@oh-my-pi/pi-catalog 17.2.3 → 17.2.4

package/CHANGELOG.md

@@ -2,9 +2,18 @@
 
 ## [Unreleased]
 
+## [17.2.4] - 2026-08-01
+
+### Added
+
+- Added `AnthropicCompat.streamIdleTimeoutMs` and propagated it through `buildAnthropicCompat` so direct Anthropic provider streams can configure their inter-event idle watchdog.
+- Fixed Ollama Cloud DeepSeek V4 Pro/Flash models (including dated tag variants such as `deepseek-v4-flash:0731`) reporting an incorrect max-output-tokens figure by pinning it to the deployment's enforced 65536-token output ceiling ([#7266](https://github.com/can1357/oh-my-pi/issues/7266)).
+
 ### Fixed
 
 - Fixed `gen:models` Codex discovery to union models across every stored OAuth account and fail closed on partial resolution, matching runtime discovery ([#6265](https://github.com/can1357/oh-my-pi/issues/6265)); restored the bundled `gpt-5.4`, `gpt-5.6-sol`, and `gpt-5.3-codex-spark` entries a single-account regen had dropped.
+- Fixed `google-antigravity` models always reporting $0 cost: Antigravity discovery carries no pricing, so the generator now back-fills each model with its Google list price (Gemini ids from the `google` provider, including `-preview` id aliases; Claude ids from `google-vertex`, falling back to `anthropic`).
+- Fixed OpenRouter `deepseek/deepseek-v4-flash-0731` exposing only `high` thinking effort by consuming the live `reasoning.supported_efforts` and `default_effort` metadata and bundling its `low`/`high`/`max` ladder. ([#7307](https://github.com/can1357/oh-my-pi/issues/7307))
 
 ## [17.2.3] - 2026-08-01
 

package/dist/types/provider-models/ollama.d.ts

@@ -5,5 +5,18 @@ export interface OllamaCloudModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * Output-token ceiling that Ollama Cloud enforces for the DeepSeek V4 Pro/Flash
+ * deployments: `/api/chat` rejects `num_predict` above it with HTTP 400
+ * (`max_tokens (...) exceeds model's maximum output tokens (65536)`) even though
+ * the model pages advertise a 1M context / 384K output. Ollama's `/api/show`
+ * never reports this cap, so the catalog pins it for the affected models
+ * (ollama/ollama#16890, #7266). The wire layer clamps `num_predict` to the same
+ * value (`OLLAMA_CLOUD_NUM_PREDICT_CAP` in `packages/ai/src/providers/ollama.ts`,
+ * #3392/#3394).
+ */
+export declare const OLLAMA_CLOUD_MAX_OUTPUT_TOKENS = 65536;
+/** Whether an Ollama Cloud model id (tagged or not) enforces the 65536 output cap. */
+export declare function isOllamaCloudOutputCapped(id: string): boolean;
 export declare function normalizeOllamaCloudBaseUrl(baseUrl?: string): string;
 export declare function ollamaCloudModelManagerOptions(config?: OllamaCloudModelManagerConfig): ModelManagerOptions<"ollama-chat">;

package/dist/types/types.d.ts

@@ -362,6 +362,15 @@ export interface OpenAICompat {
  * that proxy gateways (Vertex AI, AWS Bedrock-style fronts, etc.) reject.
  */
 export interface AnthropicCompat {
+    /**
+     * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts.
+     * Set to 0 to disable the inter-event idle watchdog entirely, matching
+     * `OpenAICompat.streamIdleTimeoutMs`.
+     *
+     * When unset, direct Anthropic streams use `PI_STREAM_IDLE_TIMEOUT_MS`,
+     * then the legacy `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+     */
+    streamIdleTimeoutMs?: number;
     /**
      * Drop the top-level `strict: true` field on tool definitions. Vertex AI's
      * Anthropic-compatible endpoint rejects unknown tool fields with
@@ -605,7 +614,13 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
  */
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs">> & {
+    /**
+     * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts; 0 disables the idle watchdog.
+     * Undefined defers to `PI_STREAM_IDLE_TIMEOUT_MS`, then the legacy
+     * `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+     */
+    streamIdleTimeoutMs?: number;
     /**
      * The configured endpoint is the official first-party Anthropic API
      * (https + exact `api.anthropic.com` host; a missing baseUrl counts as

