@gajae-code/ai 0.18.12 → 0.19.0

package/src/model-thinking.d.ts

@@ -87,12 +87,23 @@ export declare function mapEffortToAnthropicAdaptiveEffort<TApi extends Api>(mod
  * - Thinking content is omitted by default (needs display: "summarized")
  */
 export declare function hasOpus47ApiRestrictions(modelId: string): boolean;
+/**
+ * Haiku 5+ enables adaptive thinking when the `thinking` field is omitted.
+ * An explicit user-facing Off choice must therefore send `type: "disabled"`.
+ */
+export declare function requiresAnthropicThinkingDisableWhenOff(modelId: string): boolean;
+/**
+ * Haiku 5+ and Opus 4.7+ reject sampling parameters on the Anthropic Messages API.
+ * https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide#remove-sampling-parameters
+ */
+export declare function hasAnthropicModelSamplingRestrictions(modelId: string): boolean;
 /**
  * Adaptive thinking `display` is supported starting with Anthropic Opus 4.7.
  * Older adaptive-thinking models (Opus 4.6, Sonnet 4.6+) reject the field.
  * Fable (5+) postdates Opus 4.7, accepts `display`, and defaults it to
  * "omitted" — thinking tokens are billed but no content streams back — so it
  * must opt in like Opus 4.7+ (issue #2791).
+ * Haiku 5+ also accepts `display` and defaults it to "omitted".
  *
  * Shares `hasOpus47ApiRestrictions` version parsing on purpose: the two
  * predicates describe the same API generation, and a private `claude-opus-(\d+)-(\d+)`

package/src/model-thinking.ts

@@ -5,7 +5,7 @@ import {
 	isCodexGpt56Tier,
 	isCodexProductTransport,
 } from "./context-cap-policy";
-import { applyOpenAIModelPricing } from "./model-pricing";
+import { applyAnthropicModelPricing, applyOpenAIModelPricing } from "./model-pricing";
 import {
 	isAuditedOpenAIReasoningTransport,
 	parseDirectXaiReasoningEffortGeneration,
@@ -56,6 +56,17 @@ const DEFAULT_REASONING_EFFORTS_WITH_XHIGH_AND_MAX: readonly Effort[] = [
 	Effort.XHigh,
 	Effort.Max,
 ];
+// Haiku 5.5 supports low, medium, high, xhigh, and max on both Anthropic
+// Messages and Bedrock Converse; neither contract lists minimal.
+// https://platform.claude.com/docs/en/build-with-claude/effort#effort-levels
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-5-5.html
+const ANTHROPIC_HAIKU_5_5_EFFORTS: readonly Effort[] = [
+	Effort.Low,
+	Effort.Medium,
+	Effort.High,
+	Effort.XHigh,
+	Effort.Max,
+];
 const GEMINI_3_PRO_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High];
 const GEMINI_3_FLASH_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High];
 // Gemini 3.7 Flash dropped `minimal`; the official API returns an error for it.
@@ -96,7 +107,7 @@ type SemVer = {
 };
 
 type GeminiKind = "pro" | "flash" | "argon";
-type AnthropicKind = "opus" | "sonnet" | "fable";
+type AnthropicKind = "opus" | "sonnet" | "fable" | "haiku";
 type OpenAIVariant =
 	| "astra"
 	| "base"
@@ -538,12 +549,34 @@ export function hasOpus47ApiRestrictions(modelId: string): boolean {
 	return semverGte(parsed.version, "4.7") && parsed.kind === "opus";
 }
 
