How to know when your AI provider changes something
Watch one technical page per provider, not their blogs. Retirements and price changes appear on deprecations pages and API release notes, while company news and customer stories fill the blogs. Four pages checked weekly covers Meta, OpenAI, Anthropic and Google, which is where the changes that break a bot come from.
- Provider blogs and deprecations pages are different pages, and only one of them predicts an outage.
- OpenAI's news feed carries no model retirements at all, so following it gives false comfort.
- Four providers retired, removed or restricted models between June and October 2026.
- Price changes matter as much as retirements. Meta began charging for WhatsApp service replies from 1 October 2026.
- An alert that fires constantly gets filtered into a folder, so filter for removals and pricing and ignore the rest.
The outage that has no warning sign
Most things that break a WhatsApp bot or an AI calling setup give you a hint first. Messages slow down, a few calls fail, an error rate creeps up.
A model retirement does not. One day the model id is served and the next it is not, and every request fails at once. There is no gradual degradation because the model id is part of the request: once the provider stops recognising it, the call is rejected rather than downgraded to something newer.
For a bot, that failure is total. Every incoming message needs a model reply, so if the model call fails, every conversation fails. A customer messages your number and gets silence, which from the outside is indistinguishable from the business having abandoned the channel.
The warning did exist. It was published on a changelog, weeks or months earlier, on a page almost nobody subscribes to.
This is routine, not rare
In the four months to October 2026, across the providers an Indian WhatsApp or voice bot typically runs on:
- OpenAI deprecated chat models in June, legacy audio and realtime models in July, transcription models in August, and text to speech models in October with removal set for 6 January 2027.
- Anthropic deprecated Claude Sonnet 4.5 on 30 September, retiring it from the API on 30 November 2026.
- Google limited access to its Gemini 2.5 models in September to accounts that had actively used them.
- Meta began charging for WhatsApp service replies from 1 October 2026 beyond a free monthly allowance per number.
- Sarvam retired bulbul:v2 earlier in the year, which broke every call running on that voice.
That is roughly one event per provider per quarter. Plan for that rhythm rather than treating each one as an unlucky surprise.
Read the technical page, not the blog
This is the single highest value habit in this article, and it costs nothing.
Every provider publishes company news and breaking changes in different places. The blog has launches, funding, research, customer stories and policy posts. The deprecations page or API release notes carry the retirements and the access changes.
OpenAI is the clearest example. Its news feed is full of material, almost none of it operational, and it does not list model retirements. A business following that feed diligently would have seen dozens of posts and missed every deprecation. The deprecations page carries them all.
So the list to follow is short and specific: OpenAI's deprecations page, Anthropic's API release notes, the Gemini API changelog, Meta's developer blog and Graph API changelog, and your voice provider's changelog. Four or five narrow pages, once a week, is the entire discipline.
Decide what deserves to interrupt you
The failure mode of monitoring is not missing things. It is noticing too many, which trains you to ignore the channel entirely.
Three categories are worth separating. Removals, meaning anything deprecated, retired, discontinued or requiring migration, should reach you immediately, because they have a deadline and a consequence. Price changes, including a free allowance being withdrawn, should also reach you immediately, because they change your unit economics. Everything else, including new models and features, can wait for a weekly summary.
Be strict about the second category in particular. The word pricing appears in the footer and sidebar of almost every documentation page, so a naive keyword filter flags feature launches as price changes and quickly becomes noise. Filter on phrases that mean money actually moved, such as will be charged, free tier or per message.
An alert that cries wolf is worse than no alert, because when the real one arrives it is already in a folder you stopped opening.
Know what you are running
Monitoring only helps if you can act on it quickly, and the slow part of a migration is almost never the switch. It is working out which of your agents is affected.
Keep a short record: for each agent, the provider, the exact model id, and the voice engine if it makes calls. Four columns is enough. The goal is to answer in one minute when a notice names a model.
Include dormant agents. A test agent nobody opens and a seasonal campaign agent used twice a year both use live keys and will both fail loudly. The Gemini 2.5 access restriction was aimed squarely at this profile, since access depended on recent usage.
On Convo360 the provider, model and voice are per agent settings, so the agent screen is the record. That also means a workspace can deliberately run different agents on different providers, which is itself a hedge.
Or have it watched for you
If reading four changelogs weekly is not going to happen, automate the narrow version of it.
Convo360 reads the official Meta, OpenAI, Anthropic and Google changelogs four times a day. Anything that removes a feature or changes pricing is emailed to the platform owner straight away. Everything else arrives as a single weekly digest, and pure company news is dropped rather than stored, because a card full of customer success stories hides the deprecation underneath it.
That design exists because of a real outage. When Sarvam retired bulbul:v2, every Sarvam call failed and the businesses affected found out from callers rather than from the announcement, which had been public for weeks. The feature is a direct response to that.
Whether you automate it or do it by hand, the principle is the same: watch the page that predicts outages, filter hard for removals and pricing, and keep a list of what you run. Our playbook for surviving model retirements covers the migration itself once a notice arrives.
Step by step
- Bookmark one page per providerOpenAI deprecations, Anthropic API release notes, Gemini API changelog, Meta's changelog and your voice provider's.
- Unsubscribe from the blogsCompany news crowds out operational notices. Follow the technical page instead of the marketing one.
- Define what interrupts youRemovals and price changes reach you immediately. New features and launches wait for a weekly summary.
- List every agent and model idRecord provider, model id and voice engine per agent, including dormant and seasonal agents.
- Diary every deadline you seePut retirement dates in a calendar with a reminder two weeks ahead, not in a notes app.
- Keep a fallback key readyConfigure a second provider now and test it quarterly, since an unused key costs nothing.
Example conversation
Get told when a provider breaks something
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
Which pages should I actually follow?
OpenAI's deprecations page, Anthropic's API release notes, the Gemini API changelog, Meta's developer and Graph API changelogs, and your voice provider's changelog. Narrow technical pages, not company blogs.
Why is a retired model so disruptive?
The model id is part of every request. Once the provider stops serving it, calls are rejected rather than downgraded, so every conversation fails at once with no gradual warning beforehand.
How often does this happen?
About once per provider per quarter. Between June and October 2026, OpenAI, Anthropic, Google, Meta and Sarvam each made a change that affected bots running in India.
Should price changes alert me too?
Yes. Meta began charging for WhatsApp service replies from 1 October 2026 beyond a free monthly allowance, which changes unit economics even though nothing technically broke.
Does Convo360 monitor this automatically?
Yes. It reads the Meta, OpenAI, Anthropic and Google changelogs four times a day, emails the owner immediately for removals and price changes, and holds everything else for a weekly digest.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.