feat: add current-generation models across providers (v0.5.0)

[ericrallen]

Registry grows 73 → 101 models across 18 providers, and both packages bump to 0.5.0.

New models (28)

Provider	Models
Anthropic	Opus 5, Sonnet 5, Fable 5.1, Fable 5, Opus 4.8
OpenAI	GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna
Z.AI	GLM-5.3, GLM-5.3 Flash, GLM-5.2
Moonshot	Kimi K3
Qwen	3.8 Max, 3.8 Flash, 3.8 27B, 3.8 2.4T-A95B
Google	Gemini 3.8 Flash, 3.7 Flash, 3.5 Flash Lite
xAI	Grok 4.6, 4.5, 4.3
DeepSeek	V4.1 Flash, V4 Pro 0813
Mistral	Medium 3.5
NVIDIA	Nemotron 3.5 Lightning, Nemotron 3 Ultra 550B-A55B

Specs come from the live OpenRouter model API cross-checked against first-party provider docs. Where an aggregator prices a model differently from the primary provider, that is recorded as a provider-level cost_per_million_tokens override rather than flattened into one number (see gpt-5.6-sol, where OpenAI's published rate and OpenRouter's differ).

Data corrections

• gpt-5.5, gpt-5.5-pro — max_tokens was 1049999 (the context window minus one) rather than the real 128000 output cap. Added the missing knowledge_cutoff, plus cached_input for gpt-5.5. Verified against OpenAI docs; the gpt-5.4 entries already had this right.
• minimax-m3 — context_window: 512000 contradicted the entry's own description. The model serves 1048576.
• qwen3.7-plus — model_name read "Qwen 3.6 Plus".
• claude-opus-4-6-fast — marked deprecated. Fast mode is no longer offered on Opus 4.6 or 4.7; it is now limited to Opus 5 and Opus 4.8.

On deprecations

Superseded-but-still-served models are deliberately not marked deprecated. The schema defines that field as deprecation by the provider, and every other entry was verified as still being served. Marking working models would make downstream consumers hide them.

Package maintenance

• packages/typescript/src/index.ts — the hand-maintained named-export list had drifted (72 exports for 73 models). Regenerated from models/*.yaml, after confirming the naming rule reproduces all 72 existing entries byte-exact.
• packages/python/.../__init__.py — added the new models to the curated constants list. It remains a curated subset rather than a full mirror, as before.

Verification

All four validate.yml jobs were run locally before pushing, plus two extra checks:

• validate:models / validate:providers
• codegen drift (git diff --exit-code on generated/)
• tsc --noEmit
• build_registry.py + pydantic check
• all 101 models validate against the generated pydantic ModelMetadata
• Python and TypeScript registry.json are semantically identical

Release

After merge, publishing is triggered by tags: typescript-v0.5.0 → npm, python-v0.5.0 → PyPI.

🤖 Generated with Claude Code