@blockrun/clawrouter 0.12.257 → 0.12.259

package/dist/cli.js

@@ -35033,6 +35033,13 @@ ${value.slice(-(scanLimit - prefixLength))}`;
         supportsTools: true,
         supportsVision: true
       },
+      "cohere/north-mini-code": {
+        // supportsTools: not probed — fails closed
+        contextWindow: 256e3,
+        maxOutputTokens: 16384,
+        supportsTools: false,
+        supportsVision: false
+      },
       "deepseek/deepseek-chat": {
         contextWindow: 1048576,
         maxOutputTokens: 65536,
@@ -35123,8 +35130,15 @@ ${value.slice(-(scanLimit - prefixLength))}`;
         supportsTools: true,
         supportsVision: true
       },
-      "nvidia/mistral-nemotron": {
-        // supportsTools: gateway unavailable at probe time — fails closed
+      "nvidia/llama-3.2-11b-vision": {
+        // supportsTools: not probed — fails closed
+        contextWindow: 128e3,
+        maxOutputTokens: 16384,
+        supportsTools: false,
+        supportsVision: true
+      },
+      "nvidia/nemotron-3-nano-30b": {
+        // supportsTools: not probed — fails closed
         contextWindow: 131072,
         maxOutputTokens: 16384,
         supportsTools: false,
@@ -35136,21 +35150,16 @@ ${value.slice(-(scanLimit - prefixLength))}`;
         supportsTools: false,
         supportsVision: true
       },
-      "nvidia/nemotron-nano-12b-v2-vl": {
-        // supportsTools: gateway unavailable at probe time — fails closed
-        contextWindow: 131072,
-        maxOutputTokens: 16384,
-        supportsTools: false,
-        supportsVision: true
-      },
-      "nvidia/nemotron-nano-9b-v2": {
-        contextWindow: 131072,
+      "nvidia/nemotron-3-ultra-550b": {
+        // supportsTools: not probed — fails closed
+        contextWindow: 1e6,
         maxOutputTokens: 16384,
         supportsTools: false,
         supportsVision: false
       },
-      "nvidia/step-3.7-flash": {
-        contextWindow: 131072,
+      "nvidia/nemotron-3.5-lightning": {
+        // supportsTools: not probed — fails closed
+        contextWindow: 1e6,
         maxOutputTokens: 16384,
         supportsTools: false,
         supportsVision: false
@@ -35210,13 +35219,6 @@ ${value.slice(-(scanLimit - prefixLength))}`;
         supportsTools: false,
         supportsVision: true
       },
-      "openai/gpt-5.3": {
-        // supportsTools: gateway unavailable at probe time — fails closed
-        contextWindow: 128e3,
-        maxOutputTokens: 128e3,
-        supportsTools: false,
-        supportsVision: true
-      },
       "openai/gpt-5.3-codex": {
         // supportsTools: gateway unavailable at probe time — fails closed; override: 2026-08-29 probe: every request (6 plain + 3 tool attempts) returned a gateway 500, so the probe measured an incident, not the model. Codex's function calling is established by the 2026-07 Terminal-Bench / tau2 calibration trajectories in portfolio.ts. Hosts observing the 500s should drop it with RouterOptions.unavailableModels rather than this snapshot claiming the model cannot call tools.
         contextWindow: 4e5,
@@ -35322,6 +35324,13 @@ ${value.slice(-(scanLimit - prefixLength))}`;
         supportsTools: true,
         supportsVision: false
       },
+      "poolside/laguna-xs-2.1": {
+        // supportsTools: not probed — fails closed
+        contextWindow: 131072,
+        maxOutputTokens: 16384,
+        supportsTools: false,
+        supportsVision: false
+      },
       "qwen/qwen3.7-flash": {
         contextWindow: 1e6,
         maxOutputTokens: 65536,
@@ -37237,8 +37246,8 @@ ${value.slice(-(scanLimit - prefixLength))}`;
             // $0.20/$1.25, 1M ctx
             "google/gemini-2.5-flash-lite",
             // $0.10/$0.40
-            "nvidia/step-3.7-flash"
-            // FREE backstop — NVIDIA free tier (probed 2026-08-21)
+            "nvidia/nemotron-3.5-lightning"
+            // FREE backstop — NVIDIA free tier (probed 2026-08-30)
           ]
         },
         MEDIUM: {
@@ -37327,15 +37336,19 @@ ${value.slice(-(scanLimit - prefixLength))}`;
       // Eco tier configs - absolute cheapest (blockrun/eco)
       ecoTiers: {
         SIMPLE: {
-          primary: "nvidia/step-3.7-flash",
-          // FREE — NVIDIA free tier flagship
+          primary: "nvidia/nemotron-3.5-lightning",
+          // FREE — NVIDIA free tier flagship, 1M ctx
           fallback: [
-            "nvidia/nemotron-nano-9b-v2",
-            // FREE — compact + fast, high-volume light tasks
+            "nvidia/nemotron-3-nano-30b",
+            // FREE — fastest free model (~121 tok/s)
             // The free head keeps rotting with NVIDIA's hosting (deepseek-v4-flash
             // 410 2026-08-12, seed-oss-36b 410 2026-08-03, gpt-oss-120b/20b 400
-            // 2026-08-21). Each retirement retargets the two free rungs to the
-            // current free tier; the paid rungs below never move.
+            // 2026-08-21, and on 2026-08-30 FOUR of the five visible free models at
+            // once — step-3.7-flash, nemotron-nano-9b-v2 and nemotron-nano-12b-v2-vl
+            // all 410, mistral-nemotron hung). Each retirement retargets the two
+            // free rungs to the current free tier; the paid rungs below never move.
+            // The head follows blockrun's own redirect of the model it replaces, so
+            // the router and the gateway never name different models.
             "google/gemini-2.5-flash-lite",
             // $0.10/$0.40 — cheapest paid rung
             "zai/glm-5.3-flash",
@@ -37479,7 +37492,7 @@ ${value.slice(-(scanLimit - prefixLength))}`;
             // strongest open-weight reasoner
             "deepseek/deepseek-chat",
             // Cheap, reliable
-            "nvidia/step-3.7-flash"
+            "nvidia/nemotron-3.5-lightning"
             // NVIDIA free ultimate backstop
           ]
         },
@@ -37583,7 +37596,7 @@ ${value.slice(-(scanLimit - prefixLength))}`;
             // retail high-risk 3/3
             "deepseek/deepseek-chat",
             // cheap, reliable
-            "nvidia/step-3.7-flash"
+            "nvidia/nemotron-3.5-lightning"
             // NVIDIA free ultimate backstop
           ]
         },
@@ -37665,22 +37678,29 @@ var init_top_models = __esm({
       "xai/grok-4.5",
       "xai/grok-4.3",
       "xai/grok-build-0.1",
+      "zai/glm-5.3",
+      "zai/glm-5.3-flash",
       "zai/glm-5.2",
       "zai/glm-5.1",
       "zai/glm-5-turbo",
       "zai/glm-5",
+      "xiaomi/mimo-v2.5",
       "minimax/minimax-m3",
       "minimax/minimax-m2.7",
       "moonshot/kimi-k3",
       "qwen/qwen3.7-max",
+      "qwen/qwen3.8-flash",
+      "deepseek/deepseek-v4-flash-vision-exp",
       "deepseek/deepseek-v4-pro",
       "deepseek/deepseek-chat",
       "deepseek/deepseek-reasoner",
+      "free/nemotron-3.5-lightning",
+      "free/nemotron-3-nano-30b",
+      "free/laguna-xs-2.1",
+      "free/north-mini-code",
       "free/nemotron-3-nano-omni-30b-a3b-reasoning",
-      "free/mistral-nemotron",
-      "free/step-3.7-flash",
-      "free/nemotron-nano-9b-v2",
-      "free/nemotron-nano-12b-v2-vl"
+      "free/nemotron-3-ultra-550b",
+      "free/llama-3.2-11b-vision"
     ];
   }
 });
@@ -37932,6 +37952,19 @@ var init_models = __esm({
       // `kimi` on K2.7. (`grok` WAS promoted to 4.5, but only after the cost
       // tradeoff was argued explicitly; there's no such case for qwen yet.)