+function isAnthropicHaikuAdaptiveModel(modelId: string): boolean {
+	const parsed = parseAnthropicModel(getCanonicalModelId(modelId));
+	return parsed?.kind === "haiku" && semverGte(parsed.version, "5.0");
+}
+
+/**
+ * Haiku 5+ enables adaptive thinking when the `thinking` field is omitted.
+ * An explicit user-facing Off choice must therefore send `type: "disabled"`.
+ */
+export function requiresAnthropicThinkingDisableWhenOff(modelId: string): boolean {
+	return isAnthropicHaikuAdaptiveModel(modelId);
+}
+
+/**
+ * Haiku 5+ and Opus 4.7+ reject sampling parameters on the Anthropic Messages API.
+ * https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide#remove-sampling-parameters
+ */
+export function hasAnthropicModelSamplingRestrictions(modelId: string): boolean {
+	return hasOpus47ApiRestrictions(modelId) || isAnthropicHaikuAdaptiveModel(modelId);
+}
+
 /**
  * Adaptive thinking `display` is supported starting with Anthropic Opus 4.7.
  * Older adaptive-thinking models (Opus 4.6, Sonnet 4.6+) reject the field.
  * Fable (5+) postdates Opus 4.7, accepts `display`, and defaults it to
  * "omitted" — thinking tokens are billed but no content streams back — so it
  * must opt in like Opus 4.7+ (issue #2791).
+ * Haiku 5+ also accepts `display` and defaults it to "omitted".
  *
  * Shares `hasOpus47ApiRestrictions` version parsing on purpose: the two
  * predicates describe the same API generation, and a private `claude-opus-(\d+)-(\d+)`
@@ -554,7 +587,7 @@ export function hasOpus47ApiRestrictions(modelId: string): boolean {
  */
 export function supportsAnthropicAdaptiveThinkingDisplay(modelId: string): boolean {
 	if (/claude-fable-\d/.test(modelId)) return true;
-	return hasOpus47ApiRestrictions(modelId);
+	return hasOpus47ApiRestrictions(modelId) || isAnthropicHaikuAdaptiveModel(modelId);
 }
 
 function anthropicModelHasRealXHighEffort<TApi extends Api>(model: ApiModel<TApi>): boolean {
@@ -576,6 +609,7 @@ function anthropicModelHasRealXHighEffort<TApi extends Api>(model: ApiModel<TApi
 
 function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	applyOpenAIModelPricing(model);
+	applyAnthropicModelPricing(model);
 	const copilotLimits = model.provider === "github-copilot" ? COPILOT_GENERATED_LIMITS[model.id] : undefined;
 	if (copilotLimits) {
 		model.contextWindow = copilotLimits.contextWindow;
@@ -879,6 +913,19 @@ function inferDefaultEffort<TApi extends Api>(model: ApiModel<TApi>, parsedModel
 	if (model.provider === "omlx") {
 		return Effort.Medium;
 	}
+	if (
+		parsedModel.family === "anthropic" &&
+		parsedModel.kind === "haiku" &&
+		((model.provider === "anthropic" &&
+			model.api === "anthropic-messages" &&
+			semverGte(parsedModel.version, "5.0")) ||
+			(model.provider === "amazon-bedrock" &&
+				model.api === "bedrock-converse-stream" &&
+				semverGte(parsedModel.version, "5.5")))
+	) {
+		// Haiku 5.5 defaults to medium adaptive effort instead of the global high.
+		return Effort.Medium;
+	}
 	return undefined;
 }
 
@@ -1027,6 +1074,9 @@ function inferAnthropicSupportedEfforts<TApi extends Api>(
 		(model.api === "anthropic-messages" || model.api === "bedrock-converse-stream") &&
 		semverGte(parsedModel.version, "4.6")
 	) {
+		if (parsedModel.kind === "haiku" && semverGte(parsedModel.version, "5.5")) {
+			return ANTHROPIC_HAIKU_5_5_EFFORTS;
+		}
 		if (parsedModel.kind === "fable") {
 			// Fable exposes Anthropic's Messages-only xhigh preset; Bedrock
 			// Converse lacks it (same split as Opus 4.7+ below).
@@ -1105,7 +1155,7 @@ function inferThinkingControlMode<TApi extends Api>(
 			if (parsedModel.family === "anthropic") {
 				if (
 					semverGte(parsedModel.version, "4.6") &&
-					(parsedModel.kind === "opus" || parsedModel.kind === "fable")
+					(parsedModel.kind === "opus" || parsedModel.kind === "fable" || parsedModel.kind === "haiku")
 				) {
 					return "anthropic-adaptive";
 				}
@@ -1146,7 +1196,7 @@ function parseGeminiModel(modelId: string): GeminiModel | null {
 }
 
