Remove dead temperature param, add 2-tier pricing, clarify strict stripping

Author: capitanfeeder

packages/types/src/providers/mimo.ts

@@ -2,6 +2,7 @@ import type { ModelInfo } from "../model.js"
 
 // https://developer.puter.com/ai/xiaomi/mimo-v2.5-pro/
 // https://developer.puter.com/ai/xiaomi/mimo-v2.5/
+// https://platform.xiaomimimo.com/docs/en-US/quick-start/model-hyperparameters
 //
 // NOTE: mimo-v2-flash is not included here. Its thinking mode defaults to
 // disabled and it doesn't reliably handle reasoning_content passthrough
@@ -21,9 +22,18 @@ export const mimoModels = {
 		supportsPromptCache: false,
 		preserveReasoning: true,
 		inputPrice: 1.0, // $1.00/1M tokens (cache miss, ≤256K)
-		outputPrice: 3.0, // $3.00/1M tokens
-		cacheReadsPrice: 0.2, // $0.20/1M tokens (cache hit)
+		outputPrice: 3.0, // $3.00/1M tokens (≤256K)
+		cacheReadsPrice: 0.2, // $0.20/1M tokens (cache hit, ≤256K)
 		cacheWritesPrice: 0, // Free for limited time
+		// MiMo charges 2x for input >256K context
+		tiers: [
+			{
+				contextWindow: 256_001,
+				inputPrice: 2.0, // $2.00/1M tokens (>256K)
+				outputPrice: 6.0, // $6.00/1M tokens (>256K)
+				cacheReadsPrice: 0.4, // $0.40/1M tokens (>256K)
+			},
+		],
 		description:
 			"MiMo V2.5 Pro - Xiaomi's flagship reasoning model with 1M context, deep thinking, tool calling, and structured output.",
 	},
@@ -34,9 +44,18 @@ export const mimoModels = {
 		supportsPromptCache: false,
 		preserveReasoning: true,
 		inputPrice: 0.4, // $0.40/1M tokens (cache miss, ≤256K)
-		outputPrice: 2.0, // $2.00/1M tokens
-		cacheReadsPrice: 0.08, // $0.08/1M tokens (cache hit)
+		outputPrice: 2.0, // $2.00/1M tokens (≤256K)
+		cacheReadsPrice: 0.08, // $0.08/1M tokens (cache hit, ≤256K)
 		cacheWritesPrice: 0, // Free for limited time
+		// MiMo charges 2x for input >256K context
+		tiers: [
+			{
+				contextWindow: 256_001,
+				inputPrice: 0.8, // $0.80/1M tokens (>256K)
+				outputPrice: 4.0, // $4.00/1M tokens (>256K)
+				cacheReadsPrice: 0.16, // $0.16/1M tokens (>256K)
+			},
+		],
 		description:
 			"MiMo V2.5 - Full-modal understanding model (text, image, audio, video) with 1M context, deep thinking, tool calling, and structured output.",
 	},

src/api/providers/mimo.ts

@@ -54,9 +54,10 @@ export class MimoHandler extends OpenAiHandler {
 	}
 
 	/**
-	 * Strip OpenAI-specific extensions that MiMo's proxy rejects:
-	 * - strict: true on tools
-	 * - additionalProperties: false on schemas
+	 * Strip OpenAI-specific extensions that MiMo's Token Plan proxy rejects.
+	 * The official API docs list `strict` as supported (default false), but the
+	 * proxy endpoints have historically returned 400 for it. Kept as a safety
+	 * net until the proxy is confirmed to handle these fields correctly.
 	 */
 	protected override convertToolsForOpenAI(tools: any[] | undefined): any[] | undefined {
 		if (!tools) {
@@ -124,7 +125,7 @@ export class MimoHandler extends OpenAiHandler {
 		messages: any[],
 		metadata?: ApiHandlerCreateMessageMetadata,
 	): ApiStream {
-		const { id: modelId, info: modelInfo, temperature } = this.getModel()
+		const { id: modelId, info: modelInfo } = this.getModel()
 
 		// Use shared R1-format conversion with tool ID sanitization and text merging
 		const convertedMessages = convertToR1Format(messages, {
@@ -136,9 +137,10 @@ export class MimoHandler extends OpenAiHandler {
 
 		// Build request per MiMo's OpenAI-compatible API
 		// https://developer.puter.com/ai/xiaomi/mimo-v2.5-pro/
+		// Note: temperature is omitted because MiMo forces it to 1.0 when thinking mode
+		// is enabled, regardless of what is passed (see model-hyperparameters docs).
 		const params: Record<string, any> = {
 			model: modelId,
-			temperature,
 			messages: [{ role: "system", content: systemPrompt }, ...convertedMessages],
 			stream: true,
 			stream_options: { include_usage: true },