       "qwen3.7-max": "qwen/qwen3.7-max",
+      // Qwen3.8 Flash — newer generation than the whole 3.7 line, and cheaper than
+      // the 3.7-plus tier it beats. Bare `qwen` stays UNBOUND (every other qwen*
+      // shorthand resolves to a free model, so binding the short name to a paid
+      // flagship would bill callers expecting free).
+      "qwen3.8-flash": "qwen/qwen3.8-flash",
+      "qwen3-8-flash": "qwen/qwen3.8-flash",
+      "qwen-vision": "qwen/qwen3.8-flash",
+      // DeepSeek's first image-capable SKU. Bare `deepseek` stays on deepseek-chat.
+      "deepseek-vision": "deepseek/deepseek-v4-flash-vision-exp",
+      "v4-flash-vision": "deepseek/deepseek-v4-flash-vision-exp",
+      // Xiaomi MiMo V2.5 — the natively multimodal SKU, distinct from mimo-v2.5-pro
+      // (text-only upstream). `mimo` stays on the Pro entry it has always named.
+      "mimo-vision": "xiaomi/mimo-v2.5",
       "qwen-3.7-max": "qwen/qwen3.7-max",
       "qwen3-7-max": "qwen/qwen3.7-max",
       // Plus/Flash tiers (2026-08-03) — explicit pins, same rule as Max.
@@ -37946,7 +37979,15 @@ var init_models = __esm({
       hunyuan: "tencent/hy3",
       mimo: "xiaomi/mimo-v2.5-pro",
       "mimo-v2.5-pro": "xiaomi/mimo-v2.5-pro",
-      "mimo-v2.5": "xiaomi/mimo-v2.5-pro",
+      // RETARGETED 2026-08-30, Pro -> the real thing. This key used to point at
+      // `xiaomi/mimo-v2.5-pro`, which was harmless while no model owned the name —
+      // but blockrun then listed an actual `xiaomi/mimo-v2.5`, a DIFFERENT and
+      // natively-multimodal SKU at $0.14/$0.28 against Pro's $0.435/$0.87. Leaving
+      // it would have billed 3x for the text-only model when the caller named the
+      // cheaper multimodal one. (The key itself is safe: it is not equal to the
+      // catalog id, so it shadows nothing — the rule that bans `opus-5`-style keys
+      // does not apply to a bare shorthand.)
+      "mimo-v2.5": "xiaomi/mimo-v2.5",
       xiaomi: "xiaomi/mimo-v2.5-pro",
       // Google
       // gemini-3-pro-preview delisted by Google 2026-06-06 — mirror the gateway
@@ -37993,14 +38034,20 @@ var init_models = __esm({
       "xai/grok-3-fast": "xai/grok-4-fast-reasoning",
       // delisted (too expensive)
       // NVIDIA — backward compat aliases (nvidia/xxx → free/xxx)
-      // Default free model is step-3.7-flash — the same model @blockrun/router-core
-      // opens the eco SIMPLE tier on. It replaced gpt-oss-120b on 2026-08-29:
-      // gpt-oss-120b went dead upstream on 2026-08-16 (NVIDIA still LISTS it, but a
-      // completion hangs until the client gives up — blockrun #391 retargeted its
-      // whole free cascade off it, and the gateway now answers 400 Unknown model
-      // for the `free/` id). Pins that NAME gpt-oss stay routable below — the
-      // gateway server-redirects them — but nothing generic may land on it.
-      nvidia: "free/step-3.7-flash",
+      // Default free model is nemotron-3.5-lightning — the same model
+      // @blockrun/router-core opens the eco SIMPLE tier on. It replaced
+      // step-3.7-flash on 2026-08-30, when NVIDIA retired FOUR of the five visible
+      // free models in a single sweep (blockrun #448): step-3.7-flash,
+      // nemotron-nano-9b-v2 and nemotron-nano-12b-v2-vl all published 410 Gone, and
+      // mistral-nemotron went the quiet way — still listed, >150s and zero bytes.
+      // We follow blockrun's own retarget of step-3.7-flash rather than picking a
+      // different survivor, so the proxy and the gateway name the same model.
+      //
+      // gpt-oss-120b/20b RECOVERED on the same probe run and are reachable again,
+      // but they stay withheld from blockrun's public catalog over NVIDIA's
+      // prompt-retention terms — so pins that NAME gpt-oss stay routable below and
+      // nothing generic may land on it.
+      nvidia: "free/nemotron-3.5-lightning",
       "gpt-120b": "free/gpt-oss-120b",
       // names the model itself — gateway redirects
       "gpt-20b": "free/gpt-oss-20b",
@@ -38013,10 +38060,10 @@ var init_models = __esm({
       // straight to gpt-oss-120b. Ids naming v4-flash itself keep the real id (the
       // gateway redirects them); the v3.2/v4-pro ids follow blockrun's retarget
       // rather than chaining through a second dead model.
-      "nvidia/deepseek-v3.2": "free/gpt-oss-120b",
-      "free/deepseek-v3.2": "free/gpt-oss-120b",
-      "nvidia/deepseek-v4-pro": "free/gpt-oss-120b",
-      "free/deepseek-v4-pro": "free/gpt-oss-120b",
+      "nvidia/deepseek-v3.2": "free/nemotron-3.5-lightning",
+      "free/deepseek-v3.2": "free/nemotron-3.5-lightning",
+      "nvidia/deepseek-v4-pro": "free/nemotron-3.5-lightning",
+      "free/deepseek-v4-pro": "free/nemotron-3.5-lightning",
       "nvidia/deepseek-v4-flash": "free/deepseek-v4-flash",
       "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
       // qwen3-coder-480b retired (NVIDIA EOL 2026-06-14). Its server-side redirect used
@@ -38043,16 +38090,16 @@ var init_models = __esm({
       // Retired free IDs → successors (mirror blockrun's 2026-07-17 redirect map:
       // ultra-253b redirects to gpt-oss-120b; the two supers have NO server redirect —
       // they stay hidden-but-routable, so their pins pass through to the real ids).
-      "nvidia/nemotron-ultra-253b": "free/gpt-oss-120b",
+      "nvidia/nemotron-ultra-253b": "free/nemotron-3.5-lightning",
       // mistral-large-3-675b un-retired 2026-06-14, then EOL'd for good 2026-07-28:
       // blockrun's re-probe got HTTP 410 Gone from NVIDIA on both passes (baa967b).
       // Pin stays routable — the gateway redirects it to gpt-oss-120b.
       "nvidia/mistral-large-3-675b": "free/mistral-large-3-675b",
       "nvidia/qwen3.5-122b-a10b": "free/qwen3.5-122b-a10b",
       // devstral-2-123b died upstream; blockrun redirects it to gpt-oss-120b (2026-07-17 map)
-      "nvidia/devstral-2-123b": "free/gpt-oss-120b",
-      "free/nemotron-ultra-253b": "free/gpt-oss-120b",
-      "free/devstral-2-123b": "free/gpt-oss-120b",
+      "nvidia/devstral-2-123b": "free/nemotron-3.5-lightning",
+      "free/nemotron-ultra-253b": "free/nemotron-3.5-lightning",
+      "free/devstral-2-123b": "free/nemotron-3.5-lightning",
       // Explicit-ish pins — dead upstream since 2026-07-28, gateway redirects to gpt-oss-120b
       "mistral-large": "free/mistral-large-3-675b",
       "mistral-large-3-675b": "free/mistral-large-3-675b",
@@ -38063,74 +38110,128 @@ var init_models = __esm({
       // gateway's retarget (blockrun's /free-deepseek page points at gpt-oss-120b
       // too). Shorthands naming v4-flash itself stay on the real id — the gateway
       // redirects them, same treatment as seed-oss/mistral-large pins.
