Models: Sonnet 5.5 / Fable 5.1 defaults, direct-Anthropic-first lane rule

bin/agentic-runner.mjs

@@ -61,7 +61,7 @@ const buildId = args['build-id'] || '';
 // compete on specs for cents; a cheap judge ranks them) — so the implementation agent runs on
 // the low-cost tier. All-Fable implementation measured $7-9/page; tournament + Sonnet executor
 // targets $2.50-4.00 at the same gate bar. Both IDs verified live via GET /v1/models 2026-07-13.
-const model = args.model || 'claude-sonnet-5';
+const model = args.model || 'claude-sonnet-5-5';
 // Which endpoint + billing lane runs the executor (`claude -p`). Default: first-party Anthropic.
 // The credential preflight may FAIL THIS OVER to OpenRouter's Anthropic-compatible endpoint
 // (their "Anthropic Skin" — same executor model, billed to the owner's OpenRouter credits) when
@@ -271,7 +271,7 @@ async function probeCred(label, candidates, url, headers) {
   // completion (which also validates the model slug + credits in one shot, loudly).
   if (!a.ok && a.usageLimited) {
     const orCands = credCandidates('OPENROUTER_API_KEY');
-    const orModel = process.env.OPENROUTER_ANTHROPIC_MODEL || 'anthropic/claude-sonnet-5';
+    const orModel = process.env.OPENROUTER_ANTHROPIC_MODEL || 'anthropic/claude-sonnet-5.5';
     for (const cand of orCands) {
       try {
         const r = await fetch('https://openrouter.ai/api/v1/messages', {

bin/explainmyrepo.mjs

@@ -34,7 +34,7 @@ ARGUMENTS
 
 OPTIONS
   --out <dir>            build directory (default: ./explainer-builds/<repo>)
-  --model <id>           Anthropic model for the brain steps (default: claude-sonnet-5)
+  --model <id>           Anthropic model for the brain steps (default: claude-sonnet-5-5)
   --no-deploy            skip the deploy station (build + grade locally only)
   --no-publish           skip publish-repo + repo-seo (no GitHub explainer repo)
   --no-notify            skip the email notify station

src/claude.mjs

@@ -31,7 +31,9 @@ const ANTHROPIC_VERSION = '2023-06-01';
 // tonight's model-freshness sweep checked OpenAI's family and missed our own Anthropic default).
 // Sonnet is the authoring default: strong judgment for concept/content at a sane cost. Override
 // with --model or EXPLAINMYREPO_MODEL / ANTHROPIC_MODEL.
-export const DEFAULT_MODEL = 'claude-sonnet-5';
+// 2026-10-13: bumped sonnet-5 -> sonnet-5-5 (created 2026-09-28). Live 1-token probe returned HTTP 200 with
+// model echoed; the no-temperature rule above still applies.
+export const DEFAULT_MODEL = 'claude-sonnet-5-5';
 
 // ── THE AUTHORING MODEL (2026-08-08) ────────────────────────────────────────────────────────────
 // Measured, not assumed. A blind A/B on the REAL concept-authoring task (identical brief, labels
@@ -51,7 +53,11 @@ export const DEFAULT_MODEL = 'claude-sonnet-5';
 //
 // HONEST LIMIT: n=1 per model, one repo, one judge. The 15x cost gap survives that; the 95-vs-89
 // quality gap may not. Hence the automatic Anthropic fallback below rather than a hard switch.
-export const AUTHORING_MODEL = 'z-ai/glm-5.2';
+// 2026-10-13 re-measure (blind, 2 repos, 4 judge passes each, same concept prompt): glm-5.2 scored 79/61
+// (clearly behind), glm-5.3 scored 91/94 vs claude-sonnet-5-5 96/93 at near-identical $/concept
+// (glm spends its budget on hidden reasoning; 2-3x slower). So glm-5.3 replaces 5.2 as the OpenRouter
+// lane's model, but it is NOT the default whenever a direct Anthropic key exists (see resolveModel).
+export const AUTHORING_MODEL = 'z-ai/glm-5.3';
 const OPENROUTER_URL = 'https://openrouter.ai/api/v1/chat/completions';
 
