@blockrun/franklin 3.49.12 → 3.49.13

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -71,6 +71,7 @@ const MODEL_MAX_OUTPUT = {
71
71
  'deepseek/deepseek-chat': 65_536,
72
72
  'deepseek/deepseek-reasoner': 65_536,
73
73
  'deepseek/deepseek-v4-pro': 65_536,
74
+ 'deepseek/deepseek-v4.1-flash': 65_536,
74
75
  // Kimi K3 (flagship, 2026-07) supports 65K output per the BlockRun gateway
75
76
  // model entry (max_output: 65536). Without this entry the default 16K cap
76
77
  // left users with 4× headroom on the table for long-form coding outputs
@@ -264,6 +264,7 @@ const MODEL_CONTEXT_WINDOWS = {
264
264
  'deepseek/deepseek-chat': 1_000_000,
265
265
  'deepseek/deepseek-reasoner': 1_000_000,
266
266
  'deepseek/deepseek-v4-pro': 1_000_000,
267
+ 'deepseek/deepseek-v4.1-flash': 1_000_000,
267
268
  // xAI. grok-4.5 / 4.3 / build were missing until 2026-08-20: neither matches
268
269
  // any inference pattern below, so a cold catalog cache fell through to the
269
270
  // blind 128k default and compacted a 500K–1M window ~4-8x too early.
@@ -24,7 +24,9 @@
24
24
  export const FREE_DEFAULT_MODEL = 'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning';
25
25
  /** Free models in preference order, best first. See free-models.ts. */
26
26
  export const FREE_PREFERENCE_ORDER = [
27
- 'nvidia/nemotron-3-nano-30b', // 5/5 healthy, fastest
27
+ // nvidia/nemotron-3-nano-30b led this list until the gateway delisted it on
28
+ // 2026-09-08 (NVIDIA deprovisioned it). It is gone from /api/v1/models, so
29
+ // the chain-safe default leads again.
28
30
  FREE_DEFAULT_MODEL, // the floor this never drops below
29
31
  'poolside/laguna-xs-2.1', // different provider
30
32
  'cohere/north-mini-code', // different provider; needs max_tokens >= 1200
@@ -103,7 +103,15 @@ export { FREE_DEFAULT_MODEL, FREE_POOL_CONTEXT_FLOOR, FREE_PREFERENCE_ORDER };
103
103
  * chain alive through an NVIDIA-wide outage instead of it being four names
104
104
  * for one upstream.
105
105
  *
106
+ * 2026-10-20: nemotron-3-nano-30b, the Base leader in the table above, was
107
+ * delisted on 2026-09-08 and has left this list (see LEGACY_FREE_MODEL_IDS).
108
+ * The measurements stay as the record of why it led.
109
+ *
106
110
  * Not listed, and why:
111
+ * - nvidia/muse-glimmer-30b — joined the free pool 2026-10-02 and has not
112
+ * been measured on /v1/messages, the bar every rung above had to clear.
113
+ * A reasoning model that spends a small max_tokens budget entirely on
114
+ * reasoning (gateway probe). Reachable by explicit pin until measured.
107
115
  * - nvidia/nemotron-3.5-lightning — LATENCY, and only latency. Measured
108
116
  * 10.9s / 12.2s / 22.5s for a one-sentence answer at max_tokens 800,
109
117
  * against ~0.8-3s for the rest of the pool. Its name is a lie. (An
@@ -178,16 +186,18 @@ export { FREE_DEFAULT_MODEL, FREE_POOL_CONTEXT_FLOOR, FREE_PREFERENCE_ORDER };
178
186
  */
179
187
  export declare const QUARANTINED_FREE_MODELS: readonly string[];
180
188
  /**
181
- * Free ids Franklin used to name. The gateway still answers for all of them
182
- * (via substitution), but they are gone from /api/v1/models. Kept so old
183
- * session-cost records still price at $0 and so stale aliases can be remapped
184
- * rather than handed back to the user as-is.
189
+ * Free ids Franklin used to name. All are gone from /api/v1/models; the
190
+ * gateway answers for most of them by substitution. Kept DELIBERATELY: this
191
+ * list is what lets isFreeModelId() report an old session's model as $0 and
192
+ * keep a stale pin on the free side of every paid-approval gate. It is never
193
+ * routed to — FREE_PREFERENCE_ORDER is the routing list.
185
194
  */
186
195
  export declare const LEGACY_FREE_MODEL_IDS: readonly string[];
187
196
  /**
188
- * Every free id in the live Base catalog as of 2026-08-30, mapped to its
189
- * catalogued context window. Used for metadata display; NOT for sizing a
190
- * request — use FREE_POOL_CONTEXT_FLOOR for that.
197
+ * Every free id (billing_mode "free") in the live catalog as of 2026-10-20 —
198
+ * identical on Base and Solana — mapped to its catalogued context window.
