feat: reprice Stars from real provider costs and refresh the model lineup

Pricing was not derived from anything. It had inversions — DeepSeek R1, a
32B distill with no tool support, cost 80 Stars while Kimi K2.6 cost 45 — and
markups ranging from ~60x on chat to ~4000x on image generation, so the ladder
told users nothing about relative value.

Prices are now set from measured provider rates. Assuming ~$0.009 net per Star
(the mobile figure, after the app-store cut) and ~6k input / 800 output tokens
for a tool-using turn once the multi-turn loop is counted, every model is
priced at ~40x break-even:

    glm-4.7-flash          5 ->   3      nemotron-3        20 ->  18
    gemma4                 8 ->   4      kimi-k2.6         45 ->  40
    llama-3.2-vision       8 ->   4      glm-5.2            - ->  55
    llama-4-scout          - ->  10      gemini-3.1-pro   150 -> 100
    gpt-oss-120b           - ->  12      /photo           100 ->  20
    gemini-3-flash        25 ->  12      voice surcharge   20 ->   3
                                         upload             5 ->   2

Capability flags are now verified against the live API instead of assumed, and
two of them were wrong in the dangerous direction:

- glm-4.7-flash, the DEFAULT model, claimed supportsVision. The API rejects
  images outright ("GLM-4.7-Flash is not a multimodal model"), so the vision
  fallback in chargeStars never fired and photos from default users failed.
- gemma4 and llama-3.2-vision claimed supportsTools. Both accept a tools array
  and never emit a tool call.

Lineup changes: added llama-4-scout (the only model confirmed to do both
vision and tool calls, so it is also the new vision fallback), gpt-oss-120b and
glm-5.2; dropped deepseek-r1-32b (no tools, dearest output in the catalogue)
and gemini-3.1-flash-lite (unpublished third-party rate). Stored preferences
for removed models fall through to the default rather than erroring.

Photo and voice were the worst offenders: flux-1-schnell costs ~$0.0002 an
image and Whisper $0.0005 a minute, so 100 Stars made an image five times
dearer than a full conversation with the best model available.

Note: @cf/* rates come from Cloudflare's published pricing; the gemini-* models
are third-party and Cloudflare does not publish their rates, so Google list
prices are used as the basis. The 40x multiple absorbs that uncertainty.

Also fixes the Telegram link on the docs site to https://t.me/TuxRobot.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GaLeuxVikb3iH2Etw2tHe7

Author: codebam

bot

@@ -455,47 +455,90 @@ export const SYSTEM_PROMPTS = {
 
 export const DEFAULT_MODEL = 'glm-4.7-flash';
 
-/** Extra Stars charged on top of the model cost for Whisper transcription. */
-export const VOICE_SURCHARGE_STARS = 20;
+/**
+ * Extra Stars on top of the model cost for Whisper transcription.
+ * Whisper bills $0.0005 per audio minute, so the old +20 was a ~700x markup on
+ * a step that costs a fraction of the reply it feeds.
+ */
+export const VOICE_SURCHARGE_STARS = 3;
 
-/** Stars charged per web app file upload. */
-export const UPLOAD_COST_STARS = 5;
+/**
+ * Stars per web app file upload. An R2 put plus a KV write costs essentially
+ * nothing; this exists as an anti-spam deterrent, not as a revenue line.
+ */
+export const UPLOAD_COST_STARS = 2;
 
-/** Cost of one `/photo` image generation. */
-export const PHOTO_COST_STARS = 100;
+/**
+ * Cost of one `/photo` generation. flux-1-schnell bills $0.0000528 per 512px
+ * tile — about $0.0002 for a 1024x1024 image — so charging 100 Stars made an
+ * image five times dearer than a full conversation with our best model.
+ */
+export const PHOTO_COST_STARS = 20;
 
 /** Hard cap on history entries accepted from an untrusted request body. */
 export const MAX_HISTORY_MESSAGES = 40;
 