-      "deepseek-free": "free/gpt-oss-120b",
-      "deepseek-v4-pro": "free/gpt-oss-120b",
+      "deepseek-free": "free/nemotron-3.5-lightning",
+      "deepseek-v4-pro": "free/nemotron-3.5-lightning",
       // free shorthand; pro dead upstream (410)
       "deepseek-v4-flash": "free/deepseek-v4-flash",
-      "v4-pro": "free/gpt-oss-120b",
+      "v4-pro": "free/nemotron-3.5-lightning",
       "v4-flash": "free/deepseek-v4-flash",
-      "mistral-free": "free/mistral-nemotron",
-      // mistral-large-3-675b EOL'd 2026-07-28 → last live free Mistral (blockrun demoted its /free-mistral primary the same way)
-      "glm-free": "free/gpt-oss-120b",
+      // mistral-nemotron died 2026-08-30 and it was the LAST free Mistral anywhere —
+      // blockrun's own /free-mistral page now says so plainly instead of naming one.
+      // Point the generic shorthand at the free default rather than advertise a
+      // Mistral we cannot serve.
+      "mistral-free": "free/nemotron-3.5-lightning",
+      "glm-free": "free/nemotron-3.5-lightning",
       // seed-oss-36b (the prior target) EOL'd 2026-08-03
-      "llama-free": "free/gpt-oss-120b",
-      // no live free Llama left (maverick died 2026-07)
+      // A free Llama exists again: nemotron-super-49b (Llama-3.3-based) hit 410 on
+      // 2026-08-30, and a 12-model sweep of what NVIDIA still serves found Llama 3.2
+      // 11B Vision as the only one that finishes a real completion.
+      "llama-free": "free/llama-3.2-11b-vision",
       // qwen3-coder-480b retired 2026-06-14 → seed-oss-36b, which then EOL'd 2026-08-03.
       // Follow the gateway's own retarget rather than chaining to a second dead model.
-      "qwen-coder": "free/step-3.7-flash",
+      "qwen-coder": "free/nemotron-3.5-lightning",
       // no live free Qwen; follows the free default
-      "qwen-coder-free": "free/step-3.7-flash",
-      "qwen-thinking": "free/step-3.7-flash",
+      "qwen-coder-free": "free/nemotron-3.5-lightning",
+      "qwen-thinking": "free/nemotron-3.5-lightning",
       // qwen3-next died 2026-07-17; no live free Qwen left
       "qwen3-next": "free/qwen3-next-80b-a3b-instruct",
       // explicit-ish pin — gateway redirects
       "qwen3-next-80b": "free/qwen3-next-80b-a3b-instruct",
-      "mistral-small": "free/mistral-nemotron",
-      // closest live Mistral-family free model
+      "mistral-small": "free/nemotron-3.5-lightning",
+      // no free Mistral left upstream (2026-08-30)
       // New live free models (2026-06-14 BlockRun free-tier refresh)
       // seed-oss pins name the model itself — kept routable, the gateway redirects them.
       "seed-oss": "free/seed-oss-36b",
       "seed-oss-36b": "free/seed-oss-36b",
-      "coder-free": "free/step-3.7-flash",
-      // generic "a free coder" → follows the free default
+      // A free coder exists again — two of them, both sub-second. north-mini-code is
+      // the faster (607ms median) and 256K ctx against laguna's 131K.
+      "coder-free": "free/north-mini-code",
       "mistral-nemotron": "free/mistral-nemotron",
       "step-flash": "free/step-3.7-flash",
       "step-3.7-flash": "free/step-3.7-flash",
-      "nemotron-nano-9b": "free/nemotron-nano-9b-v2",
-      "nemotron-nano": "free/nemotron-nano-9b-v2",
-      "nemotron-nano-vl": "free/nemotron-nano-12b-v2-vl",
-      "nano-vl": "free/nemotron-nano-12b-v2-vl",
+      // nemotron-nano-9b-v2 hit 410 on 2026-08-30; the generic shorthands follow
+      // blockrun's redirect to its replacement, while the 9b-naming pins above stay
+      // on the real id (the gateway resolves them).
+      "nemotron-nano-9b": "free/nemotron-3-nano-30b",
+      "nemotron-nano": "free/nemotron-3-nano-30b",
+      // nemotron-nano-12b-v2-vl hit 410 the same day; vision in, vision out — the
+      // target is blockrun's own, and nano-omni is the only vision-capable free
+      // model left.
+      "nemotron-nano-vl": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
+      "nano-vl": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
       // Vision-capable free models
       "nemotron-omni": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
       "nano-omni": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
+      // `vision-free` kept for backward compatibility ONLY — it still resolves to
+      // nano-omni, which is the strongest free model, but the free tier no longer
+      // claims working image input on either chain (see the catalog note). Do not
+      // advertise this alias as a way to get free vision.
       "vision-free": "free/nemotron-3-nano-omni-30b-a3b-reasoning",
       // Retired shorthand aliases redirect to live successors. The catch-all target
       // moved llama-4-maverick → gpt-oss-120b (2026-07-17) → step-3.7-flash
-      // (2026-08-29, gpt-oss-120b hung upstream; see the NVIDIA block above).
-      nemotron: "free/nemotron-3-nano-omni-30b-a3b-reasoning",
+      // (2026-08-29) → nemotron-3.5-lightning (2026-08-30, blockrun #448).
+      nemotron: "free/nemotron-3.5-lightning",
       // strongest live Nemotron
-      "nemotron-ultra": "free/step-3.7-flash",
-      "nemotron-253b": "free/step-3.7-flash",
-      "nemotron-super": "free/step-3.7-flash",
-      "nemotron-49b": "free/step-3.7-flash",
-      "nemotron-120b": "free/step-3.7-flash",
-      devstral: "free/step-3.7-flash",
+      "nemotron-ultra": "free/nemotron-3.5-lightning",
+      "nemotron-253b": "free/nemotron-3.5-lightning",
+      "nemotron-super": "free/nemotron-3.5-lightning",
+      "nemotron-49b": "free/nemotron-3.5-lightning",
+      "nemotron-120b": "free/nemotron-3.5-lightning",
+      devstral: "free/nemotron-3.5-lightning",
       // seed-oss-36b EOL'd 2026-08-03
-      "devstral-2": "free/step-3.7-flash",
+      "devstral-2": "free/nemotron-3.5-lightning",
       maverick: "free/llama-4-maverick",
       // explicit-ish pin — gateway redirects
+      // ── The 2026-08-30 free lineup (blockrun #448) ────────────────────────────
+      // nvidia/* bridges for the four NVIDIA-hosted additions.
+      "nvidia/nemotron-3.5-lightning": "free/nemotron-3.5-lightning",
+      "nvidia/nemotron-3-nano-30b": "free/nemotron-3-nano-30b",
+      "nvidia/nemotron-3-ultra-550b": "free/nemotron-3-ultra-550b",
+      "nvidia/llama-3.2-11b-vision": "free/llama-3.2-11b-vision",
+      // The two non-NVIDIA free models keep the `free/` picker convention; their
+      // real upstream ids are accepted as pins and rewritten in toUpstreamModelId
+      // (see FREE_UPSTREAM_OVERRIDES in proxy.ts).