199
+ * Used for metadata display and as the cold-cache answer to isFreeModelId();
200
+ * NOT for sizing a request — use FREE_POOL_CONTEXT_FLOOR for that.
191
201
  */
192
202
  export declare const FREE_MODEL_CONTEXT_WINDOWS: Readonly<Record<string, number>>;
193
203
  /**
@@ -109,7 +109,15 @@ export { FREE_DEFAULT_MODEL, FREE_POOL_CONTEXT_FLOOR, FREE_PREFERENCE_ORDER };
109
109
  * chain alive through an NVIDIA-wide outage instead of it being four names
110
110
  * for one upstream.
111
111
  *
112
+ * 2026-10-20: nemotron-3-nano-30b, the Base leader in the table above, was
113
+ * delisted on 2026-09-08 and has left this list (see LEGACY_FREE_MODEL_IDS).
114
+ * The measurements stay as the record of why it led.
115
+ *
112
116
  * Not listed, and why:
117
+ * - nvidia/muse-glimmer-30b — joined the free pool 2026-10-02 and has not
118
+ * been measured on /v1/messages, the bar every rung above had to clear.
119
+ * A reasoning model that spends a small max_tokens budget entirely on
120
+ * reasoning (gateway probe). Reachable by explicit pin until measured.
113
121
  * - nvidia/nemotron-3.5-lightning — LATENCY, and only latency. Measured
114
122
  * 10.9s / 12.2s / 22.5s for a one-sentence answer at max_tokens 800,
115
123
  * against ~0.8-3s for the rest of the pool. Its name is a lie. (An
@@ -202,12 +210,16 @@ export const QUARANTINED_FREE_MODELS = [
202
210
  // it is a cascade rung other BlockRun components depend on.
203
211
  ];
204
212
  /**
205
- * Free ids Franklin used to name. The gateway still answers for all of them
206
- * (via substitution), but they are gone from /api/v1/models. Kept so old
207
- * session-cost records still price at $0 and so stale aliases can be remapped
208
- * rather than handed back to the user as-is.
213
+ * Free ids Franklin used to name. All are gone from /api/v1/models; the
214
+ * gateway answers for most of them by substitution. Kept DELIBERATELY: this
215
+ * list is what lets isFreeModelId() report an old session's model as $0 and
216
+ * keep a stale pin on the free side of every paid-approval gate. It is never
217
+ * routed to — FREE_PREFERENCE_ORDER is the routing list.
209
218
  */
210
219
  export const LEGACY_FREE_MODEL_IDS = [
220
+ // Delisted 2026-09-08 (NVIDIA deprovisioned it; gateway marks it
221
+ // unavailable and redirects to nano-omni). Led the free chain until then.
222
+ 'nvidia/nemotron-3-nano-30b',
211
223
  'nvidia/nemotron-nano-9b-v2',
212
224
  'nvidia/mistral-nemotron',
213
225
  'nvidia/nemotron-nano-12b-v2-vl',
@@ -216,17 +228,18 @@ export const LEGACY_FREE_MODEL_IDS = [
216
228
  'nvidia/nemotron-3-super-120b-a12b-free',
217
229
  ];
218
230
  /**
219
- * Every free id in the live Base catalog as of 2026-08-30, mapped to its
220
- * catalogued context window. Used for metadata display; NOT for sizing a
221
- * request — use FREE_POOL_CONTEXT_FLOOR for that.
231
+ * Every free id (billing_mode "free") in the live catalog as of 2026-10-20 —
232
+ * identical on Base and Solana — mapped to its catalogued context window.
233
+ * Used for metadata display and as the cold-cache answer to isFreeModelId();
234
+ * NOT for sizing a request — use FREE_POOL_CONTEXT_FLOOR for that.
222
235
  */
223
236
  export const FREE_MODEL_CONTEXT_WINDOWS = {
224
237
  'nvidia/nemotron-3-ultra-550b': 1_000_000,
225
238
  'nvidia/nemotron-3.5-lightning': 1_000_000,
226
239
  'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning': 256_000,
227
240
  'cohere/north-mini-code': 256_000,
228
- 'nvidia/nemotron-3-nano-30b': 131_072,
229
241
  'poolside/laguna-xs-2.1': 131_072,
242
+ 'nvidia/muse-glimmer-30b': 131_072,
230
243
  'nvidia/llama-3.2-11b-vision': 128_000,
231
244
  };
232
245
  /**
package/dist/pricing.js CHANGED
@@ -9,13 +9,13 @@ const LEGACY_PRICING = {
9
9
  // their parser mapping promotes them to Auto upstream of cost estimation.
10
10
  'blockrun/auto': { input: 0.8, output: 4.0 },
11
11
  'blockrun/free': { input: 0, output: 0 },
12
- // FREE — BlockRun gateway free tier (refreshed 2026-08-30 to match live
13
- // /api/v1/models, which rotated the entire free pool). The routing order and
12
+ // FREE — BlockRun gateway free tier (refreshed 2026-10-20 to match live
13
+ // /api/v1/models, identical on Base and Solana). The routing order and
14
14
  // the reasoning behind it live in src/free-models.ts; this table only has to
15
15
  // price every id at $0 so no free call is ever reported as a charge.
