fix: price Fable 5.1 cache reads at 0.025x and keep Sonnet 5 at $2/$10

README.md

@@ -330,7 +330,7 @@ earns its cost (the knobs themselves:
 | Situation | Model | Effort | Why this model | Why this effort |
 | --------- | ----- | ------ | -------------- | --------------- |
 | Exploration (`scout`) | sonnet | low | Finding and tracing code is retrieval, not reasoning - a cheap tier reports as well as a costly one, and the file volume stays in the subagent regardless. | The work is mechanical lookup; extra thinking buys nothing. |
-| Ordinary implementation (`implementer`) | sonnet | medium | Sonnet is near-opus quality on single-file, clear-shape coding at a fraction of the price (intro pricing through 2026-08-31 makes it ~2.5x cheaper than opus) - for work whose approach the plan already decided, the margin never changes the outcome. | The plan already decided the approach; the agent executes real logic, not design. |
+| Ordinary implementation (`implementer`) | sonnet | medium | Sonnet is near-opus quality on single-file, clear-shape coding at a fraction of the price (permanently priced at $2/$10, ~2.5x cheaper than opus) - for work whose approach the plan already decided, the margin never changes the outcome. | The plan already decided the approach; the agent executes real logic, not design. |
 | Complex implementation (`implementer` `model=opus`) | opus | medium (pinned) | Multi-file or cross-layer changes, security/money/migrations/concurrency/public contracts, or a retry after a weak sonnet result - a wrong approach is expensive, and Opus 5 stepped the tier up at UNCHANGED price, so escalate when in doubt. (Ambiguous tasks are not an escalation case: implementer stops on ambiguity by contract - clarify first.) | The `model=opus` dispatch changes the model only - the Agent tool has no effort param, so the pinned medium stays; the escalation buys the tier, not extra thinking. |
 | Code review (`reviewer`) | opus | high | Review is an asymmetric bet - one pass guards against a bug that costs far more if it ships, so it is the one place to prefer the top tier by default. | High: subtle correctness bugs hide from shallow reading. Anthropic reports Opus 5 review holds at lower effort - medium is a future eval candidate; the pin stays high until measured. |
 | Tests / builds (`test-runner`) | haiku | low | Running a command and summarizing output is mechanical; the value is keeping raw logs out of the main context, not the model doing it. | Low: no reasoning, just report. |
@@ -387,7 +387,7 @@ exact numbers live on
 | ------ | ------------ | ------------------- |
 | fable | ~2x | Frontier reasoning. A main-session choice for all-hard-reasoning days (architecture, subtle debugging hunts) - never a routing target. |
 | opus | 1x | The escalation tier: code review, multi-file/cross-layer implementation, security/money/concurrency-sensitive changes. Opus 5 made this tier a step stronger at the same price. |
-| sonnet | ~0.6x (~0.4x intro through 2026-08-31) | The workhorse: ordinary implementation from an approved plan, exploration, E2E driving. Near-opus on clear-shape coding. |
+| sonnet | ~0.4x (permanent $2/$10) | The workhorse: ordinary implementation from an approved plan, exploration, E2E driving. Near-opus on clear-shape coding. |
 | haiku | ~0.2x | Mechanical grind: test/build runs, diff sanity checks, trivial sweeps. |
 
 Generation notes (as of the Opus 5 launch, July 2026 - these rot
@@ -629,7 +629,7 @@ Attribution comes from `agent-<id>.meta.json`, the sidecar Claude Code writes be
 `tokens` also prices the volume it just counted, because tokens need translating and dollars do not. Sample output - the window rolls daily, so your figures will differ:
 
 ```text
-At API list prices (rates as of 2026-08-11), this is what the volume above would have cost on the Claude API:
+At API list prices (rates as of 2026-09-09), this is what the volume above would have cost on the Claude API:
   as it ran                                             $2,380
   had every subagent inherited its session model        $4,417
   difference                                            $2,037
@@ -638,7 +638,7 @@ At API list prices (rates as of 2026-08-11), this is what the volume above would
 
 Read it as a counterfactual, which is exactly what it is: on a subscription you pay none of this, and the difference carries the same assumption as the volume figures above - that every subagent would otherwise have inherited the session model, which agents pinned by other plugins would not.