+      "cohere/north-mini-code": "free/north-mini-code",
+      "poolside/laguna-xs-2.1": "free/laguna-xs-2.1",
+      // Shorthands.
+      lightning: "free/nemotron-3.5-lightning",
+      "nemotron-lightning": "free/nemotron-3.5-lightning",
+      "nemotron-3.5-lightning": "free/nemotron-3.5-lightning",
+      "nano-30b": "free/nemotron-3-nano-30b",
+      "nemotron-nano-30b": "free/nemotron-3-nano-30b",
+      "ultra-550b": "free/nemotron-3-ultra-550b",
+      "nemotron-ultra-550b": "free/nemotron-3-ultra-550b",
+      "llama-vision": "free/llama-3.2-11b-vision",
+      "llama-3.2-vision": "free/llama-3.2-11b-vision",
+      "north-mini": "free/north-mini-code",
+      "north-mini-code": "free/north-mini-code",
+      laguna: "free/laguna-xs-2.1",
+      "laguna-xs": "free/laguna-xs-2.1",
       // `free` = the free-tier default. Must equal router-core's ecoTiers.SIMPLE
       // primary and the head of proxy.ts FREE_MODELS so `/model free`, the eco
       // profile and the budget-cap free fallback all agree on one live model.
-      free: "free/step-3.7-flash",
+      free: "free/nemotron-3.5-lightning",
       // MiniMax (minimax → current flagship: M3)
       minimax: "minimax/minimax-m3",
       "minimax-m3": "minimax/minimax-m3",
       "minimax-m2.7": "minimax/minimax-m2.7",
       "minimax-m2.5": "minimax/minimax-m2.5",
       // Z.AI GLM-5
-      glm: "zai/glm-5.2",
-      // bare alias tracks newest flagship (launched 2026-06-16)
+      // Bare `glm` PROMOTED 5.2 → 5.3 (2026-08-30). Same rule as `opus` 4.8 → 5: the
+      // cost tradeoff is zero — identical $1.40/$4.40 AND an identical 1M ctx /
+      // 131072 maxOutput envelope, so nothing can bill or truncate differently for a
+      // caller who never typed a version. blockrun's own copy already names glm-5.3
+      // the Z.AI flagship, so leaving `glm` on 5.2 would make the proxy disagree
+      // with the gateway it fronts. `glm-5.2` and every other version pin still
+      // resolve to their own model.
+      glm: "zai/glm-5.3",
+      "glm-5.3": "zai/glm-5.3",
+      "glm-5-3": "zai/glm-5.3",
+      // GLM-5.3 Flash — Z.AI's first natively multimodal GLM-5. Also router-core's
+      // eco MEDIUM and COMPLEX primary.
+      "glm-5.3-flash": "zai/glm-5.3-flash",
+      "glm-5-3-flash": "zai/glm-5.3-flash",
+      "glm-flash": "zai/glm-5.3-flash",
       "glm-5.2": "zai/glm-5.2",
       "glm-5": "zai/glm-5",
       "glm-5.1": "zai/glm-5.1",
@@ -38192,10 +38293,10 @@ var init_models = __esm({
       },
       {
         id: "free",
-        name: "Free \u2192 Step 3.7 Flash",
+        name: "Free \u2192 Nemotron 3.5 Lightning",
         inputPrice: 0,
         outputPrice: 0,
-        contextWindow: 131072,
+        contextWindow: 1e6,
         maxOutput: 16384,
         reasoning: true
       },
@@ -38236,7 +38337,7 @@ var init_models = __esm({
         inputPrice: 0.25,
         outputPrice: 2,
         contextWindow: 2e5,
-        maxOutput: 65536,
+        maxOutput: 128e3,
         toolCalling: true
       },
       {
@@ -38246,7 +38347,7 @@ var init_models = __esm({
         inputPrice: 0.05,
         outputPrice: 0.4,
         contextWindow: 128e3,
-        maxOutput: 32768,
+        maxOutput: 128e3,
         toolCalling: true,
         deprecated: true,
         fallbackModel: "openai/gpt-5.4-nano"
@@ -38330,8 +38431,8 @@ var init_models = __esm({
         id: "openai/gpt-5.6-terra-pro",
         name: "GPT-5.6 Terra Pro",
         version: "5.6",
-        inputPrice: 1,
-        outputPrice: 6,
+        inputPrice: 2,
+        outputPrice: 12,
         contextWindow: 105e4,
         maxOutput: 128e3,
         reasoning: true,
@@ -38343,8 +38444,8 @@ var init_models = __esm({
         id: "openai/gpt-5.6-luna-pro",
         name: "GPT-5.6 Luna Pro",
         version: "5.6",
-        inputPrice: 0.1,
-        outputPrice: 0.6,
+        inputPrice: 0.2,
+        outputPrice: 1.2,
         contextWindow: 105e4,
         maxOutput: 128e3,
         reasoning: true,
@@ -38407,7 +38508,7 @@ var init_models = __esm({
         version: "5.4",
         inputPrice: 2.5,
         outputPrice: 15,
-        contextWindow: 4e5,
+        contextWindow: 105e4,
         maxOutput: 128e3,
         reasoning: true,
         vision: true,
@@ -38432,7 +38533,7 @@ var init_models = __esm({
         version: "5.4",
         inputPrice: 30,
         outputPrice: 180,
-        contextWindow: 4e5,
+        contextWindow: 105e4,
         maxOutput: 128e3,
         reasoning: true,
         vision: true,
@@ -38445,7 +38546,7 @@ var init_models = __esm({
         inputPrice: 0.2,
         outputPrice: 1.25,
         contextWindow: 105e4,
-        maxOutput: 32768,
+        maxOutput: 128e3,
         toolCalling: true
       },
       // OpenAI GPT-5.3 Family
@@ -38456,7 +38557,7 @@ var init_models = __esm({
         inputPrice: 1.75,
         outputPrice: 14,
         contextWindow: 128e3,
-        maxOutput: 16e3,
+        maxOutput: 128e3,
         reasoning: true,
         vision: true,
         agentic: true,
@@ -38482,7 +38583,7 @@ var init_models = __esm({
         inputPrice: 2,
         outputPrice: 8,
         contextWindow: 128e3,
-        maxOutput: 16384,
+        maxOutput: 32768,
         vision: true,
         toolCalling: true
       },
@@ -38493,7 +38594,7 @@ var init_models = __esm({
         inputPrice: 0.4,
         outputPrice: 1.6,
         contextWindow: 128e3,
-        maxOutput: 16384,
+        maxOutput: 32768,
         toolCalling: true
       },
       {
@@ -38503,7 +38604,7 @@ var init_models = __esm({
         inputPrice: 0.1,
         outputPrice: 0.4,
         contextWindow: 128e3,
-        maxOutput: 16384,
+        maxOutput: 32768,
         toolCalling: true
       },
       {
@@ -38569,7 +38670,7 @@ var init_models = __esm({
         inputPrice: 1.1,
         outputPrice: 4.4,
         contextWindow: 128e3,
-        maxOutput: 65536,
+        maxOutput: 1e5,
         reasoning: true,
         toolCalling: true
       },
@@ -38580,7 +38681,7 @@ var init_models = __esm({
         inputPrice: 1.1,
         outputPrice: 4.4,
         contextWindow: 128e3,
-        maxOutput: 65536,
+        maxOutput: 1e5,
         reasoning: true,
         toolCalling: true
       },
@@ -38593,7 +38694,7 @@ var init_models = __esm({
         inputPrice: 1,
         outputPrice: 5,
         contextWindow: 2e5,
-        maxOutput: 8192,
+        maxOutput: 64e3,
         vision: true,
         agentic: true,
         toolCalling: true
@@ -38617,8 +38718,8 @@ var init_models = __esm({
         version: "4.6",
         inputPrice: 3,
         outputPrice: 15,
-        contextWindow: 2e5,
-        maxOutput: 64e3,
+        contextWindow: 1e6,
+        maxOutput: 128e3,
         reasoning: true,
         vision: true,
         agentic: true,
