ugly-app 0.1.430 → 0.1.431

package/src/shared/TextGen.ts

@@ -28,7 +28,6 @@ export type TextGenTogetherModelT = (typeof textGenTogetherModels)[number];
 // Note: Llama 4 models require enterprise deployment on Fireworks, use Groq/Together instead
 export const textGenFireworksModels = [
   'fireworks_gpt_oss_120b',
-  'fireworks_gpt_oss_20b',
 ] as const;
 
 export type TextGenFireworksModelT = (typeof textGenFireworksModels)[number];
@@ -53,11 +52,9 @@ export type TextGenDeepSeekModelT = (typeof textGenDeepSeekModels)[number];
 // Groq models
 export const textGenGroqModels = [
   'groq_llama_4_scout',
-  'groq_llama_4_maverick',
   'groq_mixtral_8x7b',
   'groq_llama_3_3_70b',
   'groq_gpt_oss_120b',
-  'groq_gpt_oss_20b',
 ] as const;
 
 export type TextGenGroqModelT = (typeof textGenGroqModels)[number];
@@ -90,7 +87,6 @@ export const textGenOpenAIModels = [
   'openai_o1_mini',
   'openai_o3',
   'openai_o4_mini',
-  'openai_gpt_41_nano',
 ] as const;
 
 export type TextGenOpenAIModelT = (typeof textGenOpenAIModels)[number];
@@ -108,8 +104,11 @@ export type TextGenMiniMaxModelT = (typeof textGenMiniMaxModels)[number];
 // us a warm path when the direct upstream is throttled or down.
 export const textGenOpenRouterModels = [
   'openrouter_glm_5_1',
+  'openrouter_glm_4_6',
   'openrouter_qwen_36_plus',
+  'openrouter_qwen3_6_27b',
   'openrouter_kimi_k2_thinking',
+  'openrouter_kimi_k2_6',
   'openrouter_gemma_4_31b',
   'openrouter_gemma_4_26b',
   'openrouter_gemma_4_26b_free',
@@ -121,6 +120,7 @@ export const textGenOpenRouterModels = [
   'openrouter_o4_mini',
   'openrouter_claude_sonnet_4_6',
   'openrouter_claude_haiku_4_5',
+  'openrouter_claude_opus_4_7',
 ] as const;
 export type TextGenOpenRouterModelT = (typeof textGenOpenRouterModels)[number];
 
@@ -132,9 +132,13 @@ export const textGenDeepInfraModels = [
   'deepinfra_gemma_4_31b',
   'deepinfra_gemma_4_26b',
   'deepinfra_gpt_oss_120b',
-  'deepinfra_gpt_oss_20b',
   'deepinfra_qwen3_235b',
-  'deepinfra_kimi_k2',
+  'deepinfra_qwen3_6_27b',
+  'deepinfra_qwen3_6_35b_a3b',
+  'deepinfra_kimi_k2_6',
+  'deepinfra_glm_4_6',
+  'deepinfra_glm_5_1',
+  'deepinfra_minimax_m2_5',
   'deepinfra_deepseek_v4_pro',
   'deepinfra_deepseek_v4_flash',
 ] as const;
@@ -165,24 +169,23 @@ const textGenProviderModelValues = [
   'openai_o1_mini',
   'openai_o3',
   'openai_o4_mini',
-  'openai_gpt_41_nano',
   // Fireworks (GPT-OSS - Llama 4 requires enterprise deployment)
   'fireworks_gpt_oss_120b',
-  'fireworks_gpt_oss_20b',
   // Groq
   'groq_llama_4_scout',
-  'groq_llama_4_maverick',
   'groq_mixtral_8x7b',
   'groq_llama_3_3_70b',
   'groq_gpt_oss_120b',
-  'groq_gpt_oss_20b',
   // MiniMax
   'minimax_m2_5',
   'minimax_m2_7',
   // OpenRouter
   'openrouter_glm_5_1',
+  'openrouter_glm_4_6',
   'openrouter_qwen_36_plus',
+  'openrouter_qwen3_6_27b',
   'openrouter_kimi_k2_thinking',
+  'openrouter_kimi_k2_6',
   'openrouter_gemma_4_31b',
   'openrouter_gemma_4_26b',
   'openrouter_gemma_4_26b_free',
@@ -193,13 +196,18 @@ const textGenProviderModelValues = [
   'openrouter_o4_mini',
   'openrouter_claude_sonnet_4_6',
