rovecode 0.4.0-beta.3 → 0.4.0

package/src/providers/cache.ts

@@ -0,0 +1,223 @@
+/** Anthropic prompt-cache boundary placement — pure transform for the /messages request body.
+ *
+ *  Prompt caching is a PREFIX match: the API render order is tools → system → messages, and a
+ *  `cache_control: {type:"ephemeral"}` marker on a content block caches everything from the start
+ *  of the request up to and including that block. Any byte change before a marker invalidates it.
+ *  Therefore boundaries only pay off on STABLE prefixes — a boundary after content that changes
+ *  every turn is a wasted breakpoint (Anthropic allows at most 4 per request).
+ *
+ *  Pattern reference (hermes-agent snapshot, research/source_snapshots/hermes-agent):
+ *  - agent/prompt_caching.py:1-8 — "The default layout uses 4 cache_control breakpoints: the static
+ *    system prefix, the end of the system prompt, and the last 2 non-system messages."
+ *  - agent/prompt_caching.py:157-162 (_build_marker) — the marker is `{"type": "ephemeral"}`.
+ *  - agent/prompt_caching.py:113-121 (_apply_cache_marker) — string content is converted to a
+ *    single text block carrying the marker; list content gets the marker on its LAST block.
+ *  - agent/prompt_caching.py:99-103, 331-333 — volatile tails ride UNMARKED ("a changed ticket ID
+ *    or timestamp no longer invalidates the whole skill body"), and empty text blocks are never
+ *    emitted (HTTP 400 on native Anthropic).
+ *  - agent/prompt_caching.py:567-573 (apply_anthropic_cache_control) — idempotence contract:
+ *    re-application "can never accumulate past 4 markers".
+ *  - agent/system_prompt.py:911-924 and agent/turn_context.py:214-217 — the stability principle:
+ *    volatile per-turn facts are kept OUT of / AFTER the cached band so the prefix stays
+ *    byte-stable turn-over-turn.
+ *
+ *  Placement policy implemented here (deliberately simpler than Hermes' 4-marker layout):
+ *  (1) SYSTEM boundary — `system: string` is converted to
+ *      `[{type:"text", text, cache_control:{type:"ephemeral"}}]` when it is ≥ minChunkChars.
+ *      A system-block array already provided by the caller gets the marker on its last block
+ *      (unless any block is already marked). Because caching is prefix-based, this single marker
+ *      also covers the `tools` array rendered before `system` — no separate tools marker needed.
+ *  (2) CONVERSATION-PREFIX boundary — the marker goes on the LAST content block of the most
+ *      recent message that has ≥ 2 messages after it (index length-3): an agent turn appends
+ *      the trailing exchange (assistant tool_use + tool_result, or user + assistant), so the
+ *      tail churns every turn while the prefix up to there is CANONICALLY stable — not
+ *      byte-identical. Two rewrites make consecutive requests differ in raw bytes: the marker
+ *      moves forward each turn (dropped from the old anchor, added at the new one), and the
+ *      marked message's string content is rewritten to [{type:"text",...}] by
+ *      withMarkerOnLastBlock below. Anthropic's cache key normalizes exactly those two
+ *      (cache_control is placement metadata, and string content is equivalent to a single
+ *      text block), so the previous request's prefix still cache-hits. Pinned by the
+ *      consecutive-turn stability test in test/unit/cache.test.ts.
+ *      If the anchor message cannot carry a marker (empty content — cf. Hermes _can_carry_marker,
+ *      prompt_caching.py:127-140, "so the breakpoints land on messages that count"), the
+ *      boundary reallocates to the nearest earlier eligible message. String content is
+ *      converted to a one-block array, preserving the text byte-for-byte.
+ *  (3) BUDGET — never exceed maxBreakpoints total, counting markers already present in
+ *      system/messages/tools before placing new ones. System wins over conversation prefix.
+ *  (4) IDEMPOTENT — a position that already carries a marker is left untouched, existing markers
+ *      count toward the budget, and untouched inputs are returned by reference, so
+ *      apply(apply(body)) deep-equals apply(body).
+ *  (5) Bodies without messages/system pass through untouched (same reference), as does anything
+ *      whose shapes don't match — this is a best-effort decorator, never a validator.