16
16
  'nvidia/nemotron-3-ultra-550b': { input: 0, output: 0 }, // strongest free model, but 0/5 reachable — NOT the default
17
- 'nvidia/nemotron-3-nano-30b': { input: 0, output: 0 }, // leads the free chain on Base (see free-models.ts)
18
17
  'nvidia/nemotron-3.5-lightning': { input: 0, output: 0 },
18
+ 'nvidia/muse-glimmer-30b': { input: 0, output: 0 }, // Meta Muse Glimmer, free on NVIDIA since 2026-10-02
19
19
  'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning': { input: 0, output: 0 },
20
20
  'nvidia/llama-3.2-11b-vision': { input: 0, output: 0 },
21
21
  'cohere/north-mini-code': { input: 0, output: 0 },
@@ -28,6 +28,7 @@ const LEGACY_PRICING = {
28
28
  'nvidia/nemotron-nano-12b-v2-vl': { input: 0, output: 0 }, // -> nemotron-3-nano-omni
29
29
  'nvidia/step-3.7-flash': { input: 0, output: 0 }, // -> nemotron-3.5-lightning
30
30
  'nvidia/nemotron-3-super-120b-a12b-free': { input: 0, output: 0 }, // pool backing id, never listed
31
+ 'nvidia/nemotron-3-nano-30b': { input: 0, output: 0 }, // delisted 2026-09-08 (NVIDIA deprovisioned it)
31
32
  // Retired free models (kept at 0 for legacy session-cost records; gateway no
32
33
  // longer serves these — do NOT route to them).
33
34
  'nvidia/qwen3-next-80b-a3b-instruct': { input: 0, output: 0 }, // NVIDIA EOL 410, 2026-07-27
@@ -73,6 +74,7 @@ const LEGACY_PRICING = {
73
74
  // per-1M table cannot express — the catalog's long_context tiers carry that.
74
75
  'openai/gpt-6-astra': { input: 10.0, output: 50.0 },
75
76
  'openai/gpt-6-sol': { input: 2.0, output: 10.0 },
77
+ 'openai/gpt-6.1-sol': { input: 2.0, output: 10.0 }, // refreshed Sol, live 2026-10-20 (same base rate as 6 Sol)
76
78
  'openai/gpt-6-luna': { input: 0.1, output: 0.5 },
77
79
  'openai/gpt-5.1': { input: 1.25, output: 10.0 },
78
80
  'openai/gpt-4.1-nano': { input: 0.1, output: 0.4 },
@@ -116,6 +118,7 @@ const LEGACY_PRICING = {
116
118
  'google/gemini-3.5-flash': { input: 1.5, output: 9.0 }, // latest Flash w/ thinking, 1M ctx
117
119
  'google/gemini-3.5-flash-lite': { input: 0.3, output: 2.5 },
118
120
  'google/gemini-3.6-flash': { input: 0.75, output: 3.75 }, // added upstream 2026-08-03 (base #329); live promo rate 2026-10-02
121
+ 'google/gemini-3.8-flash': { input: 0.75, output: 3.75 }, // live 2026-10-20
119
122
  'google/gemini-2.5-pro': { input: 1.25, output: 10.0 },
120
123
  'google/gemini-3.1-pro': { input: 2.0, output: 12.0 },
121
124
  // xAI
@@ -128,9 +131,13 @@ const LEGACY_PRICING = {
128
131
  // again upstream 2026-08-07 to $0.14/$0.28 — mirrored here 2026-08-12).
129
132
  'deepseek/deepseek-chat': { input: 0.14, output: 0.28 },
130
133
  'deepseek/deepseek-reasoner': { input: 0.14, output: 0.28 },
134
+ // V4.1 Flash — live on Base 2026-10-20; the same model and price as deepseek-chat.
135
+ 'deepseek/deepseek-v4.1-flash': { input: 0.14, output: 0.28 },
131
136
  // V4 Pro (1.6T MoE / 49B active, 1M ctx, 65K out). Live list 2026-10-02;
132
137
  // the old $0.435/$0.87 promo figure no longer matches the gateway.
133
138
  'deepseek/deepseek-v4-pro': { input: 1.32, output: 3.96 },
139
+ // V4 Flash Vision (experimental): the image-capable DeepSeek SKU, live 2026-10-20.