 function parseAnthropicModel(modelId: string): AnthropicModel | null {
-	const match = /claude-(opus|sonnet|fable)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
+	const match = /claude-(opus|sonnet|fable|haiku)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
 	if (!match) {
 		return null;
 	}
@@ -1154,6 +1204,11 @@ function parseAnthropicModel(modelId: string): AnthropicModel | null {
 	if (!version) {
 		return null;
 	}
+	// Haiku is recognized from generation 5 only. Haiku 4.5 predates the
+	// adaptive-thinking policy and keeps the generic (unparsed) budget ladder.
+	if (match[1] === "haiku" && !semverGte(version, "5.0")) {
+		return null;
+	}
 	return { family: "anthropic", kind: match[1] as AnthropicKind, version };
 }
 

package/src/models.json

@@ -779,6 +779,41 @@
         "mode": "anthropic-adaptive"
       }
     },
+    "anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.01,
+        "cacheWrite": 0.125,
+        "input": 0.1,
+        "output": 0.5
+      },
+      "id": "anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.05,
+          "cacheWrite": 0.625,
+          "input": 0.5,
+          "output": 2.5
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "anthropic.claude-opus-4-6-v1": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -1098,6 +1133,41 @@
         "mode": "budget"
       }
     },
+    "au.anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.011,
+        "cacheWrite": 0.1375,
+        "input": 0.11,
+        "output": 0.55
+      },
+      "id": "au.anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.055,
+          "cacheWrite": 0.6875,
+          "input": 0.55,
+          "output": 2.75
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5 (AU)",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "au.anthropic.claude-opus-4-6-v1": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -1787,6 +1857,41 @@
         "mode": "budget"
       }
     },
+    "eu.anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.011,
+        "cacheWrite": 0.1375,
+        "input": 0.11,
+        "output": 0.55
+      },
+      "id": "eu.anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.055,
+          "cacheWrite": 0.6875,
+          "input": 0.55,
+          "output": 2.75
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5 (EU)",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "eu.anthropic.claude-opus-4-1-20250805-v1:0": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -2267,6 +2372,41 @@
         "mode": "budget"
       }
     },
+    "global.anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.01,
+        "cacheWrite": 0.125,
+        "input": 0.1,
+        "output": 0.5
+      },
+      "id": "global.anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.05,
+          "cacheWrite": 0.625,
+          "input": 0.5,
+          "output": 2.5
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5 (Global)",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "global.anthropic.claude-opus-4-5-20251101-v1:0": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -3124,6 +3264,41 @@
         "mode": "budget"
       }
     },
+    "jp.anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.011,
+        "cacheWrite": 0.1375,
+        "input": 0.11,
+        "output": 0.55
+      },
+      "id": "jp.anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.055,
+          "cacheWrite": 0.6875,
+          "input": 0.55,
+          "output": 2.75
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5 (JP)",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "jp.anthropic.claude-opus-4-7": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -4573,6 +4748,41 @@
         "mode": "budget"
       }
     },
+    "us.anthropic.claude-haiku-5-5": {
+      "api": "bedrock-converse-stream",
+      "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.011,
+        "cacheWrite": 0.1375,
+        "input": 0.11,
+        "output": 0.55
+      },
+      "id": "us.anthropic.claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.055,
+          "cacheWrite": 0.6875,
+          "input": 0.55,
+          "output": 2.75
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5 (US)",
+      "provider": "amazon-bedrock",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "us.anthropic.claude-opus-4-1-20250805-v1:0": {
       "api": "bedrock-converse-stream",
       "baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com",
@@ -5782,6 +5992,41 @@
         "mode": "budget"
       }
     },
+    "claude-haiku-5-5": {
+      "api": "anthropic-messages",
+      "baseUrl": "https://api.anthropic.com",
+      "contextWindow": 1000000,
+      "cost": {
+        "cacheRead": 0.01,
+        "cacheWrite": 0.125,
+        "input": 0.1,
+        "output": 0.5
+      },
+      "id": "claude-haiku-5-5",
+      "input": [
+        "text",
+        "image"
+      ],
+      "longContextPricing": {
+        "cost": {
+          "cacheRead": 0.05,
+          "cacheWrite": 0.625,
+          "input": 0.5,
+          "output": 2.5
+        },
+        "threshold": 100000
+      },
+      "maxTokens": 128000,
+      "name": "Anthropic Haiku 5.5",
+      "provider": "anthropic",
+      "reasoning": true,
+      "thinking": {
+        "defaultLevel": "medium",
+        "maxLevel": "max",
+        "minLevel": "low",
+        "mode": "anthropic-adaptive"
+      }
+    },
     "claude-opus-4-0": {
       "api": "anthropic-messages",
       "baseUrl": "https://api.anthropic.com",