   'openrouter_claude_haiku_4_5',
+  'openrouter_claude_opus_4_7',
   // DeepInfra
   'deepinfra_gemma_4_31b',
   'deepinfra_gemma_4_26b',
   'deepinfra_gpt_oss_120b',
-  'deepinfra_gpt_oss_20b',
   'deepinfra_qwen3_235b',
-  'deepinfra_kimi_k2',
+  'deepinfra_qwen3_6_27b',
+  'deepinfra_qwen3_6_35b_a3b',
+  'deepinfra_kimi_k2_6',
+  'deepinfra_glm_4_6',
+  'deepinfra_glm_5_1',
+  'deepinfra_minimax_m2_5',
   'deepinfra_deepseek_v4_pro',
   'deepinfra_deepseek_v4_flash',
   // DeepSeek (direct, BYO key, OpenAI-compatible)
@@ -225,13 +233,11 @@ export const textGenProviderModelsSet = new Set(textGenProviderModels);
 const textGenModelValues = [
   // Multi-provider routed models
   'llama_4_scout',
-  'llama_4_maverick',
   'llama_3_3_70b',
-  // DeepSeek V4 — direct route only (api.deepseek.com)
+  // DeepSeek V4 — direct (api.deepseek.com) + DeepInfra fallback
   'deepseek_v4_pro',
   'deepseek_v4_flash',
   'gpt_oss_120b',
-  'gpt_oss_20b',
   // Google
   'gemini_2_5_flash',
   'gemini_2_5',
@@ -239,24 +245,27 @@ const textGenModelValues = [
   'gpt_4o',
   'gpt_5',
   'gpt_5_mini',
-  'gpt_41_nano',
   'o3',
   'o4_mini',
   // Anthropic — coding-agent tiers (Anthropic direct + OpenRouter fallback)
   'claude_opus_4_7',
   'claude_sonnet_4_6',
-  'claude_3_sonnet',
   'claude_haiku_4_5',
   // Multi-provider open-weight (DeepInfra / OpenRouter / Together)
   'gemma_4_31b',
   'gemma_4_26b',
-  // MiniMax (direct)
+  // MiniMax — direct + DeepInfra fallback for M2.5
   'minimax_m2_5',
   'minimax_m2_7',
-  // OpenRouter-only frontier
+  // GLM — DeepInfra primary + OpenRouter fallback
   'glm_5_1',
+  'glm_4_6',
+  // Qwen 3.6 — OpenRouter + DeepInfra
   'qwen3_6_plus',
+  'qwen3_6_27b',
+  // Kimi — OpenRouter + DeepInfra
   'kimi_k2_thinking',
+  'kimi_k2_6',
   // Single-provider (Together — misc)
   'qwen2_72b',
   'qwen2_vision_72b',
@@ -524,25 +533,6 @@ export const TextGenOpenAIModelData: Record<
     smartness: 4,
     compactAt: 0.90,
   },
-  // GPT-4.1 Nano — cheapest viable model, 1M context
-  openai_gpt_41_nano: {
-    model: 'gpt-4.1-nano',
-    contextWindow: 1_000_000,
-    inputTokenNanoDollar: 100,
-    outputTokenNanoDollar: 400,
-    vision: true,
-    toolCalling: true,
-    jsonMode: true,
-    streaming: true,
-    parallelToolCalls: true,
-    // Coding agent
-    name: 'GPT-4.1 Nano',
-    provider: 'OpenAI',
-    speed: 'fast',
-    reasoning: 'strong',
-    smartness: 2,
-    compactAt: 0.90,
-  },
 };
 
 export const TextGenTogetherModelData: Record<
@@ -695,17 +685,6 @@ export const TextGenFireworksModelData: Record<
     streaming: true,
     parallelToolCalls: false,
   },
-  fireworks_gpt_oss_20b: {
-    model: 'accounts/fireworks/models/gpt-oss-20b',
-    contextWindow: 128000,
-    // $0.075 input / $0.30 output per 1M tokens
-    inputTokenNanoDollar: 75,
-    outputTokenNanoDollar: 300,
-    toolCalling: true,
-    jsonMode: true,
-    streaming: true,
-    parallelToolCalls: false,
-  },
 };
 
 // Groq models - ultra-fast inference
@@ -724,18 +703,6 @@ export const TextGenGroqModelData: Record<TextGenGroqModelT, TextGenModelData> =