@@ -38648,7 +38749,7 @@ var init_models = __esm({
         inputPrice: 5,
         outputPrice: 25,
         contextWindow: 2e5,
-        maxOutput: 32e3,
+        maxOutput: 64e3,
         reasoning: true,
         vision: true,
         agentic: true,
@@ -38732,7 +38833,7 @@ var init_models = __esm({
         version: "3.1",
         inputPrice: 2,
         outputPrice: 12,
-        contextWindow: 105e4,
+        contextWindow: 1048576,
         maxOutput: 65536,
         reasoning: true,
         vision: true,
@@ -38744,7 +38845,7 @@ var init_models = __esm({
         version: "3.0",
         inputPrice: 2,
         outputPrice: 12,
-        contextWindow: 105e4,
+        contextWindow: 1048576,
         maxOutput: 65536,
         reasoning: true,
         vision: true,
@@ -38796,7 +38897,7 @@ var init_models = __esm({
         version: "3.0",
         inputPrice: 0.5,
         outputPrice: 3,
-        contextWindow: 1e6,
+        contextWindow: 1048576,
         maxOutput: 65536,
         reasoning: true,
         vision: true
@@ -38807,7 +38908,7 @@ var init_models = __esm({
         version: "2.5",
         inputPrice: 1.25,
         outputPrice: 10,
-        contextWindow: 105e4,
+        contextWindow: 1048576,
         maxOutput: 65536,
         reasoning: true,
         vision: true,
@@ -38819,7 +38920,7 @@ var init_models = __esm({
         version: "2.5",
         inputPrice: 0.3,
         outputPrice: 2.5,
-        contextWindow: 1e6,
+        contextWindow: 1048576,
         maxOutput: 65536,
         vision: true,
         toolCalling: true
@@ -38830,7 +38931,7 @@ var init_models = __esm({
         version: "2.5",
         inputPrice: 0.1,
         outputPrice: 0.4,
-        contextWindow: 1e6,
+        contextWindow: 1048576,
         maxOutput: 65536,
         toolCalling: true
       },
@@ -38840,8 +38941,8 @@ var init_models = __esm({
         version: "3.1",
         inputPrice: 0.25,
         outputPrice: 1.5,
-        contextWindow: 1e6,
-        maxOutput: 8192,
+        contextWindow: 1048576,
+        maxOutput: 65536,
         toolCalling: true
       },
       // DeepSeek — V4 family (2026-04-24). The legacy deepseek-chat/reasoner
@@ -38854,8 +38955,8 @@ var init_models = __esm({
         version: "4-flash",
         inputPrice: 0.14,
         outputPrice: 0.28,
-        contextWindow: 1e6,
-        maxOutput: 8192,
+        contextWindow: 1048576,
+        maxOutput: 65536,
         toolCalling: true
       },
       {
@@ -38864,8 +38965,8 @@ var init_models = __esm({
         version: "4-flash",
         inputPrice: 0.14,
         outputPrice: 0.28,
-        contextWindow: 1e6,
-        maxOutput: 8192,
+        contextWindow: 1048576,
+        maxOutput: 65536,
         reasoning: true,
         toolCalling: true
       },
@@ -38877,8 +38978,8 @@ var init_models = __esm({
         id: "deepseek/deepseek-v4-pro",
         name: "DeepSeek V4 Pro",
         version: "4-pro",
-        inputPrice: 0.435,
-        outputPrice: 0.87,
+        inputPrice: 1.32,
+        outputPrice: 3.96,
         contextWindow: 1048576,
         maxOutput: 65536,
         reasoning: true,
@@ -38991,7 +39092,7 @@ var init_models = __esm({
         inputPrice: 0.6,
         outputPrice: 3,
         contextWindow: 262144,
-        maxOutput: 16384,
+        maxOutput: 65536,
         reasoning: true,
         vision: true,
         agentic: true,
@@ -39004,7 +39105,7 @@ var init_models = __esm({
         inputPrice: 0.6,
         outputPrice: 3,
         contextWindow: 262144,
-        maxOutput: 8192,
+        maxOutput: 16384,
         reasoning: true,
         vision: true,
         agentic: true,
@@ -39042,7 +39143,7 @@ var init_models = __esm({
         version: "4",
         inputPrice: 0.2,
         outputPrice: 0.5,
-        contextWindow: 131072,
+        contextWindow: 2e6,
         maxOutput: 16384,
         reasoning: true,
         toolCalling: true
@@ -39053,7 +39154,7 @@ var init_models = __esm({
         version: "4",
         inputPrice: 0.2,
         outputPrice: 0.5,
-        contextWindow: 131072,
+        contextWindow: 2e6,
         maxOutput: 16384,
         toolCalling: true
       },
@@ -39063,7 +39164,7 @@ var init_models = __esm({
         version: "4.1",
         inputPrice: 0.2,
         outputPrice: 0.5,
-        contextWindow: 131072,
+        contextWindow: 2e6,
         maxOutput: 16384,
         reasoning: true,
         toolCalling: true
@@ -39074,7 +39175,7 @@ var init_models = __esm({
         version: "4.1",
         inputPrice: 0.2,
         outputPrice: 0.5,
-        contextWindow: 131072,
+        contextWindow: 2e6,
         maxOutput: 16384,
         toolCalling: true
       },
@@ -39086,7 +39187,7 @@ var init_models = __esm({
         version: "4-0709",
         inputPrice: 3,
         outputPrice: 15,
-        contextWindow: 131072,
+        contextWindow: 256e3,
         maxOutput: 16384,
         reasoning: true,
         toolCalling: true
@@ -39097,7 +39198,7 @@ var init_models = __esm({
         version: "2",
         inputPrice: 2,
         outputPrice: 10,
-        contextWindow: 131072,
+        contextWindow: 32768,
         maxOutput: 16384,
         vision: true,
         toolCalling: true
@@ -39336,8 +39437,21 @@ var init_models = __esm({
         outputPrice: 0,
         contextWindow: 256e3,
         maxOutput: 16384,
-        reasoning: true,
-        vision: true
+        reasoning: true
+        // NO `vision: true` — 2026-08-31. Both catalogs advertise vision on this
+        // model and blockrun #448 cites an 8x8 PNG answering "Red", but a 64x64
+        // solid-red probe does not hold up: 1 of 4 correct on Base (the others
+        // "I'm not able to view the image" or leaked reasoning), and on sol it
+        // answered "white" twice, once with the response's own `model` field
+        // reading `nemotron-3-super-120b (fallback: ...nano-omni)` — the image is
+        // silently dropped and a text model answers.
+        //
+        // `vision: true` is what makes filterByVision() route real image turns
+        // here, so the flag does not merely describe the model, it aims traffic at
+        // it. HTTP 200 with a confident wrong colour is worse than no free vision:
+        // there is no error for a caller to branch on. Image turns go to paid
+        // vision models until a correctly-sized probe comes back right on both
+        // chains. Independently confirmed on Solana by the blockrun-sol owner.
       },
       // 2026-06-14: BlockRun re-featured these two as free flagships (catalog sweep).
       // Added to the auto-pick set behind gpt-oss to strengthen the mid/back of the
@@ -39402,7 +39516,13 @@ var init_models = __esm({
         maxOutput: 16384
       },
       {
-        // Mistral × NVIDIA hybrid: strong instruction following, 131K context.