package/src/models.ts

@@ -1,6 +1,6 @@
 import { readFileSync } from "node:fs";
 import { registerProviderSafetyStopModel } from "./adapter-internals/provider-safety-stop";
-import { getOpenAIModelCost } from "./model-pricing";
+import { getAnthropicCacheWriteOneHourRate, getAnthropicModelCost, getOpenAIModelCost } from "./model-pricing";
 import { isRetiredModelKey } from "./model-retirements";
 import { applyGeneratedModelPolicies, enrichModelThinking } from "./model-thinking";
 // `with { type: "file" }` is embedded by `bun build --compile` and resolves to
@@ -105,11 +105,18 @@ export function getBundledModels(provider: GeneratedProvider): Model<Api>[] {
 
 export function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage["cost"] {
 	const inputTokens = usage.input + usage.cacheRead + usage.cacheWrite;
-	const pricing = getOpenAIModelCost(model, inputTokens) ?? model.cost;
+	const pricing = getAnthropicModelCost(model, inputTokens) ?? getOpenAIModelCost(model, inputTokens) ?? model.cost;
 	usage.cost.input = (pricing.input / 1000000) * usage.input;
 	usage.cost.output = (pricing.output / 1000000) * usage.output;
 	usage.cost.cacheRead = (pricing.cacheRead / 1000000) * usage.cacheRead;
-	usage.cost.cacheWrite = (pricing.cacheWrite / 1000000) * usage.cacheWrite;
+	const oneHourRate = getAnthropicCacheWriteOneHourRate(model, inputTokens);
+	// Bill the response's 1-hour cache-write tokens at the 1-hour rate; everything else
+	// (5-minute writes and any unattributed remainder) uses the model's cacheWrite rate.
+	const oneHourTokens =
+		oneHourRate === undefined ? 0 : Math.min(Math.max(usage.cttl?.ephemeral1h ?? 0, 0), usage.cacheWrite);
+	usage.cost.cacheWrite =
+		(pricing.cacheWrite / 1000000) * (usage.cacheWrite - oneHourTokens) +
+		((oneHourRate ?? 0) / 1000000) * oneHourTokens;
 	usage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;
 	return usage.cost;
 }

package/src/provider-models/google.ts

@@ -14,11 +14,13 @@ export interface GoogleVertexModelManagerConfig {
 export interface GoogleAntigravityModelManagerConfig {
 	oauthToken?: string;
 	endpoint?: string;
+	project?: string;
 }
 
 export interface GoogleGeminiCliModelManagerConfig {
 	oauthToken?: string;
 	endpoint?: string;
+	project?: string;
 }
 
 const CLOUD_CODE_ASSIST_ENDPOINT = "https://cloudcode-pa.googleapis.com";
@@ -58,6 +60,7 @@ export function googleAntigravityModelManagerOptions(
 						fetchAntigravityDiscoveryModels({
 							token,
 							endpoint: config?.endpoint,
+							project: config?.project,
 							userAgent: ua,
 						}),
 				}
@@ -80,6 +83,7 @@ export function googleGeminiCliModelManagerOptions(
 						const models = await fetchAntigravityDiscoveryModels({
 							token,
 							endpoint,
+							project: config?.project,
 							targetProvider: "google-gemini-cli",
 							userAgent: ua,
 						});