+ *
+ *  minChunkChars default: 4096 chars ≈ 1024 tokens (~4 chars/token) — the documented minimum
+ *  cacheable prefix for Sonnet/Opus-class models (model-dependent, 512–4096 tokens per the API
+ *  docs). Prefixes below the model minimum are silently NOT cached, so spending a breakpoint on
+ *  them is pure waste. For the message boundary the gate compares the CUMULATIVE prefix
+ *  (system + messages[0..anchor]) against minChunkChars, since that whole span is what the
+ *  marker would cache (tools text is not counted — a conservative undercount).
+ *
+ *  Pure: the input body is never mutated; modified paths are copied copy-on-write.
+ */
+
+export interface CacheOptions {
+  /** Maximum cache_control markers allowed in the request. Anthropic's limit — and the default — is 4. */
+  maxBreakpoints?: number;
+  /** Don't spend a breakpoint on a prefix shorter than this many characters. Default 4096 (≈1024 tokens). */
+  minChunkChars?: number;
+}
+
+export interface CacheControl { type: "ephemeral" }
+
+/** Loose ("-ish") view of the body anthropicStream builds (src/providers/stream.ts:152-160):
+ *  `{ model, max_tokens, system?: string, messages: [...], tools?: [...] }`. Every field is
+ *  unknown-typed so the adapter's `Record<string, unknown>` is directly assignable; this module
+ *  narrows at runtime and passes through anything it does not recognize. */
+export interface AnthropicishBody {
+  system?: unknown;
+  messages?: unknown;
+  tools?: unknown;
+  [key: string]: unknown;
+}
+
+export const DEFAULT_MAX_BREAKPOINTS = 4;
+export const DEFAULT_MIN_CHUNK_CHARS = 4096;
+
+const EPHEMERAL: CacheControl = { type: "ephemeral" };
+
+type Rec = Record<string, unknown>;
+
+function isRec(v: unknown): v is Rec {
+  return typeof v === "object" && v !== null && !Array.isArray(v);
+}
+
+/** Marker presence on a content block / tool entry (block-level only — the shapes this module emits). */
+function hasMarker(block: unknown): boolean {
+  return isRec(block) && block["cache_control"] !== undefined;
+}
+
+function blockText(block: unknown): string {
+  if (!isRec(block)) return "";
+  return typeof block["text"] === "string" ? block["text"] : "";
+}
+
+function systemChars(system: unknown): number {
+  if (typeof system === "string") return system.length;
+  if (Array.isArray(system)) return system.reduce((n: number, b) => n + blockText(b).length, 0);
+  return 0;
+}
+
+function contentChars(content: unknown): number {
+  if (typeof content === "string") return content.length;
+  if (!Array.isArray(content)) return 0;
+  let n = 0;
+  for (const b of content) {
+    n += blockText(b).length;
+    // tool_result blocks carry their payload under "content" (string form only; arrays undercount)
+    if (isRec(b) && typeof b["content"] === "string") n += b["content"].length;
+  }
+  return n;
+}
+
+function messageMarkerCount(msg: unknown): number {
+  if (!isRec(msg)) return 0;
+  const content = msg["content"];
+  return Array.isArray(content) ? content.filter(hasMarker).length : 0;
+}
+
+function countExistingMarkers(system: unknown, messages: readonly unknown[] | null, tools: unknown): number {
+  let n = 0;
+  if (Array.isArray(system)) n += system.filter(hasMarker).length;
+  for (const m of messages ?? []) n += messageMarkerCount(m);
+  if (Array.isArray(tools)) n += tools.filter(hasMarker).length;
+  return n;
+}
+
+/** A marker on this message would actually land on the wire: non-empty string content (an empty
+ *  string would convert to an empty text block — HTTP 400), or a non-empty block array whose last
+ *  element is an object. Mirrors Hermes _can_carry_marker (prompt_caching.py:127-154). */
+function canCarryMarker(msg: unknown): boolean {
+  if (!isRec(msg)) return false;
+  const content = msg["content"];
+  if (typeof content === "string") return content.length > 0;
+  if (Array.isArray(content) && content.length > 0) return isRec(content[content.length - 1]);