       streaming: true,
       parallelToolCalls: false,
     },
-    groq_llama_4_maverick: {
-      model: 'meta-llama/llama-4-maverick-17b-128e-instruct',
-      contextWindow: 128000,
-      inputTokenNanoDollar: 200,
-      outputTokenNanoDollar: 600,
-      vision: true,
-      // Note: toolCalling marked 'unreliable' due to known 100% failure rate with structured prompts
-      toolCalling: 'unreliable',
-      jsonMode: true,
-      streaming: true,
-      parallelToolCalls: false,
-    },
     groq_mixtral_8x7b: {
       model: 'mixtral-8x7b-32768',
       contextWindow: 32768,
@@ -768,17 +735,6 @@ export const TextGenGroqModelData: Record<TextGenGroqModelT, TextGenModelData> =
       streaming: true,
       parallelToolCalls: false,
     },
-    groq_gpt_oss_20b: {
-      model: 'openai/gpt-oss-20b',
-      contextWindow: 128000,
-      // $0.075 input / $0.30 output per 1M tokens
-      inputTokenNanoDollar: 75,
-      outputTokenNanoDollar: 300,
-      toolCalling: true,
-      jsonMode: true,
-      streaming: true,
-      parallelToolCalls: false,
-    },
   };
 
 export const TextGenKieModelData: Record<TextGenKieModelT, TextGenModelData> =
@@ -1089,6 +1045,60 @@ export const TextGenOpenRouterModelData: Record<
     // doesn't assume a cache discount that never materializes.
     supportsCacheControl: false,
   },
+  // Claude Opus 4.7 fallback — primary still Anthropic direct. OpenRouter
+  // list price is ~5% over upstream. Cache passthrough verified to work
+  // for sonnet-4.6; opus-4.7 not yet verified — leave conservative.
+  openrouter_claude_opus_4_7: {
+    model: 'anthropic/claude-opus-4.7',
+    contextWindow: 200_000,
+    inputTokenNanoDollar: 15_750, // ~$15/M × 1.05
+    outputTokenNanoDollar: 78_750, // ~$75/M × 1.05
+    vision: true,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: true,
+    supportsCacheControl: false,
+  },
+  // GLM 4.6 OpenRouter route. List price is roughly the same as
+  // DeepInfra's published rate; OpenRouter wins when DeepInfra is
+  // throttled.
+  openrouter_glm_4_6: {
+    model: 'z-ai/glm-4.6',
+    contextWindow: 200_000,
+    inputTokenNanoDollar: 600,
+    outputTokenNanoDollar: 2200,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: false,
+  },
+  // Qwen 3.6 27B OpenRouter route — dense (non-MoE) variant.
+  openrouter_qwen3_6_27b: {
+    model: 'qwen/qwen3.6-27b',
+    contextWindow: 262_144,
+    inputTokenNanoDollar: 350,
+    outputTokenNanoDollar: 3360,
+    vision: true,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: false,
+  },
+  // Kimi K2.6 OpenRouter route.
+  openrouter_kimi_k2_6: {
+    model: 'moonshotai/kimi-k2.6',
+    contextWindow: 262_144,
+    inputTokenNanoDollar: 800,
+    outputTokenNanoDollar: 3700,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: false,
+  },
 };
 
 // DeepInfra — direct route, bypasses OpenRouter's shared rate limits.
@@ -1159,43 +1169,107 @@ export const TextGenDeepInfraModelData: Record<
     parallelToolCalls: false,
     supportsCacheControl: false,
   },
-  // GPT-OSS 20B — same model as Groq/Fireworks at lower price.
