Summary

models.aliases.opus is now set to anthropic:claude-opus-4-8 (released 2026-05-28). But src/core/anthropic-pricing.ts (ANTHROPIC_PRICING) only has entries for claude-opus-4-7 and claude-opus-4-6 — there is no claude-opus-4-8 row.

Impact

estimateMaxCostUsd() returns null for any model not in the table. When pricing is null, the skillopt cost preflight has no estimate to compare against --max-cost-usd, so the hard cost cap silently no-ops on 4.8 runs. A run that should refuse at the cap will instead proceed unbounded.

This also affects the dream-cycle budget meter (same pricing source) — non-priced models hit the BUDGET_METER_NO_PRICING warn path and bypass the budget gate.

Fix

Add to ANTHROPIC_PRICING in src/core/anthropic-pricing.ts:

'claude-opus-4-8': { input: 5.00, output: 25.00 },

Same per-token pricing as 4-7 ($5 in / $25 out per 1M tokens), confirmed against Anthropic's published pricing.

Repro

1. gbrain config set models.aliases.opus anthropic:claude-opus-4-8
2. gbrain skillopt <skill> --bootstrap-from-skill --dry-run
3. Cost estimate line shows no refusal even when projected spend exceeds --max-cost-usd.

Notes

Found while upgrading all gbrain tiers from Opus 4.7 → 4.8. Config flip done via models.aliases.opus override; this pricing gap is the one residual.