+  return false;
+}
+
+/** Copy of `msg` with the ephemeral marker on its last content block; string content becomes a
+ *  one-block array with the text preserved byte-for-byte (Hermes prompt_caching.py:113-121). */
+function withMarkerOnLastBlock(msg: Rec): Rec {
+  const content = msg["content"];
+  if (typeof content === "string") {
+    return { ...msg, content: [{ type: "text", text: content, cache_control: EPHEMERAL }] };
+  }
+  if (Array.isArray(content) && content.length > 0) {
+    const last = content[content.length - 1];
+    if (isRec(last)) {
+      const blocks = content.slice();
+      blocks[blocks.length - 1] = { ...last, cache_control: EPHEMERAL };
+      return { ...msg, content: blocks };
+    }
+  }
+  return msg; // callers gate on canCarryMarker(); unreachable in practice
+}
+
+export function applyAnthropicCacheBoundaries(body: AnthropicishBody, opts?: CacheOptions): AnthropicishBody {
+  const maxBreakpoints = opts?.maxBreakpoints ?? DEFAULT_MAX_BREAKPOINTS;
+  const minChunkChars = opts?.minChunkChars ?? DEFAULT_MIN_CHUNK_CHARS;
+
+  const system = body.system;
+  const messages: readonly unknown[] | null = Array.isArray(body.messages) ? body.messages : null;
+
+  const hasSystem = (typeof system === "string" && system.length > 0) || (Array.isArray(system) && system.length > 0);
+  const hasMessages = messages !== null && messages.length > 0;
+  if (!hasSystem && !hasMessages) return body; // policy (5): pass through untouched
+
+  let used = countExistingMarkers(system, messages, body.tools);
+  let out = body;
+  const copyOnWrite = (): AnthropicishBody => (out === body ? (out = { ...body }) : out);
+
+  // ---- policy (1): system boundary ------------------------------------------------------------
+  if (used < maxBreakpoints && hasSystem) {
+    if (typeof system === "string") {
+      if (system.length >= minChunkChars) {
+        copyOnWrite().system = [{ type: "text", text: system, cache_control: EPHEMERAL }];
+        used += 1;
+      }
+    } else if (Array.isArray(system) && !system.some(hasMarker)) {
+      const last = system[system.length - 1];
+      if (isRec(last) && systemChars(system) >= minChunkChars) {
+        const blocks = system.slice();
+        blocks[blocks.length - 1] = { ...last, cache_control: EPHEMERAL };
+        copyOnWrite().system = blocks;
+        used += 1;
+      }
+    }
+  }
+
+  // ---- policy (2): conversation-prefix boundary ------------------------------------------------
+  // Anchor = most recent message with ≥2 messages after it (index length-3); the trailing two
+  // messages are this turn's churn. Reallocate backwards past marker-incapable messages.
+  if (used < maxBreakpoints && messages !== null && messages.length >= 3) {
+    let anchor = -1;
+    for (let i = messages.length - 3; i >= 0; i -= 1) {
+      if (canCarryMarker(messages[i])) { anchor = i; break; }
+    }
+    const msg = anchor >= 0 ? messages[anchor] : undefined;
+    if (isRec(msg) && messageMarkerCount(msg) === 0) {
+      let prefixChars = systemChars(system);
+      for (let i = 0; i <= anchor; i += 1) {
+        const m = messages[i];
+        if (isRec(m)) prefixChars += contentChars(m["content"]);
+      }
+      if (prefixChars >= minChunkChars) {
+        const next = messages.slice();
+        next[anchor] = withMarkerOnLastBlock(msg);
+        copyOnWrite().messages = next;
+        used += 1;
+      }
+    }
+  }
+
+  return out;
+}

package/src/providers/catalog-local.ts

@@ -0,0 +1,160 @@
+/** rovecode's own price table for models the models.dev snapshot does not carry — the API ALIASES a
+ *  vendor sells under (DeepSeek's `deepseek-chat` / `deepseek-reasoner` are aliases of a release; the
+ *  snapshot lists release ids), and xAI's retired grok-4 slugs, which still resolve but are served and
+ *  billed as grok-4.3. Without these a real run on those ids is "unpriced" in /cost and `rovecode model show`.