140
+ 'deepseek/deepseek-v4-flash-vision-exp': { input: 0.3, output: 1.2 },
134
141
  // Minimax
135
142
  'minimax/minimax-m3': { input: 0.3, output: 1.2 },
136
143
  'minimax/minimax-m2.7': { input: 0.3, output: 1.2 },
@@ -138,9 +145,17 @@ const LEGACY_PRICING = {
138
145
  'qwen/qwen3.7-max': { input: 1.475, output: 4.425 },
139
146
  'qwen/qwen3.7-plus': { input: 0.32, output: 1.28 },
140
147
  'qwen/qwen3.7-flash': { input: 0.03, output: 0.13 },
148
+ // Qwen 3.8 — live 2026-10-20. Flash is the vision-capable one.
149
+ 'qwen/qwen3.8-max': { input: 2.0, output: 6.0 },
150
+ 'qwen/qwen3.8-flash': { input: 0.15, output: 0.47 },
141
151
  // New providers added upstream 2026-08 — priced from the live catalog.
142
- 'tencent/hy3': { input: 0.132, output: 0.528 }, // retired upstream (hidden) — kept for session-cost records
152
+ 'tencent/hy3': { input: 0.132, output: 0.528 }, // retired 2026-09-24 (gateway redirects to qwen3.7-flash) — kept for session-cost records
153
+ 'tencent/hy4-preview': { input: 0.834, output: 2.501 }, // live 2026-10-20 — Tencent's current model
143
154
  'xiaomi/mimo-v2.5-pro': { input: 0.435, output: 0.87 },
155
+ 'xiaomi/mimo-v2.5': { input: 0.14, output: 0.28 }, // live 2026-10-20 — vision-capable, cheaper than Pro
156
+ // Mistral + Poolside paid tiers — live 2026-10-20.
157
+ 'mistral/mistral-large-4': { input: 0.68, output: 2.09 },
158
+ 'poolside/laguna-s-2.1': { input: 0.09, output: 0.18 }, // paid sibling of the free laguna-xs-2.1
144
159
  // Moonshot — K3 is the gateway flagship (2026-07): 2.8T open MoE, 1M
145
160
  // context, multimodal (image+text), returns reasoning_content. Pricier
146
161
  // than the K2.x line it replaced ($3/$15 vs K2.7's $0.95/$4).
@@ -168,28 +183,32 @@ const LEGACY_PRICING = {
168
183
  'openai/gpt-5-nano': { input: 0.05, output: 0.4 }, // (probed)
169
184
  'openai/o1-mini': { input: 1.1, output: 4.4 },
170
185
  'google/gemini-3-pro-preview': { input: 2.0, output: 12.0 }, // GA'd as gemini-3.1-pro
171
- 'xai/grok-4-fast': { input: 0.2, output: 0.5 },
172
- 'xai/grok-4-fast-reasoning': { input: 0.2, output: 0.5 }, // (probed)
173
- 'xai/grok-4-1-fast': { input: 0.2, output: 0.5 },
174
- 'xai/grok-4-1-fast-reasoning': { input: 0.2, output: 0.5 }, // (probed)
186
+ // Retired at xAI. The gateway pins every retired Grok slug to grok-4.3 and
187
+ // bills grok-4.3's $1.25/$2.50 (blockrun models.ts + model-maps.ts
188
+ // "Retired -> grok-4.3", repriced 2026-09-07/09). The old per-id list
189
+ // prices here overstated grok-3 / grok-4-0709 by 2.4x/6x and understated
190
+ // the fast family, so they now carry what the gateway actually charges.
191
+ 'xai/grok-4-fast': { input: 0.2, output: 0.5 }, // not in the gateway catalog — historical only
192
+ 'xai/grok-4-fast-reasoning': { input: 1.25, output: 2.5 }, // served as grok-4.3
193
+ 'xai/grok-4-1-fast': { input: 0.2, output: 0.5 }, // not in the gateway catalog — historical only
194
+ 'xai/grok-4-1-fast-reasoning': { input: 1.25, output: 2.5 }, // served as grok-4.3
175
195
  // The two non-reasoning ids sit in the shared Router's SIMPLE chains and
176
- // were charged today, yet had no row here — Franklin reported them as $0
196
+ // were charged, yet once had no row here — Franklin reported them as $0
177
197
  // and the portfolio scorer, which treats an unpriced id as infinitely
178
- // expensive, could never pick them over a priced rung. Priced with their
179
- // reasoning siblings (xAI bills the pair identically).
180
- 'xai/grok-4-1-fast-non-reasoning': { input: 0.2, output: 0.5 }, // (probed)
181
- 'xai/grok-4-fast-non-reasoning': { input: 0.2, output: 0.5 }, // (probed)
182
- 'xai/grok-4-0709': { input: 3.0, output: 15.0 }, // (probed) gateway lists $3/$15 (was mispriced here at $0.2/$1.5)
183
- 'xai/grok-3-mini': { input: 0.3, output: 0.5 }, // (probed)
184
- 'xai/grok-2-vision': { input: 2.0, output: 10.0 },
185
- 'xai/grok-3': { input: 3.0, output: 15.0 }, // (probed)
198
+ // expensive, could never pick them over a priced rung.
