chore(ai): refresh the generated model catalog

[snimu]

What

Refreshes the checked-in models.generated.ts with the current output of npm run generate-models. Pure generated output, nothing hand-edited: 92 models added, 33 removed (1163 → 1222) across 14 providers.

Why

The build regenerates this file from the live provider catalogs on every CI run anyway — but the committed copy had drifted ~2,300 lines behind reality. Refreshing it keeps local keyless test runs representative of what CI actually tests, and makes future drift diffs reviewable instead of monolithic.

Highlights

Added

• DeepSeek V4 Flash 0731 / Pro 0813 dated snapshots (cloudflare-workers-ai, fireworks, huggingface)
• gemini-3.7-flash (google, github-copilot), gemini-3.6-flash (copilot)
• grok-4.5 / grok-4.6 and kimi-k3 on github-copilot
• glm-5.3 on opencode-go, qwen3.6-27b on groq, qwen3p8-max on fireworks
• GPT-5.x family fan-out on cloudflare-ai-gateway (gpt-5 through gpt-5.6, codex-spark)
• 10 new huggingface routes (Qwen2.5/3 coder + DeepSeek V3/V4 families)

Removed

• claude-opus-4-1 aliases retired everywhere (anthropic, opencode, vercel, cloudflare)
• Cloudflare gateway dropped its claude-3.x routes
• gemini-2.0-flash/-lite and gemini-3-pro-preview retired on google
• ling-3.0-flash free routes gone (openrouter, opencode, vercel)

Validation

• Full keyless packages/ai suite against the regenerated catalog: 327 passed / 731 skipped, 0 failed — so no supports-xhigh-style expectation drift is hiding in this refresh
• npm run build reproduces the file byte-identically; npm run check clean

---

Note

Medium Risk
Removed model IDs can break saved configs or tests that still reference them; widespread pricing and context-limit changes affect cost estimates and truncation behavior without runtime code changes.

Overview
Regenerates models.generated.ts from the current provider catalogs (~92 models added, ~33 removed), bringing the committed file in line with what npm run generate-models emits on build.

Retirements include claude-opus-4-1 (and dated alias) across Anthropic and gateway providers, legacy Claude 3.x routes on Cloudflare AI Gateway, gemini-2.0-flash / -lite and gemini-3-pro-preview on Google, and several free-tier routes (e.g. Ling 3.0 flash on OpenRouter/OpenCode/Vercel).

Additions and updates span new dated DeepSeek V4 snapshots, gemini-3.7-flash, Grok 4.5/4.6, Kimi K3, GLM 5.3, Qwen 3.8 family entries, expanded GPT-5.x and Workers AI routes on Cloudflare Gateway, and Hugging Face / Fireworks / OpenRouter listings. Many existing entries only change context windows, token limits, cache pricing, thinkingLevelMap (often adding "max":null for DeepSeek), display names, or API kind (e.g. OpenCode Go Qwen3.7 models moved from anthropic-messages to openai-completions).

^{Reviewed by Cursor Bugbot for commit 616463e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Refresh generated model catalog with new models and updated pricing

Regenerates models.generated.ts to reflect the current state of supported AI models across all providers.

• Adds new model entries including GPT-4.1/GPT-5 variants, Gemini 3.x Flash, Grok updates, DeepSeek dated releases, and Qwen 3.x variants
• Updates context window sizes and token limits for several models (some increasing up to 1,050,000 tokens)
• Adjusts pricing fields across many models, including minor rounding corrections and larger cost changes
• Removes or renames some existing entries (e.g. dropping (latest) suffixes, removing some Claude 3-series gateway entries)
• Behavioral Change: models with updated contextWindow, maxTokens, or cost fields will reflect new values immediately for any downstream consumers

^{Macroscope summarized 616463e.}