Discovery via each provider's /v1/models endpoint (2026-07-23) turned up:

Renames (existing entries pointing at deprecated model strings):
- deepseek-chat → deepseek-v4-flash (aliases retire 2026-07-24)
- deepseek-reasoner → deepseek-v4-pro
- glm-4.7-flash → glm-4.7 (the -flash variant was removed by Z.ai)
  Registry ids kept where possible for consumer stability; only the
  `model:` string and matching `--model` flag change. glm registry id
  itself renamed since -flash was in the identifier.

Added (live-tested healthy or paid-plan-required on this account):
- groq-llama-3.1-8b-instant — Groq's fastest tier (weak/fast/free).
- openrouter-nemotron-super-free — Nvidia 120B :free, lighter tier
  companion to the 550B ultra already bundled.
- cerebras-gemma-4-31b — Google Gemma on Cerebras (~2000 tok/s).

Skipped despite being new on /v1/models:
- Gemini `-latest` aliases + preview variants (moving targets)
- Groq gpt-oss-safeguard-20b (safety-tuned variant, marginal utility)
- Z.ai glm-5/5.1/5.2 (paid tier, no free access verified)
- OpenRouter poolside/laguna family (unknown quality)
- OpenRouter nvidia/nemotron nano/content-safety variants (redundant)

Curation over shotgun — bundled registry stays tight.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>