@just-every/ensemble 0.2.267 → 0.2.269

package/README.md

@@ -35,12 +35,13 @@ See the [demo README](demo/README.md) for detailed information about each demo.
 - 📊 **Cost & Quota Tracking** - Built-in usage monitoring and cost calculation
 - 🎯 **Smart Result Processing** - Automatic summarization and truncation for long outputs
 
-## Model Updates (Dec 2025)
+## Model Highlights
 
 - OpenAI: Added GPT-5.2 (base + chat-latest + pro) and refreshed GPT-5.1/GPT-5/Codex pricing
 - Anthropic: Claude 4.5 (Sonnet/Haiku, incl. 1M context) and Claude Opus 4.1
-- Google: Gemini 3 (Pro/Flash/Ultra) and refreshed Gemini 2.5 pricing incl. image/TTS/native-audio
+- Google: Gemini 3.6 Flash, Gemini 3.5 Flash/Lite, Gemini 3.1 Pro/Image/Live/TTS, and Gemini 2.5
 - xAI: Grok 4.5 with current `grok-build-latest` aliasing and reasoning controls, Grok 4.1 Fast / Grok 4 Fast variants, and Grok Imagine image generation/editing support (`grok-imagine-image`, `grok-imagine-image-pro`)
+- OpenRouter: Kimi K3, LongCat 2.0, Inkling, Muse Spark 1.1, and KAT-Coder Pro/Air V2.5 with current routing, capability, context, and pricing metadata
 
 \*Codex-Max pricing reflects current published rates and may change if OpenAI updates pricing.
 