 
-The arithmetic is per token type rather than per token, because they are priced an order of magnitude apart: base input at the model rate, cache reads at 0.1x, five-minute cache writes at 1.25x, one-hour writes at 2x, and output at its own rate. Cache reads dominate real transcripts, so a flat volume multiply would overstate the bill several times over. Transcripts break cache writes down by TTL; a line carrying only a total is charged at the cheaper five-minute rate.
+The arithmetic is per token type rather than per token, because they are priced an order of magnitude apart: base input at the model rate, cache reads at 0.1x (0.025x on Fable 5.1 and Mythos 5.1), five-minute cache writes at 1.25x, one-hour writes at 2x, and output at its own rate. Cache reads dominate real transcripts, so a flat volume multiply would overstate the bill several times over. Transcripts break cache writes down by TTL; a line carrying only a total is charged at the cheaper five-minute rate.
 
 Several documented modifiers push these figures DOWNWARD, so treat them as conservative rather than optimistic. Models from Opus 4.7, Sonnet 5 and Fable 5 onward use a tokenizer producing roughly 30% more tokens for the same text than Sonnet 4.6 and earlier, so re-pricing a cheap model's token count at an expensive model's rate understates what that work would really have cost there. Nothing in a usage line reveals whether fast mode (double rates on Opus 5 and Opus 4.8) or US-only inference (1.1x from 4.6 onward) applied, so neither is added. And a model absent from the price table is excluded from every figure and declared on its own line, never counted as free - a window where nothing can be priced prints that declaration instead of figures, because silence would read as "this cost nothing".
 

hooks/dispatch-counter.mjs

@@ -282,44 +282,45 @@ const shortModel = (m) => m ? m.replace(/^claude-/, "").replace(/-\d{8}$/, "") :
 // (Opus 4.1 bills at three times Opus 4.5) and a loose pattern would quietly
 // misprice a retired model. A model absent from this table is reported as
 // unpriced volume, never as zero.
-const PRICES_ASOF = "2026-08-11";
-const SONNET5_STANDARD_FROM = Date.parse("2026-09-01T00:00:00Z");
+const PRICES_ASOF = "2026-09-09";
 const PRICES = [
   // Retired families first: a looser pattern below must not claim them.
   [/opus-4-1-|opus-4-20/, () => [15, 75]],
   // Pre-4.x ids put the generation first (claude-3-5-haiku-...), so both orders
   // are matched; the row above needs the same trick for claude-opus-4-20250514.
   [/3-5-haiku|haiku-3-5/, () => [0.8, 4]],
+  // 5.1 cache reads at 0.025x base input, ahead of the looser 5.x row below.
+  [/fable-5-1|mythos-5-1/, () => [10, 50, 0.025]],
   [/fable-5|mythos-5/, () => [10, 50]],
   [/opus-5|opus-4-8|opus-4-7|opus-4-6|opus-4-5/, () => [5, 25]],
-  // The one model on the page whose price changes on a date rather than with a
-  // new id: introductory $2/$10 through 2026-08-31, standard $3/$15 after.
-  [/sonnet-5/, (at) => (at < SONNET5_STANDARD_FROM ? [2, 10] : [3, 15])],
+  // Permanent $2/$10: the scheduled increase to $3/$15 on 2026-09-01 was
+  // called off.
+  [/sonnet-5/, () => [2, 10]],
   [/sonnet-4-6|sonnet-4-5|sonnet-4-20/, () => [3, 15]],
   [/haiku-4-5/, () => [1, 5]],
 ];
 // Prompt-caching multipliers, quoted from the same page: a 5-minute cache write
-// costs 1.25x base input, a 1-hour write 2x, and a cache read 0.1x. Transcripts
-// break cache writes down by TTL, so no averaging is needed.
+// costs 1.25x base input, a 1-hour write 2x. Cache reads are 0.1x base input
+// for every model except where the page states otherwise (Fable 5.1 and
+// Mythos 5.1 read at 0.025x - the row's third element overrides CACHE_READ).
 const CACHE_WRITE_5M = 1.25, CACHE_WRITE_1H = 2, CACHE_READ = 0.1;
 
 // Billable input volume: everything the model read, however it was cached.