 // An OpenRouter model id is namespaced ("z-ai/glm-5.2"); a first-party Anthropic id is not.
@@ -61,9 +67,14 @@ export function resolveModel(env = {}, override) {
   if (override) return override;
   if (env.EXPLAINMYREPO_MODEL) return env.EXPLAINMYREPO_MODEL;
   if (env.ANTHROPIC_MODEL) return env.ANTHROPIC_MODEL;
-  // Prefer the measured champion when an OpenRouter key exists; otherwise the Anthropic default.
-  // Set EXPLAINMYREPO_AUTHORING=anthropic to pin the old behaviour without removing the key.
-  if (env.OPENROUTER_API_KEY && env.EXPLAINMYREPO_AUTHORING !== 'anthropic') return AUTHORING_MODEL;
+  // Owner directive 2026-10-13: local runs bill the owner's own Anthropic developer account directly,
+  // never via OpenRouter when a direct key exists (a pass-through reseller is never cheaper than the
+  // source). The OpenRouter authoring champion is now used only when it is the ONLY credential
+  // (nothing else could run) or when explicitly opted into with EXPLAINMYREPO_AUTHORING=openrouter.
+  // EXPLAINMYREPO_AUTHORING=anthropic still pins Anthropic unconditionally.
+  const hasAnthropicKey = Boolean(env.ANTHROPIC_API_KEY || env.CLAUDE_API_KEY);
+  const a = env.EXPLAINMYREPO_AUTHORING;
+  if (env.OPENROUTER_API_KEY && a !== 'anthropic' && (a === 'openrouter' || !hasAnthropicKey)) return AUTHORING_MODEL;
   return DEFAULT_MODEL;
 }
 
@@ -301,7 +312,8 @@ async function callClaudeOnce({ apiKey, model, system, user, maxTokens, temperat
   const j = await resp.json();
   // Anthropic does not return a charge, so this lane is DERIVED from published per-token rates and
   // is marked as such in the receipt. Only the OpenRouter lane is a measured figure.
-  const RATES = { 'claude-sonnet-5': [2, 10], 'claude-opus-5': [5, 25], 'claude-haiku-4.5': [1, 5] };
+  // sonnet-5-5 rate is the published list price as reported by third-party sources, not read from a bill.
+  const RATES = { 'claude-sonnet-5-5': [2, 10], 'claude-sonnet-5': [2, 10], 'claude-opus-5-5': [4, 20], 'claude-fable-5-1': [10, 50], 'claude-opus-5': [5, 25], 'claude-haiku-5-5': [0.1, 0.5], 'claude-haiku-4.5': [1, 5] };
   const [rin, rout] = RATES[model] || [0, 0];
   recordSpend(model, ((j.usage?.input_tokens || 0) * rin + (j.usage?.output_tokens || 0) * rout) / 1e6,
     { in: j.usage?.input_tokens, out: j.usage?.output_tokens });

tests/brain-lane.test.mjs

@@ -75,7 +75,16 @@ import { resolveModel, isOpenRouterModel, AUTHORING_MODEL, DEFAULT_MODEL } from
 
 test('authoring — an OpenRouter key selects the measured champion', () => {
   assert.equal(resolveModel({ OPENROUTER_API_KEY: 'x' }), AUTHORING_MODEL);
-  assert.equal(AUTHORING_MODEL, 'z-ai/glm-5.2');
+  assert.equal(AUTHORING_MODEL, 'z-ai/glm-5.3');
+});
+
+test('authoring — a direct Anthropic key outranks OpenRouter (owner directive 2026-10-13)', () => {
+  assert.equal(resolveModel({ OPENROUTER_API_KEY: 'x', CLAUDE_API_KEY: 'k' }), DEFAULT_MODEL);
+  assert.equal(resolveModel({ OPENROUTER_API_KEY: 'x', ANTHROPIC_API_KEY: 'k' }), DEFAULT_MODEL);
+});
+
+test('authoring — EXPLAINMYREPO_AUTHORING=openrouter is the explicit opt-in even with an Anthropic key', () => {
+  assert.equal(resolveModel({ OPENROUTER_API_KEY: 'x', CLAUDE_API_KEY: 'k', EXPLAINMYREPO_AUTHORING: 'openrouter' }), AUTHORING_MODEL);
 });
 
 test('authoring — NO OpenRouter key leaves the old behaviour exactly as it was', () => {

tools/concept-tournament.mjs

@@ -110,8 +110,8 @@ function parseSpec(text, model) {
 }
 
 const CANDIDATES = [
-  { model: 'claude-sonnet-5', call: callAnthropic, needs: anthropicKey },
-  { model: 'claude-fable-5', call: callAnthropic, needs: anthropicKey },
+  { model: 'claude-sonnet-5-5', call: callAnthropic, needs: anthropicKey },
+  { model: 'claude-fable-5-1', call: callAnthropic, needs: anthropicKey },
   { model: 'gpt-5.6-sol', call: callOpenAI, needs: openaiKey },
 ];
 const JUDGE = { model: 'gpt-5.6-terra', call: callOpenAI, needs: openaiKey };