+        // Mistral × NVIDIA hybrid, 131K context. DEAD 2026-08-30 — the QUIET kind:
+        // NVIDIA still LISTS it but a completion returns zero bytes after >150s on
+        // both of blockrun's probe passes (blockrun #448, the #391 shape). Upstream
+        // hid it and redirects callers to a live workhorse. Entry kept so explicit
+        // pins stay routable; off the picker and the FREE_MODELS cascade.
+        // There is no free Mistral left on NVIDIA — do not point a generic
+        // "a free mistral" shorthand at one.
         id: "free/mistral-nemotron",
         name: "[Free] Mistral Nemotron",
         version: "nemotron",
@@ -39412,7 +39532,13 @@ var init_models = __esm({
         maxOutput: 16384
       },
       {
-        // StepFun Step 3.7 Flash: reasoning-focused, 131K context.
+        // StepFun Step 3.7 Flash: reasoning-focused, 131K context. EOL'd 2026-08-30
+        // — NVIDIA published 410 Gone on both of blockrun's probe passes (#448), in
+        // the same sweep that took nemotron-nano-9b-v2, nemotron-nano-12b-v2-vl and
+        // (hidden) nemotron-super-49b. It had been ClawRouter's free default since
+        // 2026-08-29; the gateway redirects it to nemotron-3.5-lightning, which is
+        // exactly why nothing looked broken. Entry kept so explicit pins stay
+        // routable; off the picker and the FREE_MODELS cascade.
         id: "free/step-3.7-flash",
         name: "[Free] StepFun Step 3.7 Flash",
         version: "3.7-flash",
@@ -39424,6 +39550,8 @@ var init_models = __esm({
       },
       {
         // NVIDIA Nemotron Nano 9B v2: fast lightweight generalist, 131K context.
+        // EOL'd 2026-08-30 (410 Gone, same sweep). Gateway redirects it to
+        // nemotron-3-nano-30b. Entry kept for pins; off the picker and the cascade.
         id: "free/nemotron-nano-9b-v2",
         name: "[Free] Nemotron Nano 9B v2",
         version: "nano-9b-v2",
@@ -39435,6 +39563,9 @@ var init_models = __esm({
       },
       {
         // NVIDIA Nemotron Nano 12B v2 VL: vision-language (text + image), 131K context.
+        // EOL'd 2026-08-30 (410 Gone, same sweep). Gateway redirects it to
+        // nemotron-3-nano-omni, the only vision-capable free model left. Entry kept
+        // for pins; off the picker and the cascade.
         id: "free/nemotron-nano-12b-v2-vl",
         name: "[Free] Nemotron Nano 12B v2 VL",
         version: "nano-12b-v2-vl",
@@ -39442,12 +39573,217 @@ var init_models = __esm({
         outputPrice: 0,
         contextWindow: 131072,
         maxOutput: 16384,
+        reasoning: true
+        // Vision flag dropped 2026-08-31 with the rest of the free tier: this id is
+        // 410 Gone and the gateway redirects it to nano-omni, whose image path does
+        // not work either (see above). A pinned caller sending an image would have
+        // been told, by the flag, that it would be seen.
+      },
+      // ── 2026-08-30 free-tier rebuild (blockrun #448) ─────────────────────────
+      // NVIDIA retired FOUR of the five VISIBLE free models in one sweep. blockrun
+      // replaced them and moved two survivors onto OpenRouter's $0 ":free" pool,
+      // taking the visible free set from 5 to 7 and spanning three hosts for the
+      // first time (NVIDIA, OpenRouter, and two non-NVIDIA makers).
+      //
+      // All seven live free models were tool-probed through the gateway on
+      // 2026-08-30 and every one returned a structured `tool_calls` array with
+      // valid JSON args and finish_reason: "tool_calls" (nemotron-3-nano-30b looked
+      // like a textual leak at max_tokens:120 — that was truncation; at 300 it is
+      // structured). The evidence is recorded here on purpose, but `toolCalling` is
+      // deliberately NOT set: no free entry has ever carried it, filterByToolCalling
+      // keeps the free tier out of tool-bearing requests, and the budget pre-check
+      // at proxy.ts:5217 is written on the same assumption. Flipping the tier into
+      // agentic eligibility is its own change with its own evidence bar.
+      {
+        // Nemotron 3.5 Lightning 30B-A3B — the new free default. blockrun's own
+        // redirect target for the retired step-3.7-flash, so following it keeps the
+        // proxy and the gateway naming the same model. Served from OpenRouter's $0
+        // pool (4.9s median vs 16.3s direct) which also publishes a 1M window
+        // against the 131K the NVIDIA NIM path gives; blockrun carries a hidden
+        // "-nim" twin as its fallback. Probed 1.3s with tools.
+        id: "free/nemotron-3.5-lightning",
+        name: "[Free] Nemotron 3.5 Lightning",
+        version: "3.5-lightning",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 1e6,
+        maxOutput: 16384,
+        reasoning: true
+      },
+      {
+        // Nemotron 3 Nano 30B-A3B — the fastest free model in the catalog
+        // (~121 tok/s on a realistic workload, not a 16-token ping). Returns
+        // reasoning_content. Also the tertiary rung of blockrun's own free cascade.
+        id: "free/nemotron-3-nano-30b",
+        name: "[Free] Nemotron 3 Nano 30B",
+        version: "3-nano-30b",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 131072,
+        maxOutput: 16384,
+        reasoning: true
+      },
+      {
+        // Nemotron 3 Ultra 550B-A55B — the largest free model ever listed, 1M ctx,
+        // reachable ONLY through OpenRouter's $0 pool (build.nvidia.com answers 503).
+        // Deliberately LOW in the cascade: 16.8s on the tools probe, and blockrun
+        // measured 3 of 15 calls coming back as an HTTP 200 carrying an upstream
+        // 502/503 error object instead of choices.
+        id: "free/nemotron-3-ultra-550b",
+        name: "[Free] Nemotron 3 Ultra 550B",
+        version: "3-ultra-550b",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 1e6,
+        maxOutput: 16384,
+        reasoning: true
+      },
+      {
+        // Meta Llama 3.2 11B Vision — restores a free Llama after nemotron-super-49b
+        // (Llama-3.3-based) hit NVIDIA's 410 EOL. A 12-model sweep found exactly two
+        // survivors and only the 11B finishes a real completion. Older than the rest
+        // of the tier and deliberately so: it is a real Llama that actually answers.
+        // Slowest in the tier (~18 tok/s) — last rung of the cascade.
+        id: "free/llama-3.2-11b-vision",
+        name: "[Free] Llama 3.2 11B Vision",
+        version: "3.2-11b-vision",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 128e3,
+        maxOutput: 16384
+        // NO `vision: true` despite the name and blockrun's `categories:
+        // ["chat","vision"]` — 2026-08-31, three consecutive 64x64 PNG probes came
+        // back "I'm unable to see the image" / "you haven't provided an image",
+        // while a plain-text control on the same id answered fine. The model is
+        // alive; the image path is not. Mirroring the catalog's claim here would
+        // have routed image turns to it. See the nano-omni note above.
+      },
+      {
+        // Cohere North Mini Code, on OpenRouter's $0 pool — 607ms median, the
+        // fastest thing in the tier. Emits reasoning_content (content is clean once
+        // the budget is large enough to finish, verified live).
+        // NOTE: NOT an nvidia/* id upstream — see FREE_UPSTREAM_OVERRIDES in proxy.ts.
+        id: "free/north-mini-code",
+        name: "[Free] Cohere North Mini Code",
+        version: "north-mini-code",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 256e3,
+        maxOutput: 16384,
+        reasoning: true
+      },
+      {
+        // Poolside Laguna XS 2.1 on OUR NVIDIA key (~161 tok/s). Deliberately not
+        // the OpenRouter twin, which 429s on every attempt — so this rung and
+        // north-mini-code sit on DIFFERENT capacity pools, which is why they are
+        // adjacent in the cascade.