-  deepinfra_gpt_oss_20b: {
-    model: 'openai/gpt-oss-20b',
-    contextWindow: 128_000,
-    inputTokenNanoDollar: 30,
-    outputTokenNanoDollar: 150,
+  // Qwen3-235B-A22B-Instruct-2507 (live 2026-05-03 on DeepInfra).
+  deepinfra_qwen3_235b: {
+    model: 'Qwen/Qwen3-235B-A22B-Instruct-2507',
+    contextWindow: 256_000,
+    inputTokenNanoDollar: 71,
+    outputTokenNanoDollar: 100,
     toolCalling: true,
     jsonMode: true,
     streaming: true,
     parallelToolCalls: false,
     supportsCacheControl: false,
   },
-  // Qwen3-235B — DeepInfra at $0.071/$0.10 vs Together at $0.20/$0.60.
-  // ~3× cheaper input, ~6× cheaper output.
-  deepinfra_qwen3_235b: {
-    model: 'Qwen/Qwen3-235B-A22B-Instruct-2507',
-    contextWindow: 256_000,
-    inputTokenNanoDollar: 71,
-    outputTokenNanoDollar: 100,
+  // Qwen3.6 27B — $0.32/$3.20 per 1M (live 2026-05-03), 256k ctx,
+  // multimodal + reasoning. The dense, non-MoE variant.
+  deepinfra_qwen3_6_27b: {
+    model: 'Qwen/Qwen3.6-27B',
+    contextWindow: 262_144,
+    inputTokenNanoDollar: 320,
+    outputTokenNanoDollar: 3200,
+    vision: true,
     toolCalling: true,
     jsonMode: true,
     streaming: true,
     parallelToolCalls: false,
     supportsCacheControl: false,
   },
-  // Kimi K2 — DeepInfra at $0.50/$2.00 vs Together at $1.00/$3.00.
-  deepinfra_kimi_k2: {
-    model: 'moonshotai/Kimi-K2-Instruct',
-    contextWindow: 128_000,
-    inputTokenNanoDollar: 500,
-    outputTokenNanoDollar: 2000,
+  // Qwen3.6 35B-A3B — MoE variant, dramatically cheaper.
+  // $0.15/$0.95 per 1M (live 2026-05-03), 256k ctx.
+  deepinfra_qwen3_6_35b_a3b: {
+    model: 'Qwen/Qwen3.6-35B-A3B',
+    contextWindow: 262_144,
+    inputTokenNanoDollar: 150,
+    outputTokenNanoDollar: 950,
+    vision: true,
     toolCalling: true,
     jsonMode: true,
     streaming: true,
     parallelToolCalls: false,
     supportsCacheControl: false,
   },
+  // Kimi K2.6 — newer than the (now-removed) K2-Instruct. $0.75/$3.50
+  // per 1M (live 2026-05-03), 256k ctx, native reasoning. Cache-read
+  // is 20% of input → 150 nanoDollar/M.
+  deepinfra_kimi_k2_6: {
+    model: 'moonshotai/Kimi-K2.6',
+    contextWindow: 262_144,
+    inputTokenNanoDollar: 750,
+    outputTokenNanoDollar: 3500,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: true,
+    cacheReadTokenNanoDollar: 150,
+  },
+  // GLM 4.6 — $0.43/$1.74 per 1M (live 2026-05-03), 200k ctx.
+  // Cache-read 18.6% of input → 80 nanoDollar/M.
+  deepinfra_glm_4_6: {
+    model: 'zai-org/GLM-4.6',
+    contextWindow: 202_752,
+    inputTokenNanoDollar: 430,
+    outputTokenNanoDollar: 1740,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: true,
+    cacheReadTokenNanoDollar: 80,
+  },
+  // GLM 5.1 fallback route. Pricing $1.05/$3.50 per 1M (live
+  // 2026-05-03), 200k ctx. Slightly higher than the OpenRouter
+  // route's published list, so OpenRouter stays primary on price.
+  // Cache-read 19.5% of input → 205 nanoDollar/M.
