/Sonnet 4.6 (where xhigh maps
+ * to max effort), it shows "MAX".
  * Medium always displays as "MED".
  */
 export function getThinkingDisplayLabel(level: ThinkingLevel, modelString?: string): string {
@@ -164,10 +165,10 @@ export const ANTHROPIC_THINKING_BUDGETS: Record<ThinkingLevel, number> = {
 /**
  * Anthropic effort type - matches SDK's AnthropicProviderOptions["effort"].
  *
- * Note: Opus 4.7 introduced a native "xhigh" effort level in the API, but the
- * SDK's Zod validator still rejects "xhigh". Mux handles this by sending "max"
- * through the SDK and rewriting `output_config.effort` to "xhigh" in a fetch
- * wrapper for Opus 4.7+ when the user selected the xhigh ThinkingLevel.
+ * Note: Opus 4.7 and Sonnet 5 introduced a native "xhigh" effort level in the API,
+ * but the SDK's Zod validator still rejects "xhigh". Mux handles this by sending
+ * "max" through the SDK and rewriting `output_config.effort` to "xhigh" in a fetch
+ * wrapper for native-xhigh Anthropic models when the user selected the xhigh ThinkingLevel.
  * See `wrapFetchWithAnthropicCacheControl` and `buildRequestHeaders`.
  */
 export type AnthropicEffortLevel = "low" | "medium" | "high" | "max";
@@ -178,18 +179,18 @@ export type AnthropicEffortLevel = "low" | "medium" | "high" | "max";
  * The effort parameter controls how much computational work the model applies.
  * - Opus 4.5 supports: low, medium, high (policy clamps xhigh → high)
  * - Opus 4.6 supports: low, medium, high, max (xhigh maps to "max" effort)
- * - Opus 4.7 supports: low, medium, high, xhigh, max (xhigh requires wire override)
+ * - Opus 4.7+ and Sonnet 5+ support: low, medium, high, xhigh, max (xhigh requires wire override)
  *
  * Because the @ai-sdk/anthropic Zod schema doesn't accept "xhigh" yet, we send
- * "max" through the SDK for xhigh-on-4.7 and rewrite `output_config.effort` to
- * "xhigh" in the Anthropic fetch wrapper.
+ * "max" through the SDK for native-xhigh Anthropic models and rewrite
+ * `output_config.effort` to "xhigh" in the Anthropic fetch wrapper.
  */
 const ANTHROPIC_EFFORT: Record<ThinkingLevel, AnthropicEffortLevel> = {
   off: "low",
   low: "low",
   medium: "medium",
   high: "high",
-  xhigh: "max", // SDK placeholder; fetch wrapper rewrites to "xhigh" on Opus 4.7+
+  xhigh: "max", // SDK placeholder; fetch wrapper rewrites to "xhigh" on native-xhigh models
   max: "max",
 };
 
@@ -204,6 +205,8 @@ export function getAnthropicEffort(level: ThinkingLevel): AnthropicEffortLevel {
  * Matches:
  * - `claude-opus-4-7`, `claude-opus-4-8`, ... `claude-opus-4-99`, and any future Opus 5+
  *   (which we assume preserves or exceeds 4.7's capabilities).
+ * - `claude-sonnet-5` and any future Sonnet 5+ (Sonnet 5 introduced native xhigh effort for
+ *   the Sonnet tier; Sonnet 4.6 and earlier did not).
  * - Mythos-class models (`claude-fable-*`, `claude-mythos-*`), the tier above Opus.
  */
 export function anthropicSupportsNativeXhigh(modelString: string): boolean {
@@ -212,10 +215,11 @@ export function anthropicSupportsNativeXhigh(modelString: string): boolean {
     .toLowerCase()
     .replace(/^[a-z0-9_-]+:\s*/, "")
     .replace(/^[a-z0-9_-]+\//, "");
-  // Opus 4.7+ (4-7, 4-8, 4-9, 4-10, 4-11, ...) or any Opus 5+, plus the Mythos-class
-  // Fable / Mythos models that sit above Opus.
+  // Opus 4.7+ (4-7, 4-8, 4-9, 4-10, 4-11, ...) or any Opus 5+, Sonnet 5+ (5, 6, ... 10+),
+  // plus the Mythos-class Fable / Mythos models that sit above Opus.
   return (
     /claude-opus-(?:4-(?:[7-9]|\d{2,})|[5-9]|\d{2,})/.test(withoutPrefix) ||
+    /claude-sonnet-(?:[5-9]|\d{2,})/.test(withoutPrefix) ||
     /claude-(?:fable|mythos)-/.test(withoutPrefix)
   );
 }