+        // NOTE: NOT an nvidia/* id upstream — see FREE_UPSTREAM_OVERRIDES in proxy.ts.
+        id: "free/laguna-xs-2.1",
+        name: "[Free] Poolside Laguna XS 2.1",
+        version: "xs-2.1",
+        inputPrice: 0,
+        outputPrice: 0,
+        contextWindow: 131072,
+        maxOutput: 16384
+      },
+      // ── 2026-08-30 paid catalog refresh (blockrun #449) ─────────────────────
+      // Three additions, each probe-verified upstream with a real completion AND a
+      // real image before listing. All three carry `vision: true` on that evidence
+      // — unlike the free tier, where the same claim did not survive a probe.
+      {
+        // Alibaba's 3.8 generation: 125B MoE, one tier above the whole 3.7 line and
+        // cheaper than the qwen3.7-plus ($0.32/$1.28) it outperforms.
+        // Vision took two probes to establish upstream: the first 400'd with
+        // `invalid_parameter_error` on an 8x8 PNG because the model requires >10px
+        // per side; 64x64 answered correctly. A single 400 is not a capability gap.
+        id: "qwen/qwen3.8-flash",
+        name: "Qwen3.8 Flash",
+        version: "3.8-flash",
+        inputPrice: 0.15,
+        outputPrice: 0.47,
+        contextWindow: 1e6,
+        maxOutput: 131072,
         reasoning: true,
-        vision: true
+        vision: true,
+        toolCalling: true
+      },
+      {
+        // The first DeepSeek SKU that takes images. Priced at DeepSeek's PEAK rate
+        // on purpose: they now split peak/off-peak and off-peak is half, so listing
+        // the lower number would sell under cost for seven hours every weekday.
+        id: "deepseek/deepseek-v4-flash-vision-exp",
+        name: "DeepSeek V4 Flash Vision",
+        version: "v4-flash-vision-exp",
+        inputPrice: 0.44,
+        outputPrice: 1.32,
+        contextWindow: 1048576,
+        maxOutput: 65536,
+        reasoning: true,
+        vision: true,
+        toolCalling: true
+      },
+      {
+        // NOT a cheaper mimo-v2.5-pro — a different, natively multimodal SKU. The
+        // Pro entry is text-only upstream while this one takes images, at a third
+        // of the price. Keep both.
+        id: "xiaomi/mimo-v2.5",
+        name: "Xiaomi MiMo V2.5",
+        version: "2.5",
+        inputPrice: 0.14,
+        outputPrice: 0.28,
+        contextWindow: 1048576,
+        maxOutput: 131072,
+        reasoning: true,
+        vision: true,
+        toolCalling: true
       },
       // Z.AI GLM-5 Models
       {
-        // Launched 2026-06-16 — Z.AI's newest flagship. 1M-token context,
+        // Z.AI's flagship, live-probed by blockrun 2026-08-19 against api.z.ai
+        // (real completion in 2s, content alongside reasoning tokens). $1.40/$4.40,
+        // cached input $0.26, off the international USD price list.
+        // maxOutput is 131072, NOT the 1M context: that is the hard ceiling the API
+        // enforces (error 1210 above it), the same on every GLM-5 SKU.
+        // Thinking is ALWAYS ON here and cannot be disabled.
+        id: "zai/glm-5.3",
+        name: "GLM-5.3",
+        version: "5.3",
+        inputPrice: 1.4,
+        outputPrice: 4.4,
+        contextWindow: 1e6,
+        maxOutput: 131072,
+        reasoning: true,
+        toolCalling: true
+      },
+      {
+        // Z.AI's first natively multimodal GLM-5 — 320B/18B MoE. blockrun probed
+        // text, vision (base64 data URL → correct answer), tools
+        // (finish_reason=tool_calls with well-formed arguments) and streaming live
+        // on 2026-08-27 before listing it.
+        //
+        // MUST be in this catalog, not just router-core's: it is router-core's eco
+        // MEDIUM and COMPLEX primary, and estimateAmount() returns undefined for an
+        // id we do not carry — which makes the request skip the maxCostPerRun filter
+        // AND never accumulate into session cost. An uncatalogued routing target is
+        // a cost-cap hole, not a logging gap.
+        //
+        // Price is the LIST rate. Z.AI is running a 50% launch promo ($0.075/$0.25)
+        // that ENDS 2026-09-09; listing the promo rate would put us under COGS the
+        // morning it lapses. Thinking is always on, as on glm-5.3.
+        id: "zai/glm-5.3-flash",
+        name: "GLM-5.3 Flash",
+        version: "5.3-flash",
+        inputPrice: 0.15,
+        outputPrice: 0.5,
+        contextWindow: 1e6,
+        maxOutput: 131072,
+        reasoning: true,
+        vision: true,
+        toolCalling: true
+      },
+      {
+        // Launched 2026-06-16. Was Z.AI's flagship until glm-5.3 (above) took the
+        // slot on 2026-08-19. 1M-token context,
         // beats GPT-5.5 on long-horizon coding at a fraction of the cost.
         // Paid per-token at $1.40/$4.40 (same as glm-5.1, cached $0.26).
         id: "zai/glm-5.2",
@@ -39456,7 +39792,7 @@ var init_models = __esm({
         inputPrice: 1.4,
         outputPrice: 4.4,
         contextWindow: 1e6,
-        maxOutput: 262144,
+        maxOutput: 131072,
         reasoning: true,
         toolCalling: true
       },
@@ -89687,13 +90023,44 @@ import { homedir as homedir7 } from "os";
 import { join as join10 } from "path";
 import { mkdir as mkdir4, writeFile as writeFile2, readFile as readFile2, stat as fsStat } from "fs/promises";
 import { readFileSync as readFileSync2, existsSync as existsSync2 } from "fs";
+async function loadGatewayCatalog(apiBase) {
+  try {
+    const controller = new AbortController();
+    const timer2 = setTimeout(() => controller.abort(), GATEWAY_CATALOG_TIMEOUT_MS);
+    const res = await fetch(`${apiBase}/v1/models`, { signal: controller.signal });
+    clearTimeout(timer2);
+    if (!res.ok) return;
+    const body = await res.json();
+    const ids = (body.data ?? []).map((m) => m.id).filter((id2) => !!id2);
+    if (ids.length === 0) return;
+    gatewayModelIds = new Set(ids);
+    const unserved = [...FREE_MODELS].filter((m) => !gatewayModelIds.has(toUpstreamModelId(m)));
+    if (unserved.length > 0) {
+      console.log(
+        `[ClawRouter] Free models not served by this gateway (skipped in the cascade): ${unserved.join(", ")}`