+ *
+ *  Rules: an entry here applies ONLY when neither the live models.dev layer nor the snapshot has the model
+ *  (catalog.ts findIn) — a later snapshot entry wins by construction. Every entry names the vendor page the
+ *  numbers were read from and the day; "fetched" means the page was read on that day, "unverified" means the
+ *  page no longer documents the id and the row is the last known value. `rovecode model show` prints
+ *  "priced from rovecode's own table" so nobody mistakes them for models.dev's. Prices are USD per million
+ *  tokens. Output limits are given only where the vendor publishes one (the runtime caps at MAX_OUTPUT_CAP). */
+
+export interface LocalModel {
+  context: number;
+  output?: number;
+  /** has a reasoning mode (thinking.ts sends the dial only when this is not false) */
+  reasoning: boolean;
+  toolCall: boolean;
+  image?: boolean;
+  cost: { input: number; output: number; cacheRead?: number; cacheWrite?: number };
+  /** where the numbers were read — the reader's warrant */
+  source: string;
+  /** when, and how: "YYYY-MM-DD (fetched)" or "YYYY-MM-DD (unverified: …)" */
+  checked: string;
+  /** Take precedence over a models.dev entry for this model. Off by default and deliberately rare: the
+   *  overlay exists to fill gaps, and a snapshot that gains a model should normally win. Set it ONLY when
+   *  the vendor's own page was read and disagrees with the snapshot, and say so in `source` — a reader has
+   *  to be able to see which number rovecode chose and why. */
+  override?: true;
+}
+
+/** A vendor whose price CHANGES WITH PROMPT SIZE. models.dev has no field for this, so it lives beside the
+ *  overlay and augments a row whatever its source — the snapshot keeps owning the base prices.
+ *
+ *  `mode` is the part that is easy to get wrong and expensive to get wrong:
+ *   - "per-request": crossing the threshold reprices THE WHOLE REQUEST, every token, at the upper rate
+ *     (xAI). A 201k-token prompt costs 201k × the high rate, not 200k low + 1k high.
+ *   - "marginal": only the tokens past the threshold take the upper rate. NO vendor row uses this today —
+ *     it exists because the two shapes are not interchangeable and a cost function must branch on which one
+ *     it has, rather than assume. Do not classify a vendor as marginal without wording that says so. */
+export interface PriceTier {
+  thresholdTokens: number;
+  mode: "per-request" | "marginal";
+  above: { input: number; output: number; cacheRead?: number; cacheWrite?: number };
+  source: string;
+  checked: string;
+}
+
+/** xAI, docs.x.ai/developers/migration/may-15-retirement (fetched 2026-09-04): grok-4-0709, grok-4-fast-reasoning,
+ *  grok-4-fast-non-reasoning, grok-4-1-fast-* and grok-3 were retired on 2026-05-15; the slugs still resolve and
+ *  are SERVED BY grok-4.3 — reasoning slugs at low effort, non-reasoning ones at none — and BILLED at grok-4.3
+ *  rates ($1.25 in / $2.50 out / $0.20 cached per 1M under 200k context; docs.x.ai/docs/models/grok-4.3, fetched
+ *  the same day: 1M context, configurable reasoning none|low|medium|high). So a request to `grok-4` costs this. */
+const GROK_43_REDIRECT = { context: 1_000_000, toolCall: true, image: true, cost: { input: 1.25, output: 2.5, cacheRead: 0.2 },
+  source: "docs.x.ai May-15-2026 retirement: served by grok-4.3, billed at grok-4.3 rates (under 200k context)", checked: "2026-09-04 (fetched)" } as const;
+
+/** Tiered pricing, by provider and model id. Verified 2026-09-05 against the vendors' own pricing pages.
+ *  Nothing here changes a base price — `catalog.ts` attaches the tier to whatever row it resolved, so a
+ *  models.dev entry keeps its numbers and gains the threshold it has no field for. */
+export const PRICE_TIERS: Readonly<Record<string, Readonly<Record<string, PriceTier>>>> = {
+  // xAI: "requests over 200,000 prompt tokens are billed at the higher rate FOR THE ENTIRE REQUEST"
+  xai: {
+    "grok-4.6": { thresholdTokens: 200_000, mode: "per-request", above: { input: 4, output: 12, cacheRead: 1 },
+      source: "docs.x.ai/developers/pricing — over 200k prompt tokens the whole request bills at the high rate", checked: "2026-09-05 (fetched)" },
+    "grok-4.3": { thresholdTokens: 200_000, mode: "per-request", above: { input: 2.5, output: 5, cacheRead: 0.4 },
+      source: "docs.x.ai/developers/pricing — over 200k prompt tokens the whole request bills at the high rate", checked: "2026-09-05 (fetched)" },
+  },
+  // OpenAI, developers.openai.com/api/docs/models/gpt-5.4 and /gpt-5.5: "Prompts with >272K input tokens
+  // are priced at 2x input and 1.5x output for the full session." The multipliers are ASYMMETRIC — input
+  // doubles, output goes up by half — so these are written out rather than derived from one rate.