 // Output is counted separately because it prices an order of magnitude higher.
 const volOf = (v) => v.in + v.cr + v.cw5 + v.cw1h;
 
 // Dollars for one model's token counts, or null when the model is not on the
-// price table. `at` is the instant the rates are taken from - the price of a
-// window, not of today, since one model's rate changes on a date inside the
-// horizon these reports can cover.
+// price table. `at` is kept as the hook for a future dated rate; no row on
+// the table uses it today.
 function costOf(model, v, at) {
   const row = model ? PRICES.find(([re]) => re.test(model)) : null;
   if (!row) return null;
-  const [inRate, outRate] = row[1](at);
+  const [inRate, outRate, crMul = CACHE_READ] = row[1](at);
   const perTok = inRate / 1e6;
   return (v.in * perTok)
     + (v.cw5 * perTok * CACHE_WRITE_5M)
     + (v.cw1h * perTok * CACHE_WRITE_1H)
-    + (v.cr * perTok * CACHE_READ)
+    + (v.cr * perTok * crMul)
     + (v.out * outRate / 1e6);
 }
 
@@ -647,10 +648,8 @@ if (process.argv[2] === "tokens") {
   const perAgent = new Map();
   let metaless = 0; // agent transcripts whose sidecar named no type
   let unknownAgents = 0, unknownVol = 0; // models tierOf cannot rank
-  // Cost accounting. The rate epoch is the END of the window, not "now", so a
-  // historical window is priced at the rates that applied to it - one model on
-  // the price table changes rate on a date that falls inside the horizon these
-  // windows can reach.
+  // Cost accounting. priceAt is kept as the hook for a future dated rate; no
+  // row on the table uses it today.
   const priceAt = win.end;
   let costRan = 0, costInherited = 0, unpricedVol = 0, unpricedSessionVol = 0;
   let mainCost = 0, mainUnpricedVol = 0;

hooks/dispatch-counter.test.mjs

@@ -1025,6 +1025,22 @@ test("cost prices every token type at its own documented rate", () => {
   } finally { rmSync(cfg, { recursive: true, force: true }); }
 });
 
+test("fable-5.1 cache reads price at 0.025x, sonnet-5 stays at $2/$10", () => {
+  const cfg = freshConfigDir();
+  const dir = join(cfg, "projects", "proj", "sess-1", "subagents");
+  mkdirSync(dir, { recursive: true });
+  writeFileSync(join(cfg, "projects", "proj", "sess-1.jsonl"), usageLine("claude-opus-5", 1) + "\n");
+  // Fable 5.1 cache read: 1M x 0.025 x $10/MTok = $0.25 (vs $1.00 at the 0.1x
+  // every other model gets). Sonnet 5: 1M base input x $2/MTok = $2.00, the
+  // permanent rate - the scheduled 2026-09-01 increase to $3/$15 was called off.
+  writeFileSync(join(dir, "agent-a.jsonl"), costLine("claude-fable-5-1", { cacheRead: 1e6 }) + "\n");
+  writeFileSync(join(dir, "agent-b.jsonl"), costLine("claude-sonnet-5", { input: 1e6 }) + "\n");
+  try {
+    const out = run(["tokens"], cfg);
+    assert.match(out, /as it ran\s+\$2\.25/);
+  } finally { rmSync(cfg, { recursive: true, force: true }); }
+});
+
 test("a cache-write bucket this code has never heard of is still charged", () => {
   const cfg = freshConfigDir();
   const dir = join(cfg, "projects", "proj", "sess-1", "subagents");

skills/model-routing/SKILL.md

@@ -135,9 +135,9 @@ actually earns its cost:
   which stays in the subagent regardless of tier.
 - **Ordinary implementation -> sonnet/medium.** Sonnet is near-opus
   quality on single-file, clear-shape coding at a fraction of the price
-  (as of the Opus 5 launch, July 2026: sonnet runs at intro pricing
-  through 2026-08-31, ~2.5x cheaper than opus). For work whose approach
-  the plan already decided, that margin does not change the outcome, so
+  (sonnet is permanently priced at $2/$10, ~2.5x cheaper than opus). For
+  work whose approach the plan already decided, that margin does not
+  change the outcome, so
   sonnet stays the value default. Medium effort because the agent
   executes, it does not design.
 - **Complex implementation -> opus, still at the agent's pinned medium**