+      );
+    }
+  } catch {
+  }
+}
+function isServedByGateway(modelId) {
+  if (!gatewayModelIds) return true;
+  return gatewayModelIds.has(toUpstreamModelId(modelId));
+}
 function pickFreeModel(excludeList) {
+  for (const m of FREE_MODELS) {
+    if (excludeList?.has(m)) continue;
+    if (!isServedByGateway(m)) continue;
+    return m;
+  }
   for (const m of FREE_MODELS) {
     if (!excludeList?.has(m)) return m;
   }
   return void 0;
 }
 function toUpstreamModelId(modelId) {
+  const override = FREE_UPSTREAM_OVERRIDES[modelId];
+  if (override) return override;
   if (modelId.startsWith("free/")) {
     return "nvidia/" + modelId.slice("free/".length);
   }
@@ -90025,6 +90392,9 @@ function detectDegradedSuccessResponse(body) {
       return `degraded response: ${errorText.slice(0, 120)}`;
     }
     const choices = parsed.choices;
+    if (Array.isArray(choices) && choices.length === 0) {
+      return "degraded response: no choices returned";
+    }
     if (Array.isArray(choices) && choices.length > 0) {
       const choice = choices[0];
       const msg = choice.message ?? choice.delta;
@@ -90515,6 +90885,7 @@ async function startProxy(options) {
   } else if (paymentChain === "solana") {
     console.log(`[ClawRouter] Payment chain: Solana (${BLOCKRUN_SOLANA_API})`);
   }
+  void loadGatewayCatalog(apiBase);
   const listenPort = options.port ?? getProxyPort();
   const existingProxy = await checkExistingProxy(listenPort);
   if (existingProxy) {
@@ -93048,6 +93419,22 @@ data: [DONE]
     } else {
       modelsToTry = modelId ? [modelId] : [];
     }
+    if (modelsToTry.length > 0) {
+      const unservedFree = modelsToTry.filter((m) => FREE_MODELS.has(m) && !isServedByGateway(m));
+      if (unservedFree.length > 0) {
+        const servedFree = [...FREE_MODELS].filter(
+          (m) => isServedByGateway(m) && !excludeList?.has(m)
+        );
+        const replacement = servedFree[0];
+        if (replacement) {
+          modelsToTry = modelsToTry.map((m) => unservedFree.includes(m) ? replacement : m);
+          modelsToTry = modelsToTry.filter((m, i) => modelsToTry.indexOf(m) === i);
+          console.log(
+            `[ClawRouter] ${unservedFree.join(", ")} not served on this chain \u2014 using ${replacement}`
+          );
+        }
+      }
+    }
     if (!hasTools && routingDecision) {
       const freeFallback = pickFreeModel(excludeList);
       if (freeFallback && !modelsToTry.includes(freeFallback)) {
@@ -93928,7 +94315,7 @@ data: [DONE]
     });
   }
 }
-var paymentStore, BLOCKRUN_API, BLOCKRUN_SOLANA_API, IMAGE_DIR, AUDIO_DIR, VIDEO_DIR, AUTO_MODEL, ROUTING_PROFILES, FREE_MODELS, FREE_MODEL, MAX_MESSAGES, CONTEXT_LIMIT_KB, HEARTBEAT_INTERVAL_MS, BALANCE_CHECK_TIMEOUT_MS, DEFAULT_REQUEST_TIMEOUT_MS, PER_MODEL_TIMEOUT_MS, REASONING_MODEL_TIMEOUT_MS, REASONING_MODEL_IDS, MAX_FALLBACK_ATTEMPTS, HEALTH_CHECK_TIMEOUT_MS, RATE_LIMIT_COOLDOWN_MS, OVERLOAD_COOLDOWN_MS, PORT_RETRY_ATTEMPTS, PORT_RETRY_DELAY_MS, MODEL_BODY_READ_TIMEOUT_MS, ERROR_BODY_READ_TIMEOUT_MS, ClientDisconnectedError, rateLimitedModels, overloadedModels, perProviderErrors, BALANCE_CHECK_BUFFER, BALANCE_PREFLIGHT_OUTPUT_TOKEN_CAP, PROVIDER_ERROR_PATTERNS, DEGRADED_RESPONSE_PATTERNS, DEGRADED_LOOP_PATTERNS, VALID_ROLES, ROLE_MAPPINGS, VALID_TOOL_ID_PATTERN, KIMI_BLOCK_RE, KIMI_TOKEN_RE, THINKING_TAG_RE, THINKING_BLOCK_RE, BLOCKRUN_MODEL_BY_ID, IMAGE_PRICING, IMAGE_MODEL_IDS, IMAGE_MODEL_SIZES, IMAGE_MODEL_ALIASES, VIDEO_PRICING, PHONE_PRICING;
+var paymentStore, BLOCKRUN_API, BLOCKRUN_SOLANA_API, IMAGE_DIR, AUDIO_DIR, VIDEO_DIR, AUTO_MODEL, ROUTING_PROFILES, FREE_MODELS, gatewayModelIds, FREE_MODEL, FREE_UPSTREAM_OVERRIDES, MAX_MESSAGES, CONTEXT_LIMIT_KB, HEARTBEAT_INTERVAL_MS, BALANCE_CHECK_TIMEOUT_MS, GATEWAY_CATALOG_TIMEOUT_MS, DEFAULT_REQUEST_TIMEOUT_MS, PER_MODEL_TIMEOUT_MS, REASONING_MODEL_TIMEOUT_MS, REASONING_MODEL_IDS, MAX_FALLBACK_ATTEMPTS, HEALTH_CHECK_TIMEOUT_MS, RATE_LIMIT_COOLDOWN_MS, OVERLOAD_COOLDOWN_MS, PORT_RETRY_ATTEMPTS, PORT_RETRY_DELAY_MS, MODEL_BODY_READ_TIMEOUT_MS, ERROR_BODY_READ_TIMEOUT_MS, ClientDisconnectedError, rateLimitedModels, overloadedModels, perProviderErrors, BALANCE_CHECK_BUFFER, BALANCE_PREFLIGHT_OUTPUT_TOKEN_CAP, PROVIDER_ERROR_PATTERNS, DEGRADED_RESPONSE_PATTERNS, DEGRADED_LOOP_PATTERNS, VALID_ROLES, ROLE_MAPPINGS, VALID_TOOL_ID_PATTERN, KIMI_BLOCK_RE, KIMI_TOKEN_RE, THINKING_TAG_RE, THINKING_BLOCK_RE, BLOCKRUN_MODEL_BY_ID, IMAGE_PRICING, IMAGE_MODEL_IDS, IMAGE_MODEL_SIZES, IMAGE_MODEL_ALIASES, VIDEO_PRICING, PHONE_PRICING;
 var init_proxy = __esm({
   "src/proxy.ts"() {
     "use strict";
@@ -93979,22 +94366,31 @@ var init_proxy = __esm({
       "premium"
     ]);
     FREE_MODELS = /* @__PURE__ */ new Set([
-      "free/step-3.7-flash",
-      // reasoning-focused — free-tier flagship
-      "free/nemotron-nano-9b-v2",
-      // fast lightweight generalist (~0.7s)
-      "free/mistral-nemotron",
-      // strong instruction following
+      "free/nemotron-3.5-lightning",
+      // free-tier default — 1M ctx, thinking mode
+      "free/nemotron-3-nano-30b",
+      // fastest free model (~121 tok/s)
+      "free/laguna-xs-2.1",
+      // coding, ~161 tok/s — on our NVIDIA key
+      "free/north-mini-code",
+      // coding, 607ms median — OpenRouter $0 pool
       "free/nemotron-3-nano-omni-30b-a3b-reasoning",
       // vision (text/image/video/audio)
-      "free/nemotron-nano-12b-v2-vl"
-      // vision-language (text + image)
+      "free/nemotron-3-ultra-550b",
+      // largest free model, 1M ctx — but slow + flaky
+      "free/llama-3.2-11b-vision"
+      // only free Llama left; slowest, so last
     ]);
-    FREE_MODEL = "free/step-3.7-flash";
+    FREE_MODEL = "free/nemotron-3.5-lightning";
+    FREE_UPSTREAM_OVERRIDES = {
+      "free/north-mini-code": "cohere/north-mini-code",
+      "free/laguna-xs-2.1": "poolside/laguna-xs-2.1"
+    };
     MAX_MESSAGES = 200;
     CONTEXT_LIMIT_KB = 5120;
     HEARTBEAT_INTERVAL_MS = 2e3;
     BALANCE_CHECK_TIMEOUT_MS = 2500;
+    GATEWAY_CATALOG_TIMEOUT_MS = 5e3;
     DEFAULT_REQUEST_TIMEOUT_MS = 3e5;
     PER_MODEL_TIMEOUT_MS = 6e4;
     REASONING_MODEL_TIMEOUT_MS = 18e4;