src/common/utils/ai/models.test.ts

@@ -136,6 +136,14 @@ describe("Anthropic 1M context classification", () => {
     expect(hasNative1MContext("anthropic:claude-sonnet-4-6")).toBe(true);
   });
 
+  it("treats Sonnet 5 as a native 1M model", () => {
+    expect(getAnthropic1MContextMode("anthropic:claude-sonnet-5")).toBe("native");
+    expect(getAnthropic1MContextMode("anthropic:claude-sonnet-5-20260630")).toBe("native");
+    expect(getAnthropic1MContextMode("mux-gateway:anthropic/claude-sonnet-5")).toBe("native");
+    expect(supports1MContext("anthropic:claude-sonnet-5")).toBe(false);
+    expect(hasNative1MContext("anthropic:claude-sonnet-5")).toBe(true);
+  });
+
   it("treats Mythos-class Fable 5 / Mythos 5 as native 1M models", () => {
     expect(getAnthropic1MContextMode("anthropic:claude-fable-5")).toBe("native");
     expect(getAnthropic1MContextMode("anthropic:claude-mythos-5")).toBe("native");

src/common/utils/ai/models.ts

@@ -137,6 +137,7 @@ const ANTHROPIC_NATIVE_1M_PATTERNS = [
   new RegExp(`^claude-opus-4-8${OPTIONAL_VERSION_SUFFIX}$`, "i"),
   new RegExp(`^claude-opus-4-7${OPTIONAL_VERSION_SUFFIX}$`, "i"),
   new RegExp(`^claude-opus-4-6${OPTIONAL_VERSION_SUFFIX}$`, "i"),
+  new RegExp(`^claude-sonnet-5${OPTIONAL_VERSION_SUFFIX}$`, "i"),
   new RegExp(`^claude-sonnet-4-6${OPTIONAL_VERSION_SUFFIX}$`, "i"),
 ];
 const ANTHROPIC_BETA_1M_PATTERNS = [

src/common/utils/ai/providerOptions.test.ts

@@ -108,7 +108,7 @@ describe("buildProviderOptions - Anthropic", () => {
     });
   });
 
-  for (const model of ["claude-opus-4-6", "claude-sonnet-4-6"] as const) {
+  for (const model of ["claude-opus-4-6", "claude-sonnet-4-6", "claude-sonnet-5"] as const) {
     describe(`${model} (adaptive thinking + effort)`, () => {
       for (const { thinking, expectedThinking, effort } of [
         { thinking: "medium", expectedThinking: { type: "adaptive" }, effort: "medium" },
@@ -985,6 +985,14 @@ describe("buildRequestHeaders", () => {
         thinkingLevel: "xhigh",
         expected: { [MUX_ANTHROPIC_EFFORT_OVERRIDE_HEADER]: "xhigh" },
       },
+      {
+        // Sonnet 5 added native xhigh for the Sonnet tier, so it needs the wire rewrite too.
+        name: "emits override header for Sonnet 5",
+        model: "anthropic:claude-sonnet-5",
+        routeProvider: undefined,
+        thinkingLevel: "xhigh",
+        expected: { [MUX_ANTHROPIC_EFFORT_OVERRIDE_HEADER]: "xhigh" },
+      },
       {
         name: "does not emit override header for Opus 4.7 with thinkingLevel=max",
         model: "anthropic:claude-opus-4-7",

src/common/utils/ai/providerOptions.ts

@@ -273,23 +273,24 @@ export function buildProviderOptions(
     // serializes it to a top-level cache_control block, which adds an extra
     // breakpoint on direct Anthropic requests.
 