+  deepinfra_glm_5_1: {
+    model: 'zai-org/GLM-5.1',
+    contextWindow: 202_752,
+    inputTokenNanoDollar: 1050,
+    outputTokenNanoDollar: 3500,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: true,
+    cacheReadTokenNanoDollar: 205,
+  },
+  // MiniMax M2.5 fallback. Pricing $0.15/$1.15 per 1M (live
+  // 2026-05-03) — actually CHEAPER than direct MiniMax ($0.30/$1.20),
+  // so the price-priority sort prefers DeepInfra. Direct stays as a
+  // healthy backup. 196k ctx, cache-read 20% → 30 nanoDollar/M.
+  deepinfra_minimax_m2_5: {
+    model: 'MiniMaxAI/MiniMax-M2.5',
+    contextWindow: 196_608,
+    inputTokenNanoDollar: 150,
+    outputTokenNanoDollar: 1150,
+    toolCalling: true,
+    jsonMode: true,
+    streaming: true,
+    parallelToolCalls: false,
+    supportsCacheControl: true,
+    cacheReadTokenNanoDollar: 30,
+  },
   // DeepSeek V4 Pro fallback route — used when api.deepseek.com is
   // throttled or down. Pricing live-verified 2026-05-03 against
   // /models/deepseek-ai/DeepSeek-V4-Pro: $1.74 in / $3.48 out per 1M
@@ -1376,22 +1450,6 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
-  llama_4_maverick: [
-    // Groq deprecated llama-4-maverick-17b-128e effective March 9, 2026
-    {
-      provider: 'groq',
-      providerModel: 'groq_llama_4_maverick',
-      latencyTier: 'fast',
-      available: false,
-    },
-    // Fireworks Llama 4 requires enterprise deployment, not available on serverless
-    {
-      provider: 'together',
-      providerModel: 'together_meta_llama4_400b',
-      latencyTier: 'standard',
-      available: true,
-    },
-  ],
   // DeepSeek V4 — direct route to api.deepseek.com (BYO key, OpenAI-
   // compatible) primary, DeepInfra fallback for failover only. Direct
   // is ~6× cheaper on Pro and ~2× on Flash so the price-priority sort
@@ -1456,26 +1514,6 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
-  gpt_oss_20b: [
-    {
-      provider: 'deepinfra',
-      providerModel: 'deepinfra_gpt_oss_20b',
-      latencyTier: 'standard',
-      available: true,
-    },
-    {
-      provider: 'groq',
-      providerModel: 'groq_gpt_oss_20b',
-      latencyTier: 'fast',
-      available: true,
-    },
-    {
-      provider: 'fireworks',
-      providerModel: 'fireworks_gpt_oss_20b',
-      latencyTier: 'standard',
-      available: true,
-    },
-  ],
   // Single-provider Google models
   gemini_2_5_flash: [
     {
@@ -1569,19 +1607,11 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
-  gpt_41_nano: [
-    {
-      provider: 'openai',
-      providerModel: 'openai_gpt_41_nano',
-      latencyTier: 'fast',
-      available: true,
-    },
-  ],
   // Anthropic models — direct first, OpenRouter fallback second.
-  // claude_3_sonnet historically maps to claude-sonnet-4-20250514;
-  // the OpenRouter fallback uses claude-sonnet-4.6 (current Anthropic
-  // live model) and has verified cache_control passthrough.
-  claude_3_sonnet: [
+  // The Anthropic entry ships 4.x behind a wire-dated snapshot, and
+  // OpenRouter mirrors the same model with ~5% markup. Cache_control
+  // passthrough is verified for sonnet but unverified for opus/haiku.
+  claude_sonnet_4_6: [
     {
       provider: 'anthropic',
       providerModel: 'anthrophic_claude_3_sonnet',
@@ -1595,28 +1625,16 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
-  // Explicit current-version aliases so coding-agent callers can pin
-  // to 4.6 / 4.7 without depending on claude_3_sonnet's legacy label.
-  // Same offering chain — the Anthropic entry ships 4.x behind the
-  // wire-dated snapshot, and OpenRouter mirrors the same model.