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "17.2.3",
+	"version": "17.2.4",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,12 +35,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.1",
-		"@oh-my-pi/pi-utils": "17.2.3",
+		"@oh-my-pi/pi-utils": "17.2.4",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.2.3",
+		"@oh-my-pi/pi-ai": "17.2.4",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -172,6 +172,7 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		// id or baseUrl marker.
 		replayUnsignedThinking: !signingEndpoint && (Boolean(spec.reasoning) || modelMatchesHost(spec, "deepseekFamily")),
 		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
+		streamIdleTimeoutMs: spec.compat?.streamIdleTimeoutMs,
 	};
 	applyCompatOverrides(compat, spec.compat);
 	return compat;

package/src/model-thinking.ts

@@ -62,8 +62,8 @@ const GEMINI_3_FLASH_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort.Low, E
 const GPT_5_2_PLUS_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
 const GPT_5_1_CODEX_MINI_EFFORTS: readonly Effort[] = [Effort.Medium, Effort.High];
 const LOW_MEDIUM_HIGH_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
-/** Kimi K3's wire-exact mandatory reasoning scale. */
-const KIMI_K3_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
+/** Wire-exact `low`/`high`/`max` scale used by Kimi K3 and OpenRouter DeepSeek V4 Flash 0731. */
+const LOW_HIGH_MAX_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
 /** Wire-exact two-tier scale (`high`/`max`): GLM-5.2 on Z.ai/Umans/Ollama Cloud/Baseten, Sakana Fugu, DeepSeek. */
 const HIGH_MAX_REASONING_EFFORTS: readonly Effort[] = [Effort.High, Effort.Max];
 /** OpenRouter's DeepSeek route accepts only `high`. */
@@ -339,7 +339,7 @@ function getModelDefinedEfforts<TApi extends Api>(
 		}
 	}
 	if (isKimiK3ModelId(spec.id)) {
-		return KIMI_K3_REASONING_EFFORTS;
+		return LOW_HIGH_MAX_REASONING_EFFORTS;
 	}
 	if (isSakanaFuguReasoningModel(spec)) {
 		return HIGH_MAX_REASONING_EFFORTS;
@@ -366,9 +366,14 @@ function getModelDefinedEfforts<TApi extends Api>(
 		return OLLAMA_REASONING_EFFORTS;
 	}
 	if (isOpenAICompatReasoningApi(spec.api) && isDeepseekReasoningModel(spec)) {
-		// DeepSeek's reasoning_effort accepts only high/max; OpenRouter's
-		// DeepSeek route tops out at high.
-		return isOpenRouterThinkingFormat(compat) ? HIGH_ONLY_REASONING_EFFORTS : HIGH_MAX_REASONING_EFFORTS;
+		// OpenRouter generally exposes only high for DeepSeek, but V4 Flash 0731
+		// advertises and accepts the wire-exact low/high/max ladder.
+		if (isOpenRouterThinkingFormat(compat)) {
+			return bareModelId(spec.id) === "deepseek-v4-flash-0731"
+				? LOW_HIGH_MAX_REASONING_EFFORTS
+				: HIGH_ONLY_REASONING_EFFORTS;
+		}
+		return HIGH_MAX_REASONING_EFFORTS;
 	}
 	if (spec.provider === "baseten" && isOpenAIGptOssModelId(spec.id)) {
 		// Baseten's gpt-oss router mirrors its GLM route: high/max only.