199
+ 'xai/grok-4-1-fast-non-reasoning': { input: 1.25, output: 2.5 }, // served as grok-4.3
200
+ 'xai/grok-4-fast-non-reasoning': { input: 1.25, output: 2.5 }, // served as grok-4.3
201
+ 'xai/grok-4-0709': { input: 1.25, output: 2.5 }, // served as grok-4.3 (was $3/$15)
202
+ 'xai/grok-3-mini': { input: 1.25, output: 2.5 }, // served as grok-4.3
203
+ 'xai/grok-2-vision': { input: 1.25, output: 2.5 }, // served as grok-4.3
204
+ 'xai/grok-3': { input: 1.25, output: 2.5 }, // served as grok-4.3 (was $3/$15)
186
205
  'minimax/minimax-m2.5': { input: 0.3, output: 1.2 },
187
206
  // K2.x (probed): the bare `kimi`/`k2.*` shortcuts resolve to K3, the
188
207
  // explicit ids still answer.
189
208
  'moonshot/kimi-k2.7': { input: 0.95, output: 4.0 },
190
209
  'moonshot/kimi-k2.6': { input: 0.95, output: 4.0 },
191
210
  'moonshot/kimi-k2.5': { input: 0.6, output: 3.0 },
192
- 'nvidia/kimi-k2.5': { input: 0.55, output: 2.5 },
211
+ 'nvidia/kimi-k2.5': { input: 0.6, output: 3.0 }, // gateway list (falls back to moonshot/kimi-k2.5, same price)
193
212
  'zai/glm-5.1-turbo': { input: 1.2, output: 4.0 }, // was a client alias for zai/glm-5-turbo
194
213
  };
195
214
  export function hasHistoricalPrice(model) { return model in LEGACY_PRICING; }
@@ -69,6 +69,13 @@ const VISION_MODELS = new Set([
69
69
  'openai/gpt-6-sol',
70
70
  'openai/gpt-6-luna',
71
71
  'openai/gpt-5.1',
72
+ // 2026-10-20 catalog sync — every new id the live catalog tags vision.
73
+ 'openai/gpt-6.1-sol',
74
+ 'google/gemini-3.8-flash',
75
+ 'deepseek/deepseek-v4-flash-vision-exp',
76
+ 'qwen/qwen3.8-flash',
77
+ 'xiaomi/mimo-v2.5',
78
+ 'mistral/mistral-large-4',
72
79
  // Google — vision baked into every Gemini SKU we surface (flash-lite excepted)
73
80
  'google/gemini-3.1-pro',
74
81
  'google/gemini-3.6-flash',
@@ -50,6 +50,9 @@ const LEGACY_SHORTCUTS = {
50
50
  'gpt-6-astra': 'openai/gpt-6-astra',
51
51
  'gpt-6-sol': 'openai/gpt-6-sol',
52
52
  'gpt-6-luna': 'openai/gpt-6-luna',
53
+ // GPT-6.1 Sol (live /api/v1/models 2026-10-20, Base + Solana): same $2/$10
54
+ // as GPT-6 Sol with a newer knowledge cutoff. Explicit pin only.
55
+ 'gpt-6.1-sol': 'openai/gpt-6.1-sol',
53
56
  'gpt-5.1': 'openai/gpt-5.1',
54
57
  'gpt-5.6-sol': 'openai/gpt-5.6-sol',
55
58
  'gpt-5.6-terra': 'openai/gpt-5.6-terra',
@@ -104,16 +107,21 @@ const LEGACY_SHORTCUTS = {
104
107
  gemini: 'google/gemini-3.1-pro',
105
108
  'gemini-2.5': 'google/gemini-2.5-pro',
106
109
  'gemini-2.5-pro': 'google/gemini-2.5-pro',
107
- // `flash` follows the newest Flash generation (3.6 since 2026-08). 3.5 stays
108
- // reachable by its explicit id — it is dearer on output ($9 vs $7.5) with no
109
- // capability edge, so nothing should be pinned to it deliberately.
110
- flash: 'google/gemini-3.6-flash',
111
- 'gemini-flash': 'google/gemini-3.6-flash',
110
+ // `flash` follows the newest Flash generation — 3.8 since 2026-10-20, at
111
+ // the same $0.75/$3.75 as 3.6 and live on both chains. 3.6 stays reachable
112
+ // as `gemini-3.6`. 3.5 stays reachable by its explicit id — it is dearer
113
+ // ($1.5/$9) with no capability edge, so nothing should be pinned to it
114
+ // deliberately.