@@ -228,7 +229,7 @@ Gemini 3 native thinking levels can be requested directly:
 ```ts
 const result = await ensembleResult(
     ensembleRequest([{ type: 'message', role: 'user', content: 'Solve this carefully.' }], {
-        model: 'gemini-3-flash-preview',
+        model: 'gemini-3.6-flash',
         modelSettings: {
             thinking_level: 'high',
         },
@@ -240,6 +241,7 @@ For Gemini, `thinking_level` is mutually exclusive with numeric `thinking_budget
 When a Gemini model supports native thinking levels, suffixes like `-none`, `-minimal`, `-low`, `-medium`, `-high`, and `-max` are sent as `thinkingLevel` values instead of numeric budgets.
 For native Gemini thinking-level models, numeric `thinking_budget` values are also translated to the closest supported `thinkingLevel`; Gemini 2.5 models still receive `thinkingBudget`.
 For Gemini 3 Flash-family models, `-none` maps to the provider-native `minimal` level because Gemini does not support full thinking-off.
+Gemini 3.6 Flash and Gemini 3.5 Flash-Lite ignore deprecated sampling settings (`temperature`, `top_p`, and `top_k`), so Ensemble omits them from those requests. Their message histories must end with a user turn rather than a prefilled model response.
 
 ### Multimodal Input (Images)
 
@@ -271,7 +273,7 @@ const messages = [
     },
 ];
 
-for await (const event of ensembleRequest(messages, { model: 'gemini-3-flash-preview' })) {
+for await (const event of ensembleRequest(messages, { model: 'gemini-3.6-flash' })) {
     if (event.type === 'message_complete' && 'content' in event) {
         console.log(event.content);
     }
@@ -290,9 +292,8 @@ The example below combines **image input** with **JSON output**:
 import { ensembleRequest, ensembleResult } from '@just-every/ensemble';
 
 const agent = {
-    model: 'gemini-3-flash-preview',
+    model: 'gemini-3.6-flash',
     modelSettings: {
-        temperature: 0.2,
         json_schema: {
             name: 'image_analysis',
             type: 'json_schema',

package/dist/cjs/data/model_data.cjs

@@ -7,7 +7,7 @@ exports.MODEL_CLASSES = {
     standard: {
         models: [
             'gpt-5.5',
-            'gemini-3-flash-preview',
+            'gemini-3.6-flash',
             'claude-sonnet-5',
             'grok-4.5',
         ],
@@ -16,7 +16,7 @@ exports.MODEL_CLASSES = {
     mini: {
         models: [
             'gpt-5.4-nano',
-            'gemini-2.5-flash-lite',
+            'gemini-3.5-flash-lite',
             'claude-haiku-4-5-20251001',
             'grok-3-mini',
         ],
@@ -43,7 +43,7 @@ exports.MODEL_CLASSES = {
     reasoning_mini: {
         models: [
             'gpt-5.4-mini',
-            'gemini-3-flash-preview',
+            'gemini-3.5-flash-lite',
             'claude-sonnet-5',
             'grok-3-mini',
         ],
@@ -80,7 +80,7 @@ exports.MODEL_CLASSES = {
     writing: {
         models: [
             'gpt-5.5',
-            'gemini-3-flash-preview',
+            'gemini-3.6-flash',
             'claude-sonnet-5',
             'grok-4.5',
         ],
@@ -89,7 +89,7 @@ exports.MODEL_CLASSES = {
     summary: {
         models: [
             'gpt-5-nano',
-            'gemini-2.5-flash-lite',
+            'gemini-3.5-flash-lite',
             'claude-haiku-4-5-20251001',
             'grok-3-mini',
         ],
@@ -107,7 +107,7 @@ exports.MODEL_CLASSES = {
     vision_mini: {
         models: [
             'gpt-5.4-mini',
-            'gemini-3-flash-preview',
+            'gemini-3.5-flash-lite',
             'claude-haiku-4-5-20251001',
             'grok-3-mini',
         ],
@@ -2369,6 +2369,27 @@ exports.MODEL_REGISTRY = [
         class: 'standard',
         description: 'Gemini 3 Flash Preview - fast multimodal model with 1M context window.',
     },
+    {
+        id: 'gemini-3.6-flash',
+        aliases: ['gemini-flash-latest', 'models/gemini-3.6-flash', 'models/gemini-flash-latest'],
+        provider: 'google',
+        cost: {
+            input_per_million: 1.5,
+            output_per_million: 7.5,
+            cached_input_per_million: 0.15,
+        },
+        features: {
+            context_length: 1048576,
+            input_modality: ['text', 'image', 'video', 'audio'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            max_output_tokens: 65536,
+        },
+        class: 'standard',
+        description: 'Gemini 3.6 Flash - stable frontier multimodal model optimized for agentic execution, coding, spatial reasoning, and efficient tool use.',
+    },
     {
         id: 'gemini-3.5-flash',
         provider: 'google',
@@ -2390,6 +2411,27 @@ exports.MODEL_REGISTRY = [
         score: 88,
         description: 'Gemini 3.5 Flash - fast multimodal reasoning model with 1M context and native thinking levels.',
     },
+    {
+        id: 'gemini-3.5-flash-lite',
+        aliases: ['gemini-flash-lite-latest', 'models/gemini-3.5-flash-lite', 'models/gemini-flash-lite-latest'],
+        provider: 'google',
+        cost: {
+            input_per_million: 0.3,
+            output_per_million: 2.5,
+            cached_input_per_million: 0.03,
+        },
+        features: {
+            context_length: 1048576,
+            input_modality: ['text', 'image', 'video', 'audio'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            max_output_tokens: 65536,
+        },
+        class: 'reasoning_mini',
+        description: 'Gemini 3.5 Flash-Lite - stable low-latency multimodal model for high-throughput subagents, document parsing, extraction, and structured data tasks.',
+    },
     {
         id: 'gemini-3.1-flash-lite',
         aliases: ['gemini-3.1-flash-lite-preview'],
@@ -3923,6 +3965,143 @@ exports.MODEL_REGISTRY = [
         class: 'code',
         description: 'MoonshotAI Kimi K2.7 Code via OpenRouter. Coding-focused long-context model for software engineering and agentic coding workflows.',
     },
+    {
+        id: 'moonshotai/kimi-k3',
+        aliases: ['Kimi K3', 'kimi-k3'],
+        provider: 'openrouter',
+        openrouter_id: 'moonshotai/kimi-k3',
+        cost: {