-  claude_sonnet_4_6: [
+  claude_opus_4_7: [
     {
       provider: 'anthropic',
-      providerModel: 'anthrophic_claude_3_sonnet',
-      latencyTier: 'standard',
+      providerModel: 'anthrophic_claude_3_opus',
+      latencyTier: 'slow',
       available: true,
     },
     {
       provider: 'openrouter',
-      providerModel: 'openrouter_claude_sonnet_4_6',
-      latencyTier: 'standard',
-      available: true,
-    },
-  ],
-  claude_opus_4_7: [
-    {
-      provider: 'anthropic',
-      providerModel: 'anthrophic_claude_3_opus',
+      providerModel: 'openrouter_claude_opus_4_7',
       latencyTier: 'slow',
       available: true,
     },
@@ -1669,17 +1687,29 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
+  // Kimi K2 (original Instruct) — DeepInfra retired Kimi-K2-Instruct
+  // 2026-05-03 in favor of K2.5/K2.6 (different model). Together still
+  // serves the original K2-Instruct as the only path.
   kimi_k2: [
-    // DeepInfra at $0.50/$2.00 vs Together at $1.00/$3.00 — ~2× cheaper.
+    {
+      provider: 'together',
+      providerModel: 'together_kimi_k2',
+      latencyTier: 'standard',
+      available: true,
+    },
+  ],
+  // Kimi K2.6 — newer model. DeepInfra primary (cheaper), OpenRouter
+  // fallback. Both expose the same wire model id family.
+  kimi_k2_6: [
     {
       provider: 'deepinfra',
-      providerModel: 'deepinfra_kimi_k2',
+      providerModel: 'deepinfra_kimi_k2_6',
       latencyTier: 'standard',
       available: true,
     },
     {
-      provider: 'together',
-      providerModel: 'together_kimi_k2',
+      provider: 'openrouter',
+      providerModel: 'openrouter_kimi_k2_6',
       latencyTier: 'standard',
       available: true,
     },
@@ -1730,7 +1760,16 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
+  // MiniMax M2.5 — DeepInfra is cheaper than direct ($0.15/$1.15 vs
+  // $0.30/$1.20), so DeepInfra wins the price-priority sort. Direct
+  // stays as a healthy fallback.
   minimax_m2_5: [
+    {
+      provider: 'deepinfra',
+      providerModel: 'deepinfra_minimax_m2_5',
+      latencyTier: 'standard',
+      available: true,
+    },
     {
       provider: 'minimax',
       providerModel: 'minimax_m2_5',
@@ -1738,6 +1777,8 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
+  // MiniMax M2.7 — not on DeepInfra (only M2.5 is hosted there).
+  // Direct route only.
   minimax_m2_7: [
     {
       provider: 'minimax',
@@ -1746,6 +1787,9 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
+  // GLM 5.1 — OpenRouter is cheaper ($0.95/$3.15 vs DeepInfra
+  // $1.05/$3.50) and has been stable; OpenRouter primary, DeepInfra
+  // fallback when OR is throttled.
   glm_5_1: [
     {
       provider: 'openrouter',
@@ -1753,7 +1797,31 @@ export const textGenMultiProviderModels: Record<
       latencyTier: 'standard',
       available: true,
     },
+    {
+      provider: 'deepinfra',
+      providerModel: 'deepinfra_glm_5_1',
+      latencyTier: 'standard',
+      available: true,
+    },
   ],
+  // GLM 4.6 — DeepInfra primary at $0.43/$1.74; OpenRouter fallback.
+  glm_4_6: [
+    {
+      provider: 'deepinfra',
+      providerModel: 'deepinfra_glm_4_6',
+      latencyTier: 'standard',
+      available: true,
+    },
+    {
+      provider: 'openrouter',
+      providerModel: 'openrouter_glm_4_6',
+      latencyTier: 'standard',
+      available: true,
+    },
+  ],
+  // Qwen 3.6 Plus — OpenRouter only. The "plus" SKU isn't carried by
+  // DeepInfra (DeepInfra exposes 27B and 35B-A3B variants under
+  // separate clean names below).
   qwen3_6_plus: [
     {
       provider: 'openrouter',
@@ -1762,6 +1830,23 @@ export const textGenMultiProviderModels: Record<
       available: true,
     },
   ],
+  // Qwen 3.6 27B — dense (non-MoE). DeepInfra primary at $0.32/$3.20.
