Ask any team that runs an AI production pipeline how they found out a model they depend on was deprecated, and the answer is almost always the same: from the errors.
A model gets sunset, an endpoint changes, a pricing tier quietly disappears — and your app only breaks when traffic actually hits the affected path. Then it's two days of paging, reverting, re-testing prompts against a new model, and explaining to stakeholders why "it worked yesterday."
The worst part: every provider publishes this information. Nobody reads it. OpenAI has a deprecations page. Anthropic has one. DeepSeek and Gemini have one. They just sit there, changing silently while your pipeline depends on them.
What I built: a tiny watcher for model deprecation pages
A single-file Python CLI (pure standard library, zero dependencies) that watches the official deprecation/change pages of the major providers:
- 4 sources out of the box — OpenAI deprecations, Anthropic model-deprecations, DeepSeek news, Gemini deprecations
-
--addcustom sources — your own vendor, your own SaaS changelog, any page you depend on - Fingerprint diffing — it hashes the visible text, so it only alerts on real content changes, not on layout tweaks or popup modals
- Exit code 1 on change — plug it into cron/CI/webhook and you get alerted the day a deprecation lands, not the day your pipeline breaks
- Zero API keys, zero cost — it reads public pages only
# Build a baseline once
python model_watch.py --snapshot
# Run daily in cron; exit code 1 = something changed
0 9 * * * python /opt/model_watch/model_watch.py --json >> /var/log/model_watch.log
When it fires, it extracts the actual deprecation/retirement sentences from the page, so you know what changed, not just that something changed.
Why a watcher beats a changelog subscription
Email digests from vendors are marketing-shaped and late. A diff-based watcher on the official page is the ground truth, checked on your schedule, alerting in the format your ops stack already understands (exit codes).
If you run any multi-provider AI pipeline — or just want to stop learning about model changes from your error logs — the full kit (watcher + custom-source support + JSON output + docs) is at AgentChip.
Deprecation notices are public. Being caught by surprise is optional.
Originally published on the AgentChip blog.