package/src/models.json

@@ -11768,8 +11768,7 @@
 				],
 				"supportsDisplay": true
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"claude-opus-4-0": {
 			"id": "claude-opus-4-0",
@@ -13314,9 +13313,9 @@
 				"text"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.13,
+				"output": 0.26,
+				"cacheRead": 0.028,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -20750,7 +20749,7 @@
 		},
 		"deepseek-v4-flash-0731": {
 			"id": "deepseek-v4-flash-0731",
-			"name": "DeepSeek-V4-Flash-0731",
+			"name": "DeepSeek V4 Flash 0731",
 			"api": "openai-completions",
 			"provider": "fireworks",
 			"baseUrl": "https://api.fireworks.ai/inference/v1",
@@ -20759,13 +20758,13 @@
 				"text"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.175,
+				"output": 0.35,
+				"cacheRead": 0.035,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 1048576,
+			"maxTokens": 32768,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
@@ -20773,8 +20772,7 @@
 					"max"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"deepseek-v4-pro": {
 			"id": "deepseek-v4-pro",
@@ -24554,10 +24552,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"input": 5,
+				"output": 25,
+				"cacheRead": 0.5,
+				"cacheWrite": 6.25
 			},
 			"contextWindow": 200000,
 			"maxTokens": 64000,
@@ -24591,10 +24589,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"input": 5,
+				"output": 25,
+				"cacheRead": 0.5,
+				"cacheWrite": 6.25
 			},
 			"contextWindow": 250000,
 			"maxTokens": 64000,
@@ -24621,10 +24619,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"input": 3,
+				"output": 15,
+				"cacheRead": 0.3,
+				"cacheWrite": 3.75
 			},
 			"contextWindow": 1000000,
 			"maxTokens": 64000,
@@ -24658,10 +24656,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"input": 3,
+				"output": 15,
+				"cacheRead": 0.3,
+				"cacheWrite": 3.75
 			},
 			"contextWindow": 250000,
 			"maxTokens": 64000,
@@ -24687,9 +24685,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.3,
+				"output": 2.5,
+				"cacheRead": 0.03,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24723,9 +24721,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.1,
+				"output": 0.4,
+				"cacheRead": 0.01,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24752,9 +24750,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 1.25,
+				"output": 10,
+				"cacheRead": 0.125,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24782,9 +24780,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.5,
+				"output": 3,
+				"cacheRead": 0.05,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24826,9 +24824,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 2,
+				"output": 12,
+				"cacheRead": 0.2,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24879,9 +24877,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.25,
+				"output": 1.5,
+				"cacheRead": 0.025,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24909,9 +24907,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 2,
+				"output": 12,
+				"cacheRead": 0.2,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24947,9 +24945,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 1.5,
+				"output": 9,
+				"cacheRead": 0.15,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -24991,9 +24989,9 @@
 				"image"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 1.5,
+				"output": 7.5,
+				"cacheRead": 0.15,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -27972,6 +27970,26 @@
 				]
 			}
 		},
+		"~deepseek/deepseek-v4-flash-latest": {
+			"id": "~deepseek/deepseek-v4-flash-latest",
+			"name": "DeepSeek V4 Flash Latest",
+			"api": "openai-completions",
+			"provider": "kilo",
+			"baseUrl": "https://api.kilo.ai/api/gateway",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1048576,
+			"maxTokens": 1048576,
+			"supportsComputerUse": false
+		},
 		"~google/gemini-flash-latest": {
 			"id": "~google/gemini-flash-latest",
 			"name": "Gemini Flash Latest",
@@ -40004,8 +40022,8 @@
 			"cost": {
 				"input": 0.3,