115
+ flash: 'google/gemini-3.8-flash',
116
+ 'gemini-flash': 'google/gemini-3.8-flash',
112
117
  'gemini-3.6': 'google/gemini-3.6-flash',
113
118
  'gemini-3.6-flash': 'google/gemini-3.6-flash',
114
119
  'gemini-3.5-flash': 'google/gemini-3.5-flash',
115
120
  'gemini-2.5-flash': 'google/gemini-2.5-flash',
116
121
  'gemini-3-flash-preview': 'google/gemini-3-flash-preview',
122
+ // Gemini 3.8 Flash (live 2026-10-20, both chains) — same $0.75/$3.75 as 3.6.
123
+ 'gemini-3.8': 'google/gemini-3.8-flash',
124
+ 'gemini-3.8-flash': 'google/gemini-3.8-flash',
117
125
  // Flash Lite — thinking-mode Gemini for high-throughput work. `flash-lite`
118
126
  // follows the newest (3.5); 3.1 is cheaper still and stays explicit.
119
127
  'flash-lite': 'google/gemini-3.5-flash-lite',
@@ -134,22 +142,35 @@ const LEGACY_SHORTCUTS = {
134
142
  'grok-4.6': 'xai/grok-4.6',
135
143
  'grok-4.3': 'xai/grok-4.3',
136
144
  'grok-build': 'xai/grok-build-0.1',
137
- // grok-3 / grok-4-0709 / the grok-4-1-fast pair are hidden from
138
- // /v1/models but still served and charged (probed 2026-08-29) — explicit
139
- // pins keep resolving to the real ids. Note grok-3 and grok-4-0709 bill at
140
- // $3/$15; `grok` (4.5) is the cheaper flagship.
141
- 'grok-3': 'xai/grok-3',
142
- 'grok-4': 'xai/grok-4-0709',
143
- 'grok-fast': 'xai/grok-4-1-fast-reasoning',
144
- 'grok-4.1': 'xai/grok-4-1-fast-reasoning',
145
+ // grok-3 / grok-4-0709 / the grok-4-1-fast pair are retired at xAI. The
146
+ // gateway pins every one of them to grok-4.3 and bills grok-4.3's rate
147
+ // (blockrun model-maps.ts, "Retired -> grok-4.3"), so these pins name what
148
+ // actually answers. `grok-4` is the generation alias and follows `grok`.
149
+ 'grok-3': 'xai/grok-4.3',
150
+ 'grok-4': 'xai/grok-4.5',
151
+ 'grok-fast': 'xai/grok-4.3',
152
+ 'grok-4.1': 'xai/grok-4.3',
145
153
  // DeepSeek — paid SKUs route through deepseek/* (gateway aliases serve V4
146
154
  // Flash modes upstream); free tier routes through nvidia/*.
147
- deepseek: 'deepseek/deepseek-chat', // V4 Flash Chat (paid, $0.20/$0.40)
155
+ deepseek: 'deepseek/deepseek-v4.1-flash', // V4.1 Flash (paid, $0.14/$0.28)
148
156
  r1: 'deepseek/deepseek-reasoner', // V4 Flash Reasoner (paid)
149
157
  // V4 Pro: paid flagship, 1.6T MoE / 49B active, 1M ctx, $1.32/$3.96 (live 2026-10-02).
150
158
  'deepseek-v4-pro': 'deepseek/deepseek-v4-pro',
151
159
  'dsv4-pro': 'deepseek/deepseek-v4-pro',
152
160
  'v4-pro': 'deepseek/deepseek-v4-pro',
161
+ // V4.1 Flash (live on Base and Solana 2026-10-20, $0.14/$0.28): the
162
+ // gateway's own name for the model deepseek-chat also serves — same
163
+ // upstream, same price, same non-thinking default. `deepseek` points here
164
+ // now that both chains list it (the `glm` rule). NOTE `deepseek-v4-flash` /
165
+ // `dsv4` are FREE aliases (below) and must not move to this paid id.
166
+ 'deepseek-v4.1-flash': 'deepseek/deepseek-v4.1-flash',
167
+ 'deepseek-v4.1': 'deepseek/deepseek-v4.1-flash',
168
+ 'v4.1-flash': 'deepseek/deepseek-v4.1-flash',
169
+ // V4 Flash Vision (experimental, live 2026-10-20): the paid DeepSeek SKU
170
+ // that takes images, $0.30/$1.20. Names match the shared catalog policy.
171
+ 'deepseek-v4-flash-vision-exp': 'deepseek/deepseek-v4-flash-vision-exp',
172
+ 'deepseek-vision': 'deepseek/deepseek-v4-flash-vision-exp',
173
+ 'v4-flash-vision': 'deepseek/deepseek-v4-flash-vision-exp',
153
174
  // The free nvidia/deepseek-v4-flash SKU was EOL'd by the gateway (410).
154
175
  // Point the deepseek-free aliases at the current free default so muscle
155
176
  // memory keeps working without handing back a dead id.
@@ -192,12 +213,17 @@ const LEGACY_SHORTCUTS = {
192
213
  omni: FREE_DEFAULT_MODEL,
193
214
  'free-omni': FREE_DEFAULT_MODEL,
194
215
  'nano-omni': FREE_DEFAULT_MODEL,
195
- // Base-only free ids. They 400 on Solana, which is a clear error rather than
196
- // a silent wrong model, so the shortcuts stay honest and resolve to the id
197
- // the user typed.
