[MaxGhenis]

Claude Fable 5.2 launched on 2026-11-09. This pull request adds its serving card, list price and release date so the board can run it. It adds no result: the model joins the board in a later release.

What changes

• MODELS routes claude-fable-5.2 to claude-fable-5-2, with Anthropic's list prices ($8 / $40 per 1M tokens, $0.20 cache reads, $10 five-minute cache writes).
• The card uses the JSON contract, because the API rejects forced tool use with a 400, as on Fable 5.1. Its fields equal those of Fable 5.1, Opus 5.5 and Sonnet 5.5: JSON contract, no chunking, a 600-second timeout, the thinking budget and a 128,000-token ceiling.
• The card records two things a reader of the row needs. The board sends no reasoning-control parameter, so this row runs at the API's default effort, high, as Fable 5.1 and Sonnet 5.5 do, where Opus 5.5 and Haiku 5.5 default to medium. And the timeout is the other Claude JSON-contract rows' 600 seconds, where the onboarding gauntlet suggested the 300-second class default.
• MODEL_RELEASE_DATES gains the date in both the paper registry and the app, which a test keeps equal.

Sources

• Models API (created_at 2026-11-09), and platform.claude.com's Fable 5.2 overview, pricing page and effort documentation, all read 2026-11-09.
• policybench onboard parsed 3/3 and 16/16 on the JSON contract and estimated $0.143 per scenario.

Invariants and tests

• tests/test_model_cards.py locks the row's treatment (contract, chunking, timeout, budget) and now checks the thinking-budget field of the four Claude JSON cards.
• test_claude_fable_5_2_override_reconstructs_token_costs pins each listed rate against _reconstruct_token_cost.
• test_claude_fable_5_2_release_date_is_recorded pins the release date.

Review
An independent Claude Opus 5.5 review read the first two commits against the onboarding evidence and the peers' cards. Round 1 asked for one change, the timeout sentence, and five smaller ones; commit 3531192 makes them. Round 2 approves at 3531192. The session that carries this work moved to Claude Fable 5.2 on 2026-11-09, so a reviewer from another model reads every change that could affect this model's row.

Later, with the release
The board run's facts on the card, display names, the app roster and the note.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policybench-site	Ready	Preview	Nov 9, 2026 10:50pm UTC

[MaxGhenis]

Merge audit, head 3531192.

• Checks: gh pr checks exits 0 on this head (test, app, paper, lint, Vercel); GitHub reports MERGEABLE and CLEAN.
• Independent review: Claude Opus 5.5, read-only, against the onboarding evidence and the cards of Fable 5.1, Opus 5.5 and Sonnet 5.5. Round 1 asked for one change (the card's timeout sentence was false) and five smaller ones. Round 2 approves at 3531192: the fields equal the three Claude peers', the prices match Anthropic's pages, and the diff holds no result.
• Why another model reviewed it: the session carrying this release moved to Claude Fable 5.2 on 2026-11-09, so a reviewer from another model reads any change that could affect this model's row. The Codex lanes were closed, so the reviewer was Opus 5.5 where the default is GPT-6.1 Sol.
• Carried to the release: the board run's facts on the card (six households escalated past the 16,384-token starting budget; one call took 325 seconds, inside the card's 600 and past the 300-second class default), and a follow-up on the reasoning-token column, which reads zero on every Claude row.

🤖 Generated with Claude Code