 				"output": 1.2,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"cacheRead": 0.03,
+				"cacheWrite": 0.375
 			},
 			"contextWindow": 204800,
 			"maxTokens": 131072,
@@ -40270,8 +40288,8 @@
 			"cost": {
 				"input": 0.3,
 				"output": 1.2,
-				"cacheRead": 0,
-				"cacheWrite": 0
+				"cacheRead": 0.03,
+				"cacheWrite": 0.375
 			},
 			"contextWindow": 204800,
 			"maxTokens": 131072,
@@ -45440,7 +45458,7 @@
 			"api": "openai-completions",
 			"provider": "nanogpt",
 			"baseUrl": "https://nano-gpt.com/api/v1",
-			"reasoning": false,
+			"reasoning": true,
 			"input": [
 				"text"
 			],
@@ -45453,7 +45471,14 @@
 			"contextWindow": 1048576,
 			"maxTokens": 384000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUseConfig": false,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"high",
+					"max"
+				]
+			}
 		},
 		"deepseek/deepseek-v4-flash:thinking": {
 			"id": "deepseek/deepseek-v4-flash:thinking",
@@ -68099,7 +68124,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 1048576,
+			"maxTokens": 65536,
 			"omitMaxOutputTokens": true,
 			"thinking": {
 				"mode": "effort",
@@ -68139,7 +68164,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 8192,
+			"maxTokens": 65536,
 			"omitMaxOutputTokens": true,
 			"supportsComputerUse": false
 		},
@@ -68160,7 +68185,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 1048576,
+			"maxTokens": 65536,
 			"omitMaxOutputTokens": true,
 			"thinking": {
 				"mode": "effort",
@@ -74174,6 +74199,33 @@
 			"supportsComputerUse": false,
 			"supportsComputerUseConfig": false
 		},
+		"~deepseek/deepseek-v4-flash-latest": {
+			"id": "~deepseek/deepseek-v4-flash-latest",
+			"name": "DeepSeek V4 Flash Latest",
+			"api": "openrouter",
+			"provider": "openrouter",
+			"baseUrl": "https://openrouter.ai/api/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.09,
+				"output": 0.18,
+				"cacheRead": 0.018,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1048576,
+			"maxTokens": 65536,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"high"
+				]
+			},
+			"supportsComputerUse": false,
+			"supportsComputerUseConfig": false
+		},
 		"~google/gemini-flash-latest": {
 			"id": "~google/gemini-flash-latest",
 			"name": "Gemini Flash Latest",
@@ -76370,17 +76422,19 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.14,
-				"output": 0.28,
-				"cacheRead": 0.0028,
+				"input": 0.09,
+				"output": 0.18,
+				"cacheRead": 0.018,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 384000,
+			"maxTokens": 65536,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
-					"high"
+					"low",
+					"high",
+					"max"
 				]
 			},
 			"supportsComputerUse": false,
@@ -81508,7 +81562,7 @@
 			],
 			"cost": {
 				"input": 0.03,
-				"output": 0.14,
+				"output": 0.13,
 				"cacheRead": 0.03,
 				"cacheWrite": 0
 			},
@@ -85391,9 +85445,9 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.76006,
-				"output": 2.38876,
-				"cacheRead": 0.141154,
+				"input": 0.42,
+				"output": 1.32,
+				"cacheRead": 0.078,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
@@ -85615,8 +85669,6 @@
 			},
 			"contextWindow": 524288,
 			"maxTokens": 65536,
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
@@ -85629,7 +85681,8 @@
 					"max": "max"
 				},
 				"requiresEffort": true
-			}
+			},
+			"supportsComputerUse": false
 		},
 		"hf:nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4": {
 			"id": "hf:nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4",
@@ -85659,8 +85712,7 @@
 					"xhigh"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"hf:openai/gpt-oss-120b": {
 			"id": "hf:openai/gpt-oss-120b",
@@ -85688,8 +85740,7 @@
 					"high"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"hf:Qwen/Qwen3.6-27B": {
 			"id": "hf:Qwen/Qwen3.6-27B",
@@ -85719,8 +85770,7 @@
 					"high"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"hf:zai-org/GLM-4.7-Flash": {
 			"id": "hf:zai-org/GLM-4.7-Flash",
@@ -85750,8 +85800,7 @@
 					"xhigh"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"hf:zai-org/GLM-5.2": {