198
- 'nano-30b': 'nvidia/nemotron-3-nano-30b',
216
+ // nemotron-3-nano-30b was delisted by the gateway on 2026-09-08 (NVIDIA
217
+ // deprovisioned it), so `nano-30b` joins the retired-free-id pattern.
218
+ 'nano-30b': FREE_DEFAULT_MODEL,
219
+ // Explicit pins to other live free ids (Base + Solana, 2026-10-20) — they
220
+ // resolve to the id the user typed, never to the default.
199
221
  laguna: 'poolside/laguna-xs-2.1',
200
222
  'ultra-550b': 'nvidia/nemotron-3-ultra-550b',
223
+ // Meta's Muse Glimmer 30B, free on NVIDIA (gateway 2026-10-02). A reasoning
224
+ // model: give it a real max_tokens budget or it answers blank.
225
+ 'muse-glimmer': 'nvidia/muse-glimmer-30b',
226
+ 'muse-glimmer-30b': 'nvidia/muse-glimmer-30b',
201
227
  // Maverick left the gateway catalog on/before 2026-07-14 and the free pool
202
228
  // silently substitutes for it, so pointing users at it would promise a model
203
229
  // they don't get. Resolve to the current free default instead.
@@ -216,6 +242,13 @@ const LEGACY_SHORTCUTS = {
216
242
  'qwen-max': 'qwen/qwen3.7-max',
217
243
  'qwen3.7-max': 'qwen/qwen3.7-max',
218
244
  'qwen-3.7-max': 'qwen/qwen3.7-max',
245
+ // Qwen 3.8 (live 2026-10-20, both chains). Max is $2/$6, Flash $0.15/$0.47
246
+ // and takes images. Explicit pins — `qwen-max` / `qwen-flash` stay on 3.7.
247
+ 'qwen3.8-max': 'qwen/qwen3.8-max',
248
+ 'qwen-3.8-max': 'qwen/qwen3.8-max',
249
+ 'qwen3.8-flash': 'qwen/qwen3.8-flash',
250
+ 'qwen-3.8-flash': 'qwen/qwen3.8-flash',
251
+ 'qwen-vision': 'qwen/qwen3.8-flash',
219
252
  // Plus and Flash round out the paid Qwen line — both 1M context with
220
253
  // reasoning. Flash at $0.03/$0.13 is the cheapest paid model on the gateway.
221
254
  'qwen-plus': 'qwen/qwen3.7-plus',
@@ -223,16 +256,13 @@ const LEGACY_SHORTCUTS = {
223
256
  'qwen-flash': 'qwen/qwen3.7-flash',
224
257
  'qwen3.7-flash': 'qwen/qwen3.7-flash',
225
258
  // GLM-5.3 (2026-08) is Z.AI's flagship — same $1.4/$4.4 as 5.2 with 1M
226
- // context and always-on reasoning — but it ships on the BASE gateway ONLY
227
- // (sol.blockrun.ai lists 92 of Base's 93 models, and 5.3 is the one it's
228
- // missing; verified 2026-08-20).
229
- //
230
- // So `glm` deliberately does NOT follow the flagship here. A bare alias is
231
- // what people type from muscle memory, and it has to resolve on BOTH chains
232
- // — pointing it at a Base-only id would hand every Solana user an HTTP 400
233
- // on `/model glm`. 5.2 is on both, identically priced, one generation back.
234
- // Promote this line the day 5.3 lands on Solana.
235
- glm: 'zai/glm-5.2',
259
+ // context and always-on reasoning. It shipped Base-only, so `glm` held at 5.2
260
+ // (a bare alias has to resolve on BOTH chains, or every Solana user gets an
261
+ // HTTP 400 on `/model glm`) with a note to promote it the day 5.3 reached
262
+ // Solana. It has: sol.blockrun.ai lists zai/glm-5.3 at $1.4/$4.4 (verified
263
+ // 2026-10-20), so `glm` / `glm5` now follow the flagship like `gpt`, `grok`
264
+ // and `gemini`. 5.2 stays reachable as `glm-5.2`.
265
+ glm: 'zai/glm-5.3',
236
266
  'glm-5': 'zai/glm-5',
237
267
  'glm-5.3': 'zai/glm-5.3',
238
268
  // GLM-5.3 Flash (2026-08-29 catalog sync): 1M ctx, vision, $0.15/$0.5.
@@ -243,13 +273,20 @@ const LEGACY_SHORTCUTS = {
243
273
  // routes for anyone who wants the 200K-context build explicitly.