+  // The -pro tiers are NOT given this: their own pages do not carry the sentence, and inventing a
+  // threshold for a $30/1M model is the expensive kind of guess.
+  openai: {
+    "gpt-5.4": { thresholdTokens: 272_000, mode: "per-request", above: { input: 5, output: 22.5, cacheRead: 0.5 },
+      source: "developers.openai.com/api/docs/models/gpt-5.4 — >272K input tokens: 2x input, 1.5x output for the full session", checked: "2026-09-05 (fetched)" },
+    "gpt-5.5": { thresholdTokens: 272_000, mode: "per-request", above: { input: 10, output: 45, cacheRead: 1 },
+      source: "developers.openai.com/api/docs/models/gpt-5.5 — >272K input tokens: 2x input, 1.5x output for the full session", checked: "2026-09-05 (fetched)" },
+  },
+  // Google prices by PROMPT SIZE, not by overflow: the page's own columns read "prompts <= 200k tokens"
+  // and "prompts > 200k tokens" for BOTH input and output, so a 201k prompt bills every token — and that
+  // request's output — at the upper rate. Same shape as xAI, which is why neither is "marginal".
+  google: {
+    "gemini-3.1-pro-preview": { thresholdTokens: 200_000, mode: "per-request", above: { input: 4, output: 18 },
+      source: "ai.google.dev/gemini-api/docs/pricing — \"$4.00, prompts > 200k tokens\" / \"$18.00, prompts > 200k\"", checked: "2026-09-05 (fetched)" },
+    "gemini-2.5-pro": { thresholdTokens: 200_000, mode: "per-request", above: { input: 2.5, output: 15 },
+      source: "ai.google.dev/gemini-api/docs/pricing — \"$2.50, prompts > 200k tokens\" / \"$15.00, prompts > 200k\"", checked: "2026-09-05 (fetched)" },
+  },
+} as const;
+
+/** A price that is scheduled to change, or a condition the number depends on. Not a tier and not an
+ *  override — just the sentence a reader needs so today's number is not mistaken for a permanent one.
+ *  Attached to a row whatever its source. */
+export const PRICE_NOTES: Readonly<Record<string, Readonly<Record<string, string>>>> = {
+  google: {
+    "gemini-3.8-flash": "promotional pricing through 2026-12-31: ai.google.dev/gemini-api/docs/pricing says it doubles to $1.50 in / $7.50 out starting 2027-01-01 (checked 2026-09-05)",
+  },
+  mistral: {
+    // Checked against mistral.ai/pricing/api on 2026-09-05: the table publishes ONLY `-latest` aliases
+    // (mistral-medium-latest $1.5/$7.5, mistral-small-latest $0.15/$0.6, mistral-large-latest $0.5/$1.5).
+    // No dated id appears — not the snapshot's `mistral-medium-2604`, and not `mistral-medium-3-5` or
+    // `ministral-8b-2512` either. The prices match, the ids are unverifiable, so nothing is renamed here.
+    "mistral-medium-latest": "mistral.ai/pricing/api publishes only `-latest` aliases and NO maximum output for any model — the catalog's maxOutput equalling the context window is filler, not a vendor figure (checked 2026-09-05)",
+    "mistral-medium-2604": "this dated id is not on mistral.ai/pricing/api, which lists only `-latest` aliases; the price matches mistral-medium-latest but the id itself is unverified (checked 2026-09-05)",
+  },
+  groq: {
+    // Not "unverified" — UNPUBLISHED. Groq's docs table says "Contact Sales" for the Llama models and
+    // console.groq.com/pricing is a 404, so any number here came from a third party, not from Groq.