-    // Opus 4.5+ and Sonnet 4.6 use the effort parameter for reasoning control.
-    // Opus 4.6+ / Sonnet 4.6 use adaptive thinking (model decides when/how much to think).
+    // Opus 4.5+, Sonnet 4.6, and Sonnet 5 use the effort parameter for reasoning control.
+    // Opus 4.6+ / Sonnet 4.6 / Sonnet 5 use adaptive thinking (model decides when/how much).
     // Opus 4.5 uses enabled thinking with a budgetTokens ceiling.
     const isOpus45 = capModelName?.includes("opus-4-5") ?? false;
     const isOpus46 = capModelName?.includes("opus-4-6") ?? false;
-    const isOpus47Plus = anthropicSupportsNativeXhigh(capabilityModel);
+    // Opus 4.7+ and Sonnet 5+ — the native-xhigh tier, which also uses adaptive thinking.
+    const supportsNativeXhigh = anthropicSupportsNativeXhigh(capabilityModel);
     const isSonnet46 = capModelName?.includes("sonnet-4-6") ?? false;
-    const usesAdaptiveThinking = isOpus46 || isOpus47Plus || isSonnet46;
+    const usesAdaptiveThinking = isOpus46 || supportsNativeXhigh || isSonnet46;
 
     if (isOpus45 || usesAdaptiveThinking) {
-      // Map to SDK-accepted effort. For Opus 4.7+ with xhigh ThinkingLevel, the SDK
-      // gets "max" as a placeholder and the Anthropic fetch wrapper rewrites
+      // Map to SDK-accepted effort. For Opus 4.7+ / Sonnet 5+ with xhigh ThinkingLevel, the
+      // SDK gets "max" as a placeholder and the Anthropic fetch wrapper rewrites
       // `output_config.effort` to "xhigh" via the X-Mux-Anthropic-Effort header
       // (added in buildRequestHeaders).
       const effortLevel = getAnthropicEffort(effectiveThinking);
       const budgetTokens = ANTHROPIC_THINKING_BUDGETS[effectiveThinking];
-      // Opus 4.6+ / Sonnet 4.6: adaptive thinking when on, disabled when off
+      // Opus 4.6+ / Sonnet 4.6 / Sonnet 5: adaptive thinking when on, disabled when off
       // Opus 4.5: enabled thinking with budgetTokens ceiling (only when not "off")
       const thinking: AnthropicProviderOptions["thinking"] = usesAdaptiveThinking
         ? effectiveThinking === "off"
@@ -595,9 +596,9 @@ export function buildRequestHeaders(
     headers["anthropic-beta"] = ANTHROPIC_1M_CONTEXT_HEADER;
   }
 
-  // Opus 4.7 introduced a native "xhigh" effort level that the @ai-sdk/anthropic
-  // Zod schema doesn't accept yet. Emit a Mux-internal header so the Anthropic
-  // fetch wrapper can rewrite `output_config.effort` to "xhigh" on the wire.
+  // Native-xhigh Anthropic models use an "xhigh" effort level that the
+  // @ai-sdk/anthropic Zod schema doesn't accept yet. Emit a Mux-internal header
+  // so the Anthropic fetch wrapper can rewrite `output_config.effort` to "xhigh" on the wire.
   //
   // Only emit when the route will pass through our Anthropic fetch wrapper
   // (direct Anthropic or passthrough gateways like mux-gateway). Non-passthrough

src/common/utils/thinking/policy.test.ts

@@ -385,6 +385,36 @@ describe("getThinkingPolicyForModel", () => {
     ]);
   });
 