+  qwen3_6_27b: [
+    {
+      provider: 'deepinfra',
+      providerModel: 'deepinfra_qwen3_6_27b',
+      latencyTier: 'standard',
+      available: true,
+    },
+    {
+      provider: 'openrouter',
+      providerModel: 'openrouter_qwen3_6_27b',
+      latencyTier: 'standard',
+      available: true,
+    },
+  ],
+  // Kimi K2 Thinking — OpenRouter only. DeepInfra has K2.5/K2.6 but
+  // no specific "thinking" variant.
   kimi_k2_thinking: [
     {
       provider: 'openrouter',
@@ -1775,12 +1860,10 @@ export const textGenMultiProviderModels: Record<
 /** Model data keyed by clean model name. Use for billing/display lookups. */
 export const textGenModelData: Record<TextGenModel, TextGenModelData> = {
   llama_4_scout: TextGenGroqModelData.groq_llama_4_scout,
-  llama_4_maverick: TextGenTogetherModelData.together_meta_llama4_400b,
   llama_3_3_70b: TextGenGroqModelData.groq_llama_3_3_70b,
   deepseek_v4_pro: TextGenDeepSeekModelData.deepseek_deepseek_v4_pro,
   deepseek_v4_flash: TextGenDeepSeekModelData.deepseek_deepseek_v4_flash,
   gpt_oss_120b: TextGenGroqModelData.groq_gpt_oss_120b,
-  gpt_oss_20b: TextGenGroqModelData.groq_gpt_oss_20b,
   gemini_2_5_flash: TextGenGoogleModelData.google_gemini_2_5_flash,
   gemini_2_5: TextGenGoogleModelData.google_gemini_2_5,
   gpt_4o: TextGenOpenAIModelData.openai_gpt_4o,
@@ -1788,20 +1871,21 @@ export const textGenModelData: Record<TextGenModel, TextGenModelData> = {
   gpt_5_mini: TextGenOpenAIModelData.openai_gpt_5_mini,
   o3: TextGenOpenAIModelData.openai_o3,
   o4_mini: TextGenOpenAIModelData.openai_o4_mini,
-  gpt_41_nano: TextGenOpenAIModelData.openai_gpt_41_nano,
-  claude_3_sonnet: TextGenAnthropicModelData.anthrophic_claude_3_sonnet,
   claude_sonnet_4_6: TextGenAnthropicModelData.anthrophic_claude_3_sonnet,
   claude_opus_4_7: TextGenAnthropicModelData.anthrophic_claude_3_opus,
   claude_haiku_4_5: TextGenAnthropicModelData.anthrophic_claude_haiku_4_5,
   qwen2_72b: TextGenTogetherModelData.together_qwen2_72b,
   qwen2_vision_72b: TextGenTogetherModelData.together_qwen2_vision_72b,
   qwen3_235b: TextGenTogetherModelData.together_qwen3_235b,
+  qwen3_6_27b: TextGenDeepInfraModelData.deepinfra_qwen3_6_27b,
   kimi_k2: TextGenTogetherModelData.together_kimi_k2,
+  kimi_k2_6: TextGenDeepInfraModelData.deepinfra_kimi_k2_6,
   gemma_4_31b: TextGenTogetherModelData.together_gemma_4_31b,
   gemma_4_26b: TextGenDeepInfraModelData.deepinfra_gemma_4_26b,
-  minimax_m2_5: TextGenMiniMaxModelData.minimax_m2_5,
+  minimax_m2_5: TextGenDeepInfraModelData.deepinfra_minimax_m2_5,
   minimax_m2_7: TextGenMiniMaxModelData.minimax_m2_7,
   glm_5_1: TextGenOpenRouterModelData.openrouter_glm_5_1,
+  glm_4_6: TextGenDeepInfraModelData.deepinfra_glm_4_6,
   qwen3_6_plus: TextGenOpenRouterModelData.openrouter_qwen_36_plus,
   kimi_k2_thinking: TextGenOpenRouterModelData.openrouter_kimi_k2_thinking,
 };