 			"id": "hf:zai-org/GLM-5.2",
@@ -85781,8 +85830,7 @@
 					"xhigh"
 				]
 			},
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUse": false
 		},
 		"syn:large:text": {
 			"id": "syn:large:text",
@@ -86536,7 +86584,7 @@
 			"cost": {
 				"input": 0.6,
 				"output": 3.6,
-				"cacheRead": 0,
+				"cacheRead": 0.35,
 				"cacheWrite": 0
 			},
 			"contextWindow": 262144,
@@ -86621,7 +86669,7 @@
 			"cost": {
 				"input": 1.25,
 				"output": 3.75,
-				"cacheRead": 0,
+				"cacheRead": 0.125,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1000000,
@@ -86728,7 +86776,7 @@
 			"cost": {
 				"input": 1.4,
 				"output": 4.4,
-				"cacheRead": 0,
+				"cacheRead": 0.26,
 				"cacheWrite": 0
 			},
 			"contextWindow": 202752,
@@ -87675,26 +87723,28 @@
 		},
 		"deepseek-v4-flash-0731": {
 			"id": "deepseek-v4-flash-0731",
-			"name": "deepseek-v4-flash-0731",
+			"name": "DeepSeek V4 Flash 0731",
 			"api": "openai-completions",
 			"provider": "venice",
 			"baseUrl": "https://api.venice.ai/api/v1",
-			"reasoning": false,
+			"reasoning": true,
 			"input": [
 				"text"
 			],
 			"cost": {
-				"input": 0,
-				"output": 0,
-				"cacheRead": 0,
+				"input": 0.175,
+				"output": 0.35,
+				"cacheRead": 0.035,
 				"cacheWrite": 0
 			},
 			"contextWindow": 1000000,
-			"maxTokens": 1048576,
-			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
-			"compat": {
-				"supportsUsageInStreaming": false
+			"maxTokens": 32768,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"high",
+					"max"
+				]
 			}
 		},
 		"deepseek-v4-pro": {
@@ -90547,9 +90597,9 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.125,
-				"output": 0.5,
-				"cacheRead": 0,
+				"input": 0.06,
+				"output": 0.4,
+				"cacheRead": 0.01,
 				"cacheWrite": 0
 			},
 			"contextWindow": 128000,
@@ -93104,7 +93154,8 @@
 			"baseUrl": "https://ai-gateway.vercel.sh",
 			"reasoning": true,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0.15,
@@ -93174,7 +93225,8 @@
 			"baseUrl": "https://ai-gateway.vercel.sh",
 			"reasoning": true,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0.74,
@@ -93904,7 +93956,8 @@
 			"provider": "vercel-ai-gateway",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0.09999999999999999,
@@ -93924,7 +93977,8 @@
 			"provider": "vercel-ai-gateway",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0.15,
@@ -94017,7 +94071,8 @@
 			"baseUrl": "https://ai-gateway.vercel.sh",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0.15,
@@ -98492,7 +98547,6 @@
 			"contextWindow": 2000000,
 			"maxTokens": 2000000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98524,7 +98578,6 @@
 			"contextWindow": 2000000,
 			"maxTokens": 2000000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98568,7 +98621,6 @@
 				}
 			},
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98613,7 +98665,6 @@
 				}
 			},
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98658,7 +98709,6 @@
 				}
 			},
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98690,7 +98740,6 @@
 			"contextWindow": 512000,
 			"maxTokens": 512000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98722,7 +98771,6 @@
 			"contextWindow": 256000,
 			"maxTokens": 256000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -98753,7 +98801,6 @@
 			"contextWindow": 200000,
 			"maxTokens": 200000,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -105626,6 +105673,37 @@
 				]
 			}
 		},
+		"glm-5.2-highspeed": {
+			"id": "glm-5.2-highspeed",
+			"name": "GLM-5.2 Highspeed",
+			"api": "openai-completions",
+			"provider": "zhipu-coding-plan",
+			"baseUrl": "https://open.bigmodel.cn/api/coding/paas/v4",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 131072,
+			"compat": {
+				"thinkingFormat": "zai",
+				"reasoningContentField": "reasoning_content",
+				"supportsDeveloperRole": false
+			},
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"high",
+					"max"
+				]
+			}
+		},
 		"glm-5.2-highspeed[1m]": {
 			"id": "glm-5.2-highspeed[1m]",
 			"name": "GLM-5.2 Highspeed",

