Skip to content

chore(routing): route balanced-tier work to Claude Sonnet 5.5 - #268

Merged
pacphi merged 1 commit into
mainfrom
chore/model-pricing-2026-09-28
Sep 29, 2026
Merged

pacphi merged 1 commit into
mainfrom
chore/model-pricing-2026-09-28

Conversation

@pacphi

@pacphi pacphi commented Sep 29, 2026

Copy link
Copy Markdown
Owner

Summary

  • Claude Sonnet 5.5 (released 2026-09-28) keeps Sonnet 5's per-token price ($2/$10, standard 0.1x cache-read multiplier) but leads it on every first-party benchmark checked — GDPval-AA 1844 vs 1449, OSWorld 2.1 80.1% vs 57%, Terminal-Bench 4.0 70.6% vs 10.3%, CursorBench 4.0 55.5% vs 34.1%, Chartography 61.6% vs 15.6% — with lower tool-use system-prompt overhead (286 vs 354/474 tokens).
  • Swaps ak's balanced-tier default (specification, review, release) from claude-sonnet-5 to claude-sonnet-5-5, mirroring how Opus 5 → Opus 5.5 moved in feat(models): Opus 5.5 and GPT-6 Sol/Luna pricing and current-generation defaults #234: the new model becomes the catalog's balanced entry, claude-sonnet-5 demotes to tier: 'prior' and stays listed for user pins.
  • Scope check: no sonnet reference exists in qeCourt.mjs or any qe-court config, so QE-Court seat models (upstream/AQE-owned) and ruflo's internal router are unaffected. What changes is ak's own per-activity policy, consumed by ak run and projected into AQE's agentOverrides for the QE agents mapped to review/specification.
  • Full cited research write-up: docs/plans/2026-09-28-sonnet-5-5-routing-refresh.md (moves to docs/archive/ per repo convention once this merges).

Test plan

  • pnpm test (full suite, coverage-enforced) — green
  • pnpm run typecheck — green
  • pnpm run lint — green (no new warnings)
  • pnpm run lint:md — green
  • New pricing-row test (tests/kit/pricing.test.mjs) mirrors the existing Opus 5.5 own-key test
  • Updated hardcoded expectations in tests/kit/routing-primary.test.mjs and tests/kit/model-discovery-claude-codex.test.mjs that followed the default/catalog move

🤖 Generated with Claude Code

Sonnet 5.5 (released 2026-09-28) keeps Sonnet 5's per-token price ($2/$10)
but leads it on every first-party benchmark checked (GDPval-AA, OSWorld,
Terminal-Bench, CursorBench, Chartography) with lower tool-use overhead.
Swap the balanced-tier default (specification, review, release) the same
way Opus 5 -> Opus 5.5 moved in #234: new model becomes the catalog's
balanced entry, claude-sonnet-5 demotes to tier:'prior' for user pins.

- src/lib/pricing.mjs: own price row for claude-sonnet-5-5 (prevents future
  silent mis-pricing via the prefix matcher even though rates match today)
- src/lib/routing.mjs: MODEL_CATALOG + DEFAULT_ROUTES swap
- src/lib/model-inventory/discovery/anthropic-catalog.mjs: new catalog entry
- docs/providers.md, docs/adr/0006, README.md, src/commands/x/host.mjs:
  model-id references
- tests: new pricing-row test, updated hardcoded expectations that followed
  the default/catalog move

Full research + citations in docs/plans/2026-09-28-sonnet-5-5-routing-refresh.md.
qe-court seat models and ruflo's internal router are unaffected (upstream/
AQE-owned, not routed through this policy).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@pacphi
pacphi merged commit 3c6241c into main Sep 29, 2026
16 checks passed
@pacphi
pacphi deleted the chore/model-pricing-2026-09-28 branch September 29, 2026 04:06
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant