feat(models): add GPT-5.6 (Sol/Terra/Luna) and Claude Fable 5 to engine dropdowns

[aaight]

Summary

Purely additive model-catalog change: adds the most recent SOTA models from OpenAI and Anthropic to the engine dropdowns. No default changes, no DB/schema/frontend edits.

• OpenAI GPT-5.6 family (gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna) → Codex engine dropdown.
• Anthropic Claude Fable 5 (claude-fable-5) → Claude Code engine dropdown.

Mirrors the 5-file precedent from PR #1199 ("add Opus 4.7, GPT-5.5, and 1M context variants").

Issue: https://linear.app/issue/MNG-1751
(Linear: https://linear.app/mongrel/issue/MNG-1751/lets-research-and-add-most-recent-sota-models-from-openai-and)

Changes

File	Change
src/backends/codex/models.ts	Prepend gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna to CODEX_MODELS (6 → 9 entries). CODEX_MODEL_IDS derives automatically.
src/backends/claude-code/models.ts	Prepend claude-fable-5 to CLAUDE_CODE_MODELS (10 → 11 entries).
src/utils/llmMetrics.ts	Add openai:gpt-5.6-* pricing rows (CI-enforced by llmMetrics-codex.test.ts + required by resolveCodexModel) and the anthropic:claude-fable-5 row.
src/config/rateLimits.ts	Add an anthropic:claude-fable-5 rate-limit profile mirroring the Opus tier-1 throttle.
tests/unit/backends/claude-code.test.ts	Bump CLAUDE_CODE_MODELS length 10 → 11; add Fable 5 toContain + resolveClaudeModel pass-through assertions.
tests/unit/backends/codex.test.ts	Add resolveCodexModel pass-through assertions for the GPT-5.6 tiers (bare + openai:-prefixed).

Pricing (confirmed against official sources, per the researched plan)

OpenAI GPT-5.6 (per-1M USD):

Model	Input	Cached	Output
gpt-5.6-sol	5.00	0.50	30.00
gpt-5.6-terra	2.50	0.25	15.00
gpt-5.6-luna	1.00	0.10	6.00

Anthropic Claude Fable 5 (per-1M USD): input 10.00 / cached 1.00 / output 50.00 (2× Opus). The pricing key anthropic:claude-fable-5 equals toPricingKey('claude-fable-5') (no trailing 8-digit date to strip), so cost is not silently zero.

Design decisions

• No GPT-5.6 Codex-specific model — OpenAI ships the Sol/Terra/Luna tiers; Codex runs the same three, so no -codex/-mini variant is added.
• No claude-fable-5[1m] variant — Fable 5 is 1M-context by default (max = default), so a [1m] suffix would be redundant (unlike Opus/Sonnet). A test pins not.toContain('claude-fable-5[1m]').
• No claude-mythos-5 — Project Glasswing invitation-only; not selectable by most orgs.
• Defaults unchanged — DEFAULT_CODEX_MODEL = 'gpt-5.4', DEFAULT_CLAUDE_CODE_MODEL = 'claude-sonnet-4-5-20250929' — keeps the change additive and avoids catalog.ts defaultValueLabel edits.
• No Codex rate-limit row — precedent omits it (the Codex CLI self-throttles; unmatched models fall back to { enabled: false }).
• Out of scope untouched — CUSTOM_MODELS (OpenRouter/llmist) and the frontend (dropdowns are served at runtime via trpc.agentConfigs.engines → getEngineCatalog()).

Operator note (Fable 5 deployment)

Fable 5 requires ≥30-day data retention — a project on a zero-data-retention (ZDR) org will get 400 invalid_request_error on every request. Its safety classifiers can also refuse (stop_reason: "refusal"). These are operator/config concerns, not code bugs; the additive change leaves defaults unchanged.

Testing

• npx vitest run --project unit-backends tests/unit/backends/claude-code.test.ts tests/unit/backends/codex.test.ts tests/unit/backends/codex-cost.test.ts → 181 passed
• npx vitest run --project unit-core tests/unit/utils/llmMetrics-codex.test.ts tests/unit/utils/llmMetrics.test.ts tests/unit/config/customModels.test.ts → 46 passed (the pricing-coverage loop now iterates 9 Codex models, covering the 3 new tiers)
• npm run typecheck → clean
• biome check on changed files → clean

🤖 Generated with Claude Code

🕵️ claude-code · claude-opus-4-8 · run details

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[nhopeatall]

Summary

APPROVE — clean, purely-additive model-catalog change that faithfully implements the confirmed research plan in MNG-1751 and mirrors the 5-file precedent from PR #1199. Verified against the work item's ground-truth pricing/ID tables, not just the PR description.

Verification performed

• Pricing values match the work item exactly — openai:gpt-5.6-sol (5/0.5/30), -terra (2.5/0.25/15), -luna (1/0.1/6), anthropic:claude-fable-5 (10/1/50 = 2× Opus 4.8). All four rows agree with the confirmed research table.
• No silent zero-cost trap — confirmed toPricingKey('claude-fable-5') → anthropic:claude-fable-5 (the /-\d{8}$/ strip is a no-op since there's no trailing date), so the Claude cost key resolves. Codex side is additionally CI-guarded by llmMetrics-codex.test.ts, which now iterates 9 models.
• Resolver pass-through is correct — resolveCodexModel accepts the 3 new IDs bare and openai:-prefixed (they're in CODEX_MODEL_IDS); resolveClaudeModel passes claude-fable-5 through via the CLAUDE_CODE_MODEL_IDS/claude- paths.
• Rate-limit key shape — anthropic:claude-fable-5 matches the exact-match lookup in getRateLimitForModel; consistent with existing Anthropic rows.
• No collateral breakage — catalog.ts derives options directly from the arrays and its defaultValueLabel strings are correctly left untouched (defaults unchanged). catalog.test.ts asserts options.length > 0, not an exact count, so additive entries are safe.
• Decisions honored — no -codex GPT-5.6 variant, no redundant claude-fable-5[1m] (pinned by not.toContain), no invitation-only claude-mythos-5, defaults unchanged.
• Tests green locally — claude-code.test.ts + codex.test.ts (172), llmMetrics(-codex) + rateLimits (33). CI 7/7 passing.

No blocking or should-fix issues found.

🕵️ claude-code · claude-opus-4-8 · run details