feat(anthropic+openrouter): add Opus 4.8 with adaptive-thinking effort

Adds claude-opus-4-8 to the static Anthropic catalog and wires it
through the provider switch + OpenRouter passthrough.

Opus 4.8 facts (from platform.claude.com):
  - 1M context, 128k max output, $5/$25 per MTok (same as 4.7)
  - Adaptive thinking only; manual budget_tokens returns a 400 error
  - Effort values: low / medium / high / xhigh / max
  - Default effort: high (Anthropic explicitly recommends as best
    balance -- yields token-spend similar to Opus 4.7's xhigh default
    but with better performance)

For OpenRouter, the dynamic fetcher reports supportsReasoningEffort
as a boolean which collapses the UI to the legacy [low/medium/high]
fallback (same gap we fixed for gpt-5.5). Mirror the static
capability array onto anthropic/claude-opus-4.7 AND
anthropic/claude-opus-4.8 so xhigh and max are reachable in both.
Disable supportsReasoningBudget since both reject budget_tokens.

Parametrized fetcher test covers both ids.

Author: Zeph Gillen

packages/types/src/providers/anthropic.ts

@@ -111,6 +111,30 @@ export const anthropicModels = {
 		supportsTemperature: false,
 		supportsReasoningDisplay: true,
 	},
+	"claude-opus-4-8": {
+		maxTokens: 128_000,
+		// Native 1M-token context window; flat pricing (200k on Microsoft Foundry only).
+		// Source: https://platform.claude.com/docs/en/about-claude/models/overview
+		contextWindow: 1_000_000,
+		supportsImages: true,
+		supportsPromptCache: true,
+		inputPrice: 5.0,
+		outputPrice: 25.0,
+		cacheWritesPrice: 6.25,
+		cacheReadsPrice: 0.5,
+		// Opus 4.8 uses adaptive thinking only; manual `thinking: {type: "enabled", budget_tokens}`
+		// returns a 400 error. Effort controls thinking depth.
+		// Per https://platform.claude.com/docs/en/build-with-claude/effort:
+		//   - Effort defaults to `high` on the API and Claude Code; Anthropic explicitly states
+		//     this is "the best overall balance of quality and user experience" and yields
+		//     similar token usage to Opus 4.7's xhigh default but with better performance.
+		//   - `xhigh` is recommended for difficult tasks and long-running async workflows.
+		supportsReasoningEffort: ["low", "medium", "high", "xhigh", "max"],
+		requiredReasoningEffort: true,
+		reasoningEffort: "high",
+		supportsTemperature: false,
+		supportsReasoningDisplay: true,
+	},
 	"claude-opus-4-5-20251101": {
 		maxTokens: 32_000, // Overridden to 8k if `enableReasoningEffort` is false.
 		contextWindow: 200_000,

src/api/providers/anthropic.ts

@@ -90,6 +90,7 @@ export class AnthropicHandler extends BaseProvider implements SingleCompletionHa
 			case "claude-sonnet-4-20250514":
 			case "claude-opus-4-6":
 			case "claude-opus-4-7":
+			case "claude-opus-4-8":
 			case "claude-opus-4-5-20251101":
 			case "claude-opus-4-1-20250805":
 			case "claude-opus-4-20250514":
@@ -159,6 +160,7 @@ export class AnthropicHandler extends BaseProvider implements SingleCompletionHa
 								case "claude-sonnet-4-20250514":
 								case "claude-opus-4-6":
 								case "claude-opus-4-7":
+								case "claude-opus-4-8":
 								case "claude-opus-4-5-20251101":
 								case "claude-opus-4-1-20250805":
 								case "claude-opus-4-20250514":

src/api/providers/fetchers/__tests__/openrouter.spec.ts

@@ -266,6 +266,34 @@ describe("OpenRouter API", () => {
 	})
 
 	describe("parseOpenRouterModel", () => {
+		it.each([["anthropic/claude-opus-4.7"], ["anthropic/claude-opus-4.8"]])(
+			"mirrors static capability for %s on OpenRouter",
+			(id) => {
+				const mockModel = {
+					name: id,
+					description: "Test model",
+					context_length: 1_000_000,
+					max_completion_tokens: 128000,
+					pricing: { prompt: "0.000005", completion: "0.000025", input_cache_read: "0.0000005" },
+				}
+
+				const result = parseOpenRouterModel({
+					id,
+					model: mockModel,
+					inputModality: ["text", "image"],
+					outputModality: ["text"],
+					maxTokens: 128000,
+					supportedParameters: ["reasoning", "tools"],
+				})
+
+				expect(result.maxTokens).toBe(128000)
+				expect(result.supportsReasoningEffort).toEqual(["low", "medium", "high", "xhigh", "max"])
+				expect(result.requiredReasoningEffort).toBe(true)
+				expect(result.supportsReasoningBudget).toBe(false)
+				expect(result.supportsTemperature).toBe(false)
+			},
+		)
+
 		it("sets claude-sonnet-4.6 model to Anthropic max tokens", () => {
 			const mockModel = {
 				name: "Claude Sonnet 4.6",

src/api/providers/fetchers/openrouter.ts

@@ -279,6 +279,22 @@ export const parseOpenRouterModel = ({
 		modelInfo.maxTokens = anthropicModels["claude-opus-4-6"].maxTokens
 	}
 
+	// Opus 4.7 and 4.8 via OpenRouter: mirror static capability arrays so the UI
+	// surfaces xhigh / max effort levels (boolean true would collapse to the
+	// legacy low/medium/high fallback). Both models use adaptive thinking only
+	// and reject budget_tokens, so disable supportsReasoningBudget.
+	if (id === "anthropic/claude-opus-4.7" || id === "anthropic/claude-opus-4.8") {
+		const staticId = id === "anthropic/claude-opus-4.7" ? "claude-opus-4-7" : "claude-opus-4-8"
+		const staticDef = anthropicModels[staticId]
+		modelInfo.maxTokens = staticDef.maxTokens
+		modelInfo.supportsReasoningEffort = staticDef.supportsReasoningEffort
+		modelInfo.reasoningEffort = staticDef.reasoningEffort
+		modelInfo.requiredReasoningEffort = staticDef.requiredReasoningEffort
+		modelInfo.supportsReasoningBudget = false
+		modelInfo.supportsReasoningDisplay = staticDef.supportsReasoningDisplay
+		modelInfo.supportsTemperature = staticDef.supportsTemperature
+	}
+
 	// Ensure correct reasoning handling for Claude Haiku 4.5 on OpenRouter
 	// Use budget control and disable effort-based reasoning fallback
 	if (id === "anthropic/claude-haiku-4.5") {