package/src/providers/amazon-bedrock.d.ts

@@ -8,7 +8,7 @@
  */
 import { parseBedrockClaudeGeneration } from "../bedrock-claude-cache-policy";
 import type { Effort } from "../model-thinking";
-import type { Model, StreamFunction, StreamOptions, ThinkingBudgets, Tool, ToolChoice } from "../types";
+import type { AssistantMessage, CacheRetention, Model, StreamFunction, StreamOptions, ThinkingBudgets, Tool, ToolChoice } from "../types";
 export type BedrockThinkingDisplay = "summarized" | "omitted";
 export interface BedrockOptions extends StreamOptions {
     region?: string;
@@ -51,7 +51,18 @@ interface WireToolConfig {
     tools: WireToolSpec[];
     toolChoice?: WireToolChoice;
 }
+interface MetadataEvent {
+    usage?: {
+        inputTokens?: number;
+        outputTokens?: number;
+        cacheReadInputTokens?: number;
+        cacheWriteInputTokens?: number;
+        totalTokens?: number;
+    };
+}
 export declare const streamBedrock: StreamFunction<"bedrock-converse-stream">;
+/** Exported for unit tests of the usage-to-cost path. */
+export declare function handleMetadata(event: MetadataEvent, model: Model<"bedrock-converse-stream">, output: AssistantMessage, cacheRetention: CacheRetention): void;
 /**
  * Check if the model supports prompt caching.
  * Supported: Claude 3.5 Haiku, Claude 3.7 Sonnet, and every later Claude
@@ -71,5 +82,9 @@ export declare function stripBedrockForcedToolChoiceForRetry<T extends {
     toolConfig?: {
         toolChoice?: unknown;
     };
-}>(body: T): T;
+    additionalModelRequestFields?: Record<string, unknown>;
+}>(body: T, thinking?: {
+    forced: Record<string, unknown> | undefined;
+    restored: Record<string, unknown>;
+}): T;
 export declare function convertToolConfig(tools: Tool[] | undefined, toolChoice: BedrockOptions["toolChoice"]): WireToolConfig | undefined;

package/src/providers/amazon-bedrock.ts

@@ -10,10 +10,12 @@
 import { $credentialEnv, $env, $flag, extractHttpStatusFromError, fetchWithRetry } from "@gajae-code/utils";
 import { assertAwsRegionLabel } from "../adapter-internals/aws-region";
 import { parseBedrockClaudeGeneration, supportsBedrockClaudePromptCaching } from "../bedrock-claude-cache-policy";
+import { getAnthropicCacheWriteOneHourRate } from "../model-pricing";
 import type { Effort } from "../model-thinking";
 import {
 	mapEffortToAnthropicAdaptiveEffort,
 	requireSupportedEffort,
+	requiresAnthropicThinkingDisableWhenOff,
 	supportsAnthropicAdaptiveThinkingDisplay as supportsAdaptiveThinkingDisplay,
 } from "../model-thinking";
 import { calculateCost } from "../models";
@@ -213,12 +215,30 @@ export const streamBedrock: StreamFunction<"bedrock-converse-stream"> = (
 			const resolvedToolChoice = resolveToolChoice(model, options.toolChoice);
 			const toolConfig = convertToolConfig(context.tools, resolvedToolChoice.resolvedChoice);
 			let additionalModelRequestFields = buildAdditionalModelRequestFields(model, options);
+			// Selected reasoning fields before forced-tool normalization, restored if the forced
+			// tool choice is later dropped on retry (Haiku 5+ only; see below).
+			let unforcedModelRequestFields: Record<string, unknown> | undefined;
 
 			// Bedrock rejects thinking + forced tool_choice ("any" or specific tool).
 			// When the resolved tool_choice forces tool use, disable thinking to avoid API errors.
 			if (toolConfig?.toolChoice && additionalModelRequestFields) {
 				const tc = toolConfig.toolChoice;
-				if (tc.any || tc.tool) additionalModelRequestFields = undefined;
+				if ((tc.any || tc.tool) && requiresAnthropicThinkingDisableWhenOff(model.id)) {
+					// Haiku 5+ enables adaptive thinking when the field is omitted, which a forced
+					// tool choice rejects, and applies the effort to tool-call arguments too: send an
+					// explicit `disabled` and keep the selected effort (capped at `high` while
+					// thinking is disabled).
+					const effort = (additionalModelRequestFields.output_config as { effort?: string } | undefined)?.effort;
+					unforcedModelRequestFields = additionalModelRequestFields;
+					additionalModelRequestFields = {
+						thinking: { type: "disabled" },
+						...(effort
+							? { output_config: { effort: effort === "xhigh" || effort === "max" ? "high" : effort } }
+							: {}),
+					};
+				} else if (tc.any || tc.tool) {
+					additionalModelRequestFields = undefined;
+				}
 			}
 