+  test("returns all 6 levels for Sonnet 5 (native xhigh)", () => {
+    // Sonnet 5 introduced the native xhigh effort level for the Sonnet tier, so it exposes
+    // all 6 levels (unlike Sonnet 4.6, which maps xhigh -> "max" and stops at 5).
+    expect(getThinkingPolicyForModel("anthropic:claude-sonnet-5")).toEqual([
+      "off",
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max",
+    ]);
+    expect(getThinkingPolicyForModel("anthropic:claude-sonnet-5-20260630")).toEqual([
+      "off",
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max",
+    ]);
+    // Behind gateway
+    expect(getThinkingPolicyForModel("mux-gateway:anthropic/claude-sonnet-5")).toEqual([
+      "off",
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max",
+    ]);
+  });
+
   test("returns 5 levels including xhigh for Sonnet 4.6", () => {
     expect(getThinkingPolicyForModel("anthropic:claude-sonnet-4-6")).toEqual([
       "off",

src/common/utils/tokens/models-extra.ts

@@ -155,6 +155,27 @@ export const modelsExtra: Record<string, ModelData> = {
     supports_response_schema: true,
   },
 
+  // Claude Sonnet 5 - Released June 30, 2026
+  // Native 1M context. Standard pricing $3/M input, $15/M output (same as Sonnet 4.6).
+  // Introductory pricing of $2/$10 per MTok applies through Aug 31, 2026, but we list the
+  // standard rate so cost estimates stay correct once the promo ends.
+  // 128K max output (up from Sonnet 4.6's 64K); supports adaptive thinking + effort (incl. xhigh).
+  "claude-sonnet-5": {
+    max_input_tokens: 1000000,
+    max_output_tokens: 128000,
+    input_cost_per_token: 0.000003, // $3 per million input tokens
+    output_cost_per_token: 0.000015, // $15 per million output tokens
+    cache_creation_input_token_cost: 0.00000375, // $3.75 per million tokens
+    cache_read_input_token_cost: 0.0000003, // $0.30 per million tokens
+    litellm_provider: "anthropic",
+    mode: "chat",
+    supports_function_calling: true,
+    supports_vision: true,
+    supports_pdf_input: true,
+    supports_reasoning: true,
+    supports_response_schema: true,
+  },
+
   // Claude Sonnet 4.6 - Released February 2026
   // Native 1M context at standard pricing: $3/M input, $15/M output.
   // 64K max output tokens, supports adaptive thinking + effort parameter.

src/common/utils/tools/tools.test.ts

@@ -6,7 +6,11 @@ import { Ok } from "@/common/types/result";
 import type { InitStateManager } from "@/node/services/initStateManager";
 import type { DesktopSessionManager } from "@/node/services/desktop/DesktopSessionManager";
 import { LocalRuntime } from "@/node/runtime/LocalRuntime";
-import { getToolsForModel, type WorkspaceHeartbeatToolService } from "./tools";
+import {
+  getToolsForModel,
+  supportsAnthropicNativeWebFetch,
+  type WorkspaceHeartbeatToolService,
+} from "./tools";
 
 const DESKTOP_TOOL_NAMES = [
   "desktop_screenshot",
@@ -59,6 +63,32 @@ function createDesktopSessionManager(options: { available: boolean }) {
   };
 }
 
+describe("supportsAnthropicNativeWebFetch", () => {
+  test.each([
+    // Major-only "5 generation" IDs (dateless naming) — Sonnet 5, Fable 5, Mythos 5.
+    ["claude-sonnet-5", true],
+    ["claude-fable-5", true],
+    ["claude-mythos-5", true],
+    // Two-segment IDs at/after the 4.6 cutoff.
+    ["claude-sonnet-4-6", true],
+    ["claude-opus-4-6", true],
+    ["claude-opus-4-8", true],
+    ["claude-opus-4-6-20260201", true],
+    // Below the 4.6 cutoff.
+    ["claude-sonnet-4-5", false],
+    ["claude-haiku-4-5", false],
+    // Date-based pre-4.6 ID must parse as major=4 / no minor and stay unsupported,
+    // not misread the 8-digit date as a minor version.
+    ["claude-sonnet-4-20250514", false],
+    // Older Claude 3.x IDs encode the family before the variant; do not misread
+    // `3-5` or `3-7` as variant=3, major=5/7.
+    ["claude-3-5-sonnet-20241022", false],
+    ["claude-3-7-sonnet-20250219", false],
+  ] as const)("%s -> %s", (modelId, expected) => {
+    expect(supportsAnthropicNativeWebFetch(modelId)).toBe(expected);
+  });
+});
+
 describe("getToolsForModel", () => {
   test("only includes agent_report when enableAgentReport=true", async () => {
     const runtime = new LocalRuntime(process.cwd());