+    "llama-3.3-70b-versatile": "Groq does not publish a price for this model — its docs table says \"Contact Sales\" and there is no public pricing page; the catalog's figure is models.dev's, not the vendor's (checked 2026-09-05)",
+  },
+  anthropic: {
+    // true of every Anthropic row; recorded on the one most runs use
+    "claude-mythos-5-1": "invite-only access (platform.claude.com/docs/en/models/mythos-5-1/overview); outside the Batches 300k beta, which covers Opus 5/4.8/4.7/4.6 and Sonnet 5/4.6 only (checked 2026-09-05)",
+    "claude-fable-5-1": "outside the Batches 300k beta, which covers Opus 5/4.8/4.7/4.6 and Sonnet 5/4.6 only (checked 2026-09-05)",
+    "claude-opus-5": "thinking tokens bill as OUTPUT and count inside max_tokens; from Opus 4.5 / Sonnet 4.6 / Fable 5.x earlier thinking blocks are preserved and bill as INPUT on the next turn (platform.claude.com/docs, checked 2026-09-05)",
+  },
+} as const;
+
+export const LOCAL_MODELS: Readonly<Record<string, Readonly<Record<string, LocalModel>>>> = {
+  deepseek: {
+    // api-docs.deepseek.com/quick_start/pricing, fetched 2026-09-05. These three rows OVERRIDE models.dev,
+    // which carries roughly a third of the vendor's own peak numbers for the same ids (flash 0.14/0.28 vs
+    // 0.44/1.32). The page is the seller; the snapshot is a mirror that has drifted, so the page wins here
+    // and says so. Peak rates are used: the page's own off-peak wording is self-contradictory (it labels
+    // "peak" 01:00-04:00 and 06:00-10:00 UTC, which are its cheap hours elsewhere), so rovecode quotes the
+    // HIGHER of the two — an estimate that surprises nobody with a bigger bill. Do not build discount logic
+    // on that wording until the page is unambiguous.
+    "deepseek-v4-flash": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, override: true,
+      cost: { input: 0.44, output: 1.32, cacheRead: 0.014 },
+      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
+    "deepseek-v4-pro": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, override: true,
+      cost: { input: 1.32, output: 3.96, cacheRead: 0.044 },
+      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
+    "deepseek-v4-flash-vision-exp": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, image: true, override: true,
+      cost: { input: 0.44, output: 1.32, cacheRead: 0.014 },
+      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
+    // The two API aliases. The pricing page lists only the v4-* ids, so nothing here is verified: the
+    // numbers are the last ones DeepSeek published for them and are marked as such rather than being
+    // quietly refreshed to a v4 price they may not be billed at.
+    "deepseek-chat": { context: 128_000, output: 8_000, reasoning: true, toolCall: true, cost: { input: 0.28, output: 0.42, cacheRead: 0.028 }, source: "api-docs.deepseek.com pricing (V3.2 era)", checked: "2026-09-05 (unverified: the alias is not on the pricing page, which lists deepseek-v4-* only)" },
+    "deepseek-reasoner": { context: 128_000, output: 64_000, reasoning: true, toolCall: true, cost: { input: 0.28, output: 0.42, cacheRead: 0.028 }, source: "api-docs.deepseek.com pricing (V3.2 era)", checked: "2026-09-05 (unverified: not on the pricing page; DeepSeek now exposes thinking as a MODE of the v4 models, not a separate id)" },
+  },
+  openai: {
+    // developers.openai.com/api/docs/pricing, fetched 2026-09-05. Not in the snapshot at all, so this is a
+    // gap fill, not an override.
+    "gpt-6-astra": { context: 1_050_000, output: 128_000, reasoning: true, toolCall: true, image: true,
+      cost: { input: 10, output: 50, cacheRead: 1 },
+      source: "developers.openai.com/api/docs/pricing", checked: "2026-09-05 (fetched)" },
+  },
+  xai: {
+    "grok-4": { ...GROK_43_REDIRECT, reasoning: true },
+    "grok-4-0709": { ...GROK_43_REDIRECT, reasoning: true },
+    "grok-4-fast": { ...GROK_43_REDIRECT, reasoning: true },
+    "grok-4-fast-reasoning": { ...GROK_43_REDIRECT, reasoning: true },
+    "grok-4-fast-non-reasoning": { ...GROK_43_REDIRECT, reasoning: false },
+    "grok-4-1-fast-reasoning": { ...GROK_43_REDIRECT, reasoning: true },
+    "grok-4-1-fast-non-reasoning": { ...GROK_43_REDIRECT, reasoning: false },
+    "grok-3": { ...GROK_43_REDIRECT, reasoning: false },
+    // grok-3-mini: neither on docs.x.ai/docs/models nor in the retirement list on 2026-09-04 — no row, stays unpriced
+  },
+};