 			let commandInput: ConverseStreamRequest = {
@@ -307,7 +327,12 @@ export const streamBedrock: StreamFunction<"bedrock-converse-stream"> = (
 					reason,
 					registryKey: resolvedToolChoice.registryKey,
 				});
-				stripBedrockForcedToolChoiceForRetry(commandInput);
+				stripBedrockForcedToolChoiceForRetry(
+					commandInput,
+					unforcedModelRequestFields
+						? { forced: additionalModelRequestFields, restored: unforcedModelRequestFields }
+						: undefined,
+				);
 				const retryBodyText = JSON.stringify(commandInput);
 				const retryBody = new TextEncoder().encode(retryBodyText);
 				if (rawRequestDump) rawRequestDump.body = commandInput;
@@ -429,7 +454,7 @@ export const streamBedrock: StreamFunction<"bedrock-converse-stream"> = (
 							break;
 						}
 						case "metadata": {
-							handleMetadata(payload as MetadataEvent, model, output);
+							handleMetadata(payload as MetadataEvent, model, output, cacheRetention);
 							break;
 						}
 						default:
@@ -580,13 +605,28 @@ function handleContentBlockDelta(
 	}
 }
 
-function handleMetadata(event: MetadataEvent, model: Model<"bedrock-converse-stream">, output: AssistantMessage): void {
+/** Exported for unit tests of the usage-to-cost path. */
+export function handleMetadata(
+	event: MetadataEvent,
+	model: Model<"bedrock-converse-stream">,
+	output: AssistantMessage,
+	cacheRetention: CacheRetention,
+): void {
 	if (event.usage) {
 		output.usage.input = event.usage.inputTokens || 0;
 		output.usage.output = event.usage.outputTokens || 0;
 		output.usage.cacheRead = event.usage.cacheReadInputTokens || 0;
 		output.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;
 		output.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;
+		// Converse reports aggregate cache writes only. Where 1-hour writes have their own rate
+		// (Haiku 5.5), the request's `ttl: "1h"` cache points mean every write was 1-hour.
+		if (
+			cacheRetention === "long" &&
+			output.usage.cacheWrite > 0 &&
+			getAnthropicCacheWriteOneHourRate(model, 0) !== undefined
+		) {
+			output.usage.cttl = { ephemeral1h: output.usage.cacheWrite };
+		}
 		calculateCost(model, output.usage);
 	}
 }
@@ -829,10 +869,18 @@ function convertMessages(
 	return result;
 }
 
-export function stripBedrockForcedToolChoiceForRetry<T extends { toolConfig?: { toolChoice?: unknown } }>(body: T): T {
+export function stripBedrockForcedToolChoiceForRetry<
+	T extends { toolConfig?: { toolChoice?: unknown }; additionalModelRequestFields?: Record<string, unknown> },
+>(body: T, thinking?: { forced: Record<string, unknown> | undefined; restored: Record<string, unknown> }): T {
 	if (body.toolConfig) {
 		body.toolConfig = { ...body.toolConfig, toolChoice: undefined };
 	}
+	// The forced request replaced the selected thinking fields with a disabled block; with the
+	// forced constraint gone, put the selected reasoning back. A payload the caller replaced via
+	// onPayload (not the object this provider sent) is left untouched.
+	if (thinking && body.additionalModelRequestFields === thinking.forced) {
+		body.additionalModelRequestFields = thinking.restored;
+	}
 	return body;
 }
 
@@ -888,6 +936,9 @@ function buildAdditionalModelRequestFields(
 	options: BedrockOptions,
 ): Record<string, unknown> | undefined {
 	const reasoning = options.reasoning;
+	if (requiresAnthropicThinkingDisableWhenOff(model.id) && (!reasoning || !model.reasoning)) {
+		return { thinking: { type: "disabled" } };
+	}
 	if (!reasoning || !model.reasoning) return undefined;
 
 	const mode = model.thinking?.mode;