package/src/provider-models/ollama.ts

@@ -23,6 +23,36 @@ type OllamaShowResponse = {
 };
 
 const OLLAMA_RETRY_DELAYS_MS = [2_000, 5_000, 10_000];
+/**
+ * Output-token ceiling that Ollama Cloud enforces for the DeepSeek V4 Pro/Flash
+ * deployments: `/api/chat` rejects `num_predict` above it with HTTP 400
+ * (`max_tokens (...) exceeds model's maximum output tokens (65536)`) even though
+ * the model pages advertise a 1M context / 384K output. Ollama's `/api/show`
+ * never reports this cap, so the catalog pins it for the affected models
+ * (ollama/ollama#16890, #7266). The wire layer clamps `num_predict` to the same
+ * value (`OLLAMA_CLOUD_NUM_PREDICT_CAP` in `packages/ai/src/providers/ollama.ts`,
+ * #3392/#3394).
+ */
+export const OLLAMA_CLOUD_MAX_OUTPUT_TOKENS = 65_536;
+
+/**
+ * Untagged base ids whose Ollama Cloud deployment enforces
+ * {@link OLLAMA_CLOUD_MAX_OUTPUT_TOKENS}. Only DeepSeek V4 Pro/Flash are known
+ * to cap output below their advertised window (ollama/ollama#16890); other cloud
+ * models keep their discovered limits.
+ */
+const OLLAMA_CLOUD_OUTPUT_CAPPED_BASE_IDS: Record<string, true> = {
+	"deepseek-v4-flash": true,
+	"deepseek-v4-pro": true,
+};
+
+/** Whether an Ollama Cloud model id (tagged or not) enforces the 65536 output cap. */
+export function isOllamaCloudOutputCapped(id: string): boolean {
+	const separator = id.indexOf(":");
+	const baseId = separator > 0 ? id.slice(0, separator) : id;
+	return OLLAMA_CLOUD_OUTPUT_CAPPED_BASE_IDS[baseId] === true;
+}
+
 const OLLAMA_CLOUD_GLM_52_THINKING: ThinkingConfig = {
 	mode: "effort",
 	efforts: [Effort.High, Effort.Max],
@@ -133,10 +163,10 @@ export function ollamaCloudModelManagerOptions(
 					}
 					const capabilities = metadata?.capabilities;
 					const discoveredContextWindow = getContextWindow(metadata?.model_info);
-					// `/api/show` is the only trustworthy Ollama-owned source for size caps.
-					// When it is unavailable (or returns only coarse capabilities), do NOT
-					// inherit giant budgets from bundled fallback metadata sourced from a
-					// different catalog; keep the historical safe fallback instead.
+					// `/api/show` reports the context length but never a per-model output
+					// cap. DeepSeek V4 Pro/Flash deployments enforce a 65536 output ceiling
+					// (ollama/ollama#16890, #7266); every other id keeps the trusted
+					// reference limit, falling back to the historical safe cap otherwise.
 					const contextWindow = discoveredContextWindow ?? 128000;
 					const reasoning = capabilities ? capabilities.includes("thinking") : (reference?.reasoning ?? false);
 					const thinking = capabilities ? getThinkingConfig(id, capabilities) : reference?.thinking;
@@ -157,8 +187,9 @@ export function ollamaCloudModelManagerOptions(
 						input,
 						cost: reference?.cost ?? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 						contextWindow,
-						maxTokens:
-							discoveredContextWindow !== null && discoveredContextWindow !== undefined
+						maxTokens: isOllamaCloudOutputCapped(id)
+							? Math.min(contextWindow, OLLAMA_CLOUD_MAX_OUTPUT_TOKENS)
+							: discoveredContextWindow !== null && discoveredContextWindow !== undefined
 								? (providerReference?.maxTokens ?? Math.min(contextWindow, 8192))
 								: Math.min(contextWindow, 8192),
 						omitMaxOutputTokens: true,