src/common/utils/tools/tools.ts

@@ -489,19 +489,25 @@ async function getDesktopTools(config: ToolConfiguration): Promise<Record<string
 /**
  * Returns true when an Anthropic model supports webFetch_20250910 (Claude 4.6+).
  *
- * Generation-based IDs: claude-{variant}-{major}-{minor} (e.g. claude-sonnet-4-6)
- * Pinned generation IDs: claude-{variant}-{major}-{minor}-{date} (e.g. claude-opus-4-6-20260201)
- * Date-based pre-4.6 IDs: claude-{variant}-{major}-{date} (e.g. claude-sonnet-4-20250514)
+ * Two-segment IDs:    claude-{variant}-{major}-{minor} (e.g. claude-sonnet-4-6, claude-opus-4-8)
+ * Pinned two-segment: claude-{variant}-{major}-{minor}-{date} (e.g. claude-opus-4-6-20260201)
+ * Date-based pre-4.6: claude-{variant}-{major}-{date} (e.g. claude-sonnet-4-20250514)
+ * Major-only IDs:     claude-{variant}-{major} (e.g. claude-sonnet-5, claude-fable-5,
+ *                     claude-mythos-5) — the dateless naming adopted for the 5 generation.
  *
- * The \d{1,2} constraint on the minor segment accepts 1-2 digit version numbers (1–99) while
- * rejecting 8-digit date suffixes. The (?:-|$) lookahead allows an optional pinned date to follow.
+ * The minor segment is optional so major-only IDs (Sonnet 5, Fable 5, Mythos 5, future Opus 5+)
+ * are recognized; those are all > 4 and qualify. The variant segment must be alphabetic so older
+ * third-generation IDs like claude-3-5-sonnet-20241022 do not get misread as major=5. The \d{1,2}
+ * constraint accepts 1-2 digit version numbers (1–99) while rejecting 8-digit date suffixes, so
+ * date-based pre-4.6 IDs like claude-sonnet-4-20250514 parse as major=4 / no minor and correctly
+ * stay unsupported. The (?:-|$) lookahead allows an optional pinned date to follow.
  */
-function supportsAnthropicNativeWebFetch(modelId: string): boolean {
-  const match = /^claude-\w+-(\d+)-(\d{1,2})(?:-|$)/.exec(modelId);
+export function supportsAnthropicNativeWebFetch(modelId: string): boolean {
+  const match = /^claude-[a-z]+-(\d+)(?:-(\d{1,2}))?(?:-|$)/.exec(modelId);
   if (!match) return false;
   const major = parseInt(match[1], 10);
-  const minor = parseInt(match[2], 10);
-  return major > 4 || (major === 4 && minor >= 6);
+  const minor = match[2] != null ? parseInt(match[2], 10) : undefined;
+  return major > 4 || (major === 4 && minor !== undefined && minor >= 6);
 }
 
 export async function getToolsForModel(

src/node/services/agentSkills/builtInSkillContent.generated.ts

@@ -3101,7 +3101,7 @@ export const BUILTIN_SKILL_FILES: Record<string, Record<string, string>> = {
       "| ---------------------- | ----------------------------- | ------------------------------------------------------------ | ------- |",
       "| Fable 5                | anthropic:claude-fable-5      | `fable`                                                      |         |",
       "| Opus 4.8               | anthropic:claude-opus-4-8     | `opus`                                                       | ✓       |",
-      "| Sonnet 4.6             | anthropic:claude-sonnet-4-6   | `sonnet`                                                     |         |",
+      "| Sonnet 5               | anthropic:claude-sonnet-5     | `sonnet`                                                     |         |",
       "| Haiku 4.5              | anthropic:claude-haiku-4-5    | `haiku`                                                      |         |",
       "| GPT-5.5                | openai:gpt-5.5                | `gpt`, `gpt-5.5`                                             |         |",
       "| GPT-5.5 Pro            | openai:gpt-5.5-pro            | `gpt-pro`, `gpt-5.5-pro`                                     |         |",