AI-credits guard rejects claude-fable-5: pricing missing from curated table and stale bundled models.dev catalog

[raman-anyware]

Summary

With apiProxy.maxAiCredits active (the default), the api-proxy rejects every request for claude-fable-5 with HTTP 400 unknown_model_ai_credits, so any gh-aw workflow pinned to that model fails with 0 agent turns. The model is missing from both pricing sources the AI-credits guard consults, while the rest of the stack already knows the model.

Root cause

containers/api-proxy/guards/ai-credits-guard.js resolves pricing in this order: curated table (ai-credits-pricing.js) → canonical/prefix match → bundled catalog (models.dev.catalog.json) → AWF_DEFAULT_AI_CREDITS_PRICING → reject. For claude-fable-5, every step misses:

• containers/api-proxy/ai-credits-pricing.js — the Claude entries stop at claude-opus-4-8; no Fable/Mythos rows.
• containers/api-proxy/models.dev.catalog.json — vendored on 2026-06-09 (Bundle models.dev catalog for AI credits model resolution #4589) and not re-synced since; it predates the Claude Fable 5 release. The upstream models.dev catalog carries claude-fable-5 now.
• No prefix match is possible (claude-fable-* shares no prefix with any priced key).

Meanwhile the model is already recognized elsewhere:

• src/copilot-model.ts includes claude-fable-5 in the Copilot model allowlist (synced 2026-07-01, Sync Copilot model allowlist with model API mapping #5756).
• github/gh-aw's actions/setup/js/models.json gained claude-fable-5 pricing on 2026-07-03 (feat(models): 2026-07-03 model inventory update — new models, alias additions, mirror sync gh-aw#43063) — but that file feeds cost reporting only, not this guard.

So the firewall's vendored pricing snapshot is the last component in the stack that still rejects the model.

Impact

Workflows on engine: copilot with model: claude-fable-5 fail 100% at the first inference call unless the operator disables the credits budget (max-ai-credits: -1 frontmatter, or GH_AW_DEFAULT_MAX_AI_CREDITS=0) — which throws away cost enforcement precisely on the most expensive model.

Suggested fix

Either (ideally both):

1. Re-sync the vendored models.dev.catalog.json (it has been frozen for ~4 weeks; a periodic re-bundle would prevent this class of failure for every future model launch — same class as the Smoke Gemini gemini-3.1-flash-tts-preview failure, [aw-failures] P1: Smoke Gemini uses TTS preview model gemini-3.1-flash-tts-preview → unknown_model_ai_credits, 100% agent fail [Content truncated due to length] gh-aw#39172).
2. Add curated rows, matching Anthropic's published pricing ($/1M tokens):

'claude-fable-5':  { input: 10.00, cachedInput: 1.00, cacheWrite: 12.50, output: 50.00 },
'claude-mythos-5': { input: 10.00, cachedInput: 1.00, cacheWrite: 12.50, output: 50.00 },

(input $10, output $50, cache read 0.1x, cache write 1.25x — per https://platform.claude.com/docs/en/about-claude/models/overview)

Reproduction

Any gh-aw workflow with:

engine:
  id: copilot
  model: claude-fable-5

and maxAiCredits left at its default fails its first model call with:

HTTP 400 {"type":"unknown_model_ai_credits","message":"Model \"claude-fable-5\" has no AI credits pricing and no default pricing is configured. ..."}

Observed with AWF v0.27.2 through v0.27.26 (current); the pricing files on main today still lack the model.

🤖 Filed with the help of Claude Code