feat: refresh Foundry model catalog

[anandpant]

Summary

Bring the catalog in line with foundry-cli models list --json as of September 24, 2026.

• Add GPT-6 Astra (GA), GPT-6 Luna, GPT-6 Sol, Codex Auto Review, text-embedding-ada-002, Claude Opus 5.5, Grok 4.7, and DeepSeek V4.1 Flash. Promote gemini-3.8-flash from Experimental to GA. Remove claude-opus-4.1, now Sunset with a 2027-01-07 retirement, and add it to the removed-alias list.
• Four listed candidates were probed and left out. GPT-5.2 Pro and GPT-5.4 Pro return DisabledForUser on Responses with no Chat Completions route; o1 (Azure OpenAI) and Schematic 7B (Palantir Hub) return LanguageModelNotAvailable because their backends serve neither OPEN_AI_RESPONSES nor GPT_CHAT_COMPLETION. They are recorded under docs Findings rather than dropped silently.
• DeepSeek V4.1 Flash is cataloged as third-party with the verified openai-chat transport. Native @ai-sdk/deepseek exposes reasoning text but breaks tool continuation at the proxy, so no optional peer was added; the comparison lands with the results docs.
• inputTypes follow each family's siblings and were confirmed by probes: the GPT-6 trio matches gpt-5.6-*, Codex Auto Review is Responses-only because the Chat Completions route rejects it, and Grok 4.7 matches its siblings.
• ModelPerformance.cost becomes optional alongside speed, and ModelClass gains SPECIALIZED_EMBEDDING, so the catalog can mirror what the listing publishes instead of inventing values. Grok 4.7 now omits cost, and the embeddings carry their real SPECIALIZED_EMBEDDING class.
• Re-diffed all 69 entries against the listing and corrected 28 stale fields on pre-existing models: Claude Opus 4.5-4.8 cost, Opus 4.7/4.8, Opus 5 and Sonnet 5 cutoffs, GPT-5.5 cutoff, GPT-5.3 Codex class, Gemini 3.5 Flash cost/class/URL, Gemini 3.5 Flash-Lite display name, embedding costs and URLs, and realtime cutoffs and URLs. The diff is now zero.
• The CLI catalog fixture now carries displayName, trainingCutoffDate and performance as well, and the enrollment test checks them, so future metadata drift fails instead of passing silently. The harness embedding probe list derives from the catalog so new embeddings are surveyed automatically. Opus 5.5 joins the adaptive-thinking set; the proxy rejects budget-based thinking for it with HTTP 400.

Final catalog: 69 entries — 63 language aliases, three embeddings, three realtime.

Validation

pnpm run verify passed, covering release policy, lint, tests, typecheck, build, skill validation, and the package audit. Each candidate was probed against its raw RID before cataloging, and the eight additions then passed a focused live harness run. The new drift assertions were confirmed load-bearing by perturbing a display name and watching the enrollment test fail. No package version or changelog changes.

[anandpant]

• docs: publish September 24 live results #79
• fix: keep native structured output for Claude Opus 5.5 #78
• feat: refresh Foundry model catalog #77 👈 (View in Graphite)
• main

---

How to use the Graphite Merge Queue

Add either label to this PR to merge it via the merge queue:

• merge - adds this PR to the back of the merge queue
• fast - for urgent changes, fast-track this PR to the front of the merge queue

You must have a Graphite account in order to use the merge queue. Sign up using this link.

An organization admin has enabled the Graphite Merge Queue in this repository.

Please do not merge from GitHub as this will restart CI on PRs being processed by the merge queue.

This stack of pull requests is managed by Graphite. Learn more about stacking.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[graphite-app]

Merge activity

• Sep 25, 5:48 AM UTC: anandpant added this pull request to the Graphite merge queue.
• Sep 25, 5:50 AM UTC: Merged by the Graphite merge queue.