+            input_per_million: 3.0,
+            cached_input_per_million: 0.3,
+            output_per_million: 15.0,
+        },
+        features: {
+            context_length: 1_048_576,
+            input_modality: ['text', 'image'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            structured_output: true,
+            reasoning_output: true,
+        },
+        class: 'reasoning',
+        description: 'MoonshotAI Kimi K3 via OpenRouter. Open-weight multimodal model for complex coding, knowledge work, and long-horizon agentic workflows with 1M context.',
+    },
+    {
+        id: 'meituan/longcat-2.0',
+        aliases: ['LongCat 2.0', 'LongCat-2.0', 'longcat-2.0', 'longcat-2'],
+        provider: 'openrouter',
+        openrouter_id: 'meituan/longcat-2.0',
+        cost: {
+            input_per_million: 0.3,
+            cached_input_per_million: 0.006,
+            output_per_million: 1.2,
+        },
+        features: {
+            context_length: 1_048_576,
+            input_modality: ['text'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: false,
+            max_output_tokens: 262144,
+            reasoning_output: true,
+        },
+        class: 'reasoning',
+        description: 'Meituan LongCat 2.0 via OpenRouter. Sparse MoE model for repository-level coding, long-horizon problem solving, and agentic workflows with 1M context.',
+    },
+    {
+        id: 'thinkingmachines/inkling',
+        aliases: ['Inkling', 'inkling'],
+        provider: 'openrouter',
+        openrouter_id: 'thinkingmachines/inkling',
+        cost: {
+            input_per_million: 1.0,
+            cached_input_per_million: 0.17,
+            output_per_million: 4.05,
+        },
+        features: {
+            context_length: 1_048_576,
+            input_modality: ['text', 'image', 'audio'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: false,
+            reasoning_output: true,
+        },
+        class: 'reasoning',
+        description: 'Thinking Machines Inkling via OpenRouter. Open-weight multimodal model for general reasoning, coding, tool use, and agentic applications with 1M context.',
+    },
+    {
+        id: 'meta/muse-spark-1.1',
+        aliases: ['Muse Spark 1.1', 'Muse-Spark-1.1', 'muse-spark-1.1'],
+        provider: 'openrouter',
+        openrouter_id: 'meta/muse-spark-1.1',
+        cost: {
+            input_per_million: 1.25,
+            cached_input_per_million: 0.15,
+            output_per_million: 4.25,
+        },
+        features: {
+            context_length: 1_048_576,
+            input_modality: ['text', 'image', 'audio', 'video'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            structured_output: true,
+            reasoning_output: true,
+        },
+        class: 'reasoning',
+        description: 'Meta Muse Spark 1.1 via OpenRouter. Multimodal reasoning model for agentic workflows, coding, tool use, and visual-to-code generation with 1M context.',
+    },
+    {
+        id: 'kwaipilot/kat-coder-pro-v2.5',
+        aliases: ['KAT-Coder-Pro V2.5', 'KAT-Coder-Pro-V2.5', 'kat-coder-pro-v2.5', 'kat-coder-pro-v2-5'],
+        provider: 'openrouter',
+        openrouter_id: 'kwaipilot/kat-coder-pro-v2.5',
+        cost: {
+            input_per_million: 0.74,
+            cached_input_per_million: 0.15,
+            output_per_million: 2.96,
+        },
+        features: {
+            context_length: 256000,
+            input_modality: ['text'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            structured_output: true,
+            max_output_tokens: 80000,
+        },
+        class: 'code',
+        description: 'Kwaipilot KAT-Coder Pro V2.5 via OpenRouter. Agentic coding model for autonomous repository-level software engineering and tool-driven workflows.',
+    },
+    {
+        id: 'kwaipilot/kat-coder-air-v2.5',
+        aliases: ['KAT-Coder-Air V2.5', 'KAT-Coder-Air-V2.5', 'kat-coder-air-v2.5', 'kat-coder-air-v2-5'],
+        provider: 'openrouter',
+        openrouter_id: 'kwaipilot/kat-coder-air-v2.5',
+        cost: {
+            input_per_million: 0.15,
+            cached_input_per_million: 0.03,
+            output_per_million: 0.6,
+        },
+        features: {
+            context_length: 256000,
+            input_modality: ['text'],
+            output_modality: ['text'],
+            tool_use: true,
+            streaming: true,
+            json_output: true,
+            structured_output: true,
+            max_output_tokens: 80000,
+        },
+        class: 'code',
+        description: 'Kwaipilot KAT-Coder Air V2.5 via OpenRouter. Cost-efficient agentic coding model for autonomous repository changes and tool-driven workflows.',
+    },
 ];
 function findModel(modelId) {
     const externalModel = (0, external_models_js_1.getExternalModel)(modelId);

package/dist/cjs/data/model_data.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model_data.d.ts","sourceRoot":"","sources":["../../../data/model_data.ts"],"names":[],"mappings":"AAYA,OAAO,EACH,YAAY,EACZ,eAAe,EACf,WAAW,EACX,cAAc,EACd,aAAa,EACb,SAAS,EACT,aAAa,EACb,UAAU,EACV,UAAU,EACV,UAAU,EACb,MAAM,mBAAmB,CAAC;AAM3B,YAAY,EACR,YAAY,EACZ,eAAe,EACf,WAAW,EACX,cAAc,EACd,aAAa,EACb,SAAS,EACT,aAAa,EACb,UAAU,EACV,UAAU,EACV,UAAU,GACb,CAAC;AAOF,eAAO,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0MzB,CAAC;AAGF,eAAO,MAAM,cAAc,EAAE,UAAU,EAs7HtC,CAAC;AAQF,wBAAgB,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,UAAU,GAAG,SAAS,CAuBjE"}
+{"version":3,"file":"model_data.d.ts","sourceRoot":"","sources":["../../../data/model_data.ts"],"names":[],"mappings":"AAYA,OAAO,EACH,YAAY,EACZ,eAAe,EACf,WAAW,EACX,cAAc,EACd,aAAa,EACb,SAAS,EACT,aAAa,EACb,UAAU,EACV,UAAU,EACV,UAAU,EACb,MAAM,mBAAmB,CAAC;AAM3B,YAAY,EACR,YAAY,EACZ,eAAe,EACf,WAAW,EACX,cAAc,EACd,aAAa,EACb,SAAS,EACT,aAAa,EACb,UAAU,EACV,UAAU,EACV,UAAU,GACb,CAAC;AAOF,eAAO,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0MzB,CAAC;AAGF,eAAO,MAAM,cAAc,EAAE,UAAU,EA4nItC,CAAC;AAQF,wBAAgB,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,UAAU,GAAG,SAAS,CAuBjE"}