244
274
  'glm-5.1': 'zai/glm-5.1',
245
275
  'glm-turbo': 'zai/glm-5-turbo',
246
- 'glm5': 'zai/glm-5.2',
276
+ 'glm5': 'zai/glm-5.3',
247
277
  // Tencent + Xiaomi joined the gateway in 2026-08 — cheap reasoning at long
248
- // context, below the frontier tier.
249
- hy3: 'tencent/hy3',
250
- tencent: 'tencent/hy3',
278
+ // context, below the frontier tier. tencent/hy3 was retired 2026-09-24 (the
279
+ // gateway now redirects it to qwen3.7-flash); Hy4 Preview is Tencent's live
280
+ // model, so `hy3` / `tencent` follow it the way the K2.x pins follow K3.
281
+ hy3: 'tencent/hy4-preview',
282
+ tencent: 'tencent/hy4-preview',
283
+ hy4: 'tencent/hy4-preview',
284
+ 'hy4-preview': 'tencent/hy4-preview',
251
285
  mimo: 'xiaomi/mimo-v2.5-pro',
252
286
  'mimo-v2.5-pro': 'xiaomi/mimo-v2.5-pro',
287
+ // MiMo V2.5 (non-Pro): cheaper and the vision-capable one ($0.14/$0.28).
288
+ 'mimo-v2.5': 'xiaomi/mimo-v2.5',
289
+ 'mimo-vision': 'xiaomi/mimo-v2.5',
253
290
  xiaomi: 'xiaomi/mimo-v2.5-pro',
254
291
  kimi: 'moonshot/kimi-k3',
255
292
  k3: 'moonshot/kimi-k3',
@@ -260,6 +297,16 @@ const LEGACY_SHORTCUTS = {
260
297
  'k2.6': 'moonshot/kimi-k3',
261
298
  'kimi-k2.5': 'moonshot/kimi-k3',
262
299
  'k2.5': 'moonshot/kimi-k3',
300
+ // Mistral Large 4 (live 2026-10-20, both chains): $0.68/$2.09, reasoning +
301
+ // vision. Bare `mistral` is safe to take — `mistral-small` / `mistral-nemotron`
302
+ // are the free aliases and stay free.
303
+ mistral: 'mistral/mistral-large-4',
304
+ 'mistral-large': 'mistral/mistral-large-4',
305
+ 'mistral-large-4': 'mistral/mistral-large-4',
306
+ // Poolside Laguna S 2.1 — the PAID sibling of the free laguna-xs ($0.09/$0.18).
307
+ // `laguna` stays on the free XS model: a free alias never moves to a paid one.
308
+ 'laguna-s': 'poolside/laguna-s-2.1',
309
+ 'laguna-s-2.1': 'poolside/laguna-s-2.1',
263
310
  };
264
311
  /**
265
312
  * Resolve a model name — supports shortcuts. Returns the canonical model id.
@@ -320,6 +367,9 @@ const PROVIDER_ORDER = [
320
367
  'deepseek',
321
368
  'tencent',
322
369
  'xiaomi',
370
+ 'mistral',
371
+ 'poolside',
372
+ 'cohere',
323
373
  'nvidia',
324
374
  ];
325
375
  const PROVIDER_LABELS = {
@@ -334,6 +384,9 @@ const PROVIDER_LABELS = {
334
384
  deepseek: 'DeepSeek',
335
385
  tencent: 'Tencent / Hunyuan',
336
386
  xiaomi: 'Xiaomi / MiMo',
387
+ mistral: 'Mistral',
388
+ poolside: 'Poolside / Laguna',
389
+ cohere: 'Cohere',
337
390
  nvidia: 'Free / NVIDIA',
338
391
  };
339
392
  /**
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@blockrun/franklin",
3
- "version": "3.49.12",
3
+ "version": "3.49.13",
4
4
  "description": "Franklin Agent — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
5
5
  "type": "module",
6
6
  "exports": {
@@ -76,8 +76,8 @@
76
76
  },
77
77
  "dependencies": {
78
78
  "@blockrun/llm": "^3.19.1",
79
- "@blockrun/model-catalog": "https://codeload.github.com/BlockRunAI/model-catalog/tar.gz/465c45941fb8ea9d6858fe8402abc8ba8934be90",
80
- "@blockrun/router-core": "https://codeload.github.com/BlockRunAI/router-core/tar.gz/65a92c85ea1f187a64e7611302ef6add236b0cba",
79
+ "@blockrun/model-catalog": "https://codeload.github.com/BlockRunAI/model-catalog/tar.gz/fd635936832b414e40f8fe6c91fc1ad57347cd63",
80
+ "@blockrun/router-core": "https://codeload.github.com/BlockRunAI/router-core/tar.gz/2cd5a1026f755cd39a3f0d6e61cab23a5ad391a9",
81
81
  "@colbymchenry/codegraph": "^1.5.0",
82
82
  "@modelcontextprotocol/sdk": "^1.29.0",
83
83
  "@polymarket/builder-relayer-client": "^0.0.10",