+/**
+ * Model catalogue.
+ *
+ * `cost` is in Telegram Stars and is derived from real provider token prices
+ * rather than picked by feel — see docs/pricing notes below. A Star nets us
+ * roughly $0.009 after the mobile app-store cut, and a tool-using turn in this
+ * bot costs about 6k input / 800 output tokens once the multi-turn tool loop is
+ * accounted for. Prices are set at ~40x that break-even, rounded, which leaves
+ * headroom for long contexts and retries while keeping the ladder proportional
+ * to what each model actually costs.
+ *
+ * `supportsTools` and `supportsVision` are VERIFIED against the live API, not
+ * assumed. Getting these wrong is not cosmetic: a model wrongly flagged as
+ * vision-capable never triggers the fallback in chargeStars, so the user's
+ * photo silently fails. Both flags fail safe — when unsure, leave them off.
+ */
 export const AVAILABLE_MODELS: Record<
 	string,
 	{ id: string; cost: number; supportsTools?: boolean; supportsVision?: boolean }
 > = {
-	gemma4: { id: '@cf/google/gemma-4-26b-a4b-it', cost: 8, supportsTools: true, supportsVision: true },
-	'google/gemini-3-flash': { id: 'google/gemini-3-flash', cost: 25, supportsTools: true, supportsVision: true },
-	'google/gemini-3.1-flash-lite': {
-		id: 'google/gemini-3.1-flash-lite',
-		cost: 12,
+	// Cheap default. Emits real tool calls; explicitly NOT multimodal
+	// ("GLM-4.7-Flash is not a multimodal model").
+	'glm-4.7-flash': { id: '@cf/zai-org/glm-4.7-flash', cost: 3, supportsTools: true },
+	// Accepts a tools array but never emits a tool call, so don't advertise it.
+	gemma4: { id: '@cf/google/gemma-4-26b-a4b-it', cost: 4 },
+	// Budget vision. Same story on tools: accepted, never used.
+	'llama-3.2-vision': { id: '@cf/meta/llama-3.2-11b-vision-instruct', cost: 4, supportsVision: true },
+	// The all-rounder: confirmed tool calls *and* image understanding, which is
+	// why it is also the automatic fallback for images.
+	'llama-4-scout': {
+		id: '@cf/meta/llama-4-scout-17b-16e-instruct',
+		cost: 10,
 		supportsTools: true,
 		supportsVision: true
 	},
-	'google/gemini-3.1-pro': { id: 'google/gemini-3.1-pro', cost: 150, supportsTools: true, supportsVision: true },
-	'llama-3.2-vision': {
-		id: '@cf/meta/llama-3.2-11b-vision-instruct',
-		cost: 8,
+	'gpt-oss-120b': { id: '@cf/openai/gpt-oss-120b', cost: 12, supportsTools: true },
+	'google/gemini-3-flash': {
+		id: 'google/gemini-3-flash',
+		cost: 12,
 		supportsTools: true,
 		supportsVision: true
 	},
-	'kimi-k2.6': { id: '@cf/moonshotai/kimi-k2.6', cost: 45, supportsTools: true, supportsVision: true },
-	'glm-4.7-flash': { id: '@cf/zai-org/glm-4.7-flash', cost: 5, supportsTools: true, supportsVision: true },
-	'deepseek-r1-32b': {
-		id: '@cf/deepseek-ai/deepseek-r1-distill-qwen-32b',
-		cost: 80,
-		supportsTools: false
-	},
-	'nemotron-3': { id: '@cf/nvidia/nemotron-3-120b-a12b', cost: 20, supportsTools: true }
+	'nemotron-3': { id: '@cf/nvidia/nemotron-3-120b-a12b', cost: 18, supportsTools: true },
+	'kimi-k2.6': { id: '@cf/moonshotai/kimi-k2.6', cost: 40, supportsTools: true },
+	'glm-5.2': { id: '@cf/zai-org/glm-5.2', cost: 55, supportsTools: true },
+	// Priciest by a wide margin ($2/$12 per 1M at list), and priced to match.
+	'google/gemini-3.1-pro': {
+		id: 'google/gemini-3.1-pro',
+		cost: 100,
+		supportsTools: true,
+		supportsVision: true
+	}
 };
 
+/**
+ * Model used when the user's choice cannot handle an image. Must support both
+ * vision and tools, and must have a known price — hence a first-party model
+ * rather than a third-party one whose rate Cloudflare does not publish.
+ */
+export const VISION_FALLBACK_MODEL = 'llama-4-scout';
+
 /** Look up a model entry by its provider-facing id (e.g. `@cf/...`). */
 export function modelConfigById(id: string | undefined) {
 	if (!id) return undefined;

packages/shared/src/index.ts

@@ -221,7 +221,7 @@ const html = `<!doctype html>
       Cloudflare Workers, Workers AI, Durable Objects and Pages.
     </p>
     <div class="cta">
-      <a class="btn primary" href="https://t.me/tux_robot_bot">Open in Telegram</a>
+      <a class="btn primary" href="https://t.me/TuxRobot">Open in Telegram</a>
       <a class="btn" href="${REPO}">Source on GitHub</a>
     </div>
   </div>