package/src/provider-models/openai-compat.ts

@@ -2465,6 +2465,24 @@ export interface OpenRouterModelManagerConfig {
 	fetch?: FetchImpl;
 }
 
+function mapOpenRouterThinking(entry: OpenAICompatibleModelRecord): ThinkingConfig | undefined {
+	const reasoning = entry.reasoning;
+	if (!isRecord(reasoning)) return undefined;
+	const supportedEfforts = reasoning.supported_efforts;
+	if (!Array.isArray(supportedEfforts)) return undefined;
+	const efforts = THINKING_EFFORTS.filter(effort => supportedEfforts.includes(effort));
+	if (efforts.length === 0) return undefined;
+	const defaultLevel =
+		typeof reasoning.default_effort === "string"
+			? THINKING_EFFORTS.find(effort => effort === reasoning.default_effort)
+			: undefined;
+	return {
+		mode: "effort",
+		efforts,
+		...(defaultLevel !== undefined && efforts.includes(defaultLevel) ? { defaultLevel } : {}),
+	};
+}
+
 export function openrouterModelManagerOptions(
 	config?: OpenRouterModelManagerConfig,
 ): ModelManagerOptions<"openrouter"> {
@@ -2496,6 +2514,7 @@ export function openrouterModelManagerOptions(
 					const baseModel = mapWithBundledReference(entry, defaults, reference);
 					const pricing = entry.pricing as Record<string, unknown> | undefined;
 					const params = Array.isArray(entry.supported_parameters) ? (entry.supported_parameters as string[]) : [];
+					const thinking = mapOpenRouterThinking(entry);
 					const modality = String((entry.architecture as Record<string, unknown> | undefined)?.modality ?? "");
 					const topProvider = entry.top_provider as Record<string, unknown> | undefined;
 
@@ -2504,6 +2523,7 @@ export function openrouterModelManagerOptions(
 					return {
 						...baseModel,
 						reasoning: params.includes("reasoning"),
+						...(thinking !== undefined ? { thinking } : {}),
 						input: modality.includes("image") ? ["text", "image"] : ["text"],
 						cost: {
 							input: parseFloat(String(pricing?.prompt ?? "0")) * 1_000_000,

package/src/types.ts

@@ -401,6 +401,15 @@ export interface OpenAICompat {
  * that proxy gateways (Vertex AI, AWS Bedrock-style fronts, etc.) reject.
  */
 export interface AnthropicCompat {
+	/**
+	 * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts.
+	 * Set to 0 to disable the inter-event idle watchdog entirely, matching
+	 * `OpenAICompat.streamIdleTimeoutMs`.
+	 *
+	 * When unset, direct Anthropic streams use `PI_STREAM_IDLE_TIMEOUT_MS`,
+	 * then the legacy `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+	 */
+	streamIdleTimeoutMs?: number;
 	/**
 	 * Drop the top-level `strict: true` field on tool definitions. Vertex AI's
 	 * Anthropic-compatible endpoint rejects unknown tool fields with
@@ -712,7 +721,13 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs">> & {
+	/**
+	 * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts; 0 disables the idle watchdog.
+	 * Undefined defers to `PI_STREAM_IDLE_TIMEOUT_MS`, then the legacy
+	 * `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+	 */
+	streamIdleTimeoutMs?: number;
 	/**
 	 * The configured endpoint is the official first-party Anthropic API
 	 * (https + exact `api.anthropic.com` host; a missing baseUrl counts as