📋 Playbooks

How to survive AI model retirements on a BYOK platform

Updated 2026-10-05 · 5 min read · By the Convo360 team

In short

On a bring-your-own-keys platform, the models your bot depends on are retired by someone else on their schedule. A retired model id fails outright rather than degrading, so the bot goes silent. The playbook is to pin known model ids, watch the deprecations pages rather than the blogs, and keep a tested fallback ready.

Key takeaways

A retired model does not get worse. It stops answering, so the failure is total and arrives without a warning sign.
Provider blogs and deprecations pages are different pages. Only one of them predicts an outage.
Four providers retired or restricted models between June and October 2026, so this is routine, not exceptional.
Keep a second provider's key already configured. The cost of having it is nothing; the cost of not having it is a day of downtime.
Write down which agent uses which model. Most outages are extended by not knowing what is affected.

Why this is the price of bringing your own keys

Bringing your own AI keys is the right trade for most businesses. You pay OpenAI, Anthropic, Google or Sarvam directly at their rates, with no per message markup from the platform in between, and you keep the account and the data relationship.

The cost of that control is that you are a direct customer of a fast moving supplier. Model versions are retired on the provider's timetable, announced on the provider's changelog, and nobody is standing between you and that decision to absorb it.

The alternative is a platform that owns the keys and swaps models silently for you. That removes the admin and replaces it with a different problem: you do not know what model is answering your customers, you cannot pin a version you have tested, and the markup on every message pays for that convenience. Most businesses are better off keeping control and spending twenty minutes a quarter on this playbook.

What 2026 actually looked like

This is not a hypothetical risk. In the four months to October 2026, on 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, the last of those due for removal on 6 January 2027.
Anthropic deprecated Claude Sonnet 4.5 on 30 September, with retirement 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.
Sarvam retired bulbul:v2 earlier in the year, which broke every call running on that voice.

Four providers, four months, four separate things to notice. Assume roughly one event per provider per quarter and plan for that rhythm instead of treating each one as a surprise.

Know what you are running before you need to

The slowest part of a model outage is usually not fixing it. It is working out what is broken. A workspace with six agents, two of them half finished experiments, is hard to audit while customers are waiting.

So keep a short list, and keep it current. For each agent, note the provider, the exact model id, and the voice engine if it makes calls. Four columns in a spreadsheet is enough. The point is not documentation for its own sake, it is being able to answer the question in one minute when a deprecation notice names a model.

On Convo360 the provider and model are a per agent setting, so the screen is the list: open each agent and the model and voice are shown on it. Include dormant agents. A test agent that nobody opens still uses a real key, and will still fail loudly.

Watch the right page, not the blog

This is the single highest value habit, and almost nobody does it.

Every provider publishes company news and model retirements in different places. The blog carries launches, funding, customer stories and policy posts. The deprecations or release notes page carries the retirements. OpenAI's news feed, for example, does not list model removals at all, so following it gives false comfort: lots of updates, none of them the one that matters.

Subscribe to the narrow page for each provider you use. That is OpenAI's deprecations page, Anthropic's API release notes, the Gemini API changelog, and your voice provider's changelog. Four pages, checked once a week, is the whole discipline.

Convo360 does this automatically for platform owners. It reads the Meta, OpenAI, Anthropic and Google changelogs four times a day, emails immediately when an entry removes a feature or changes pricing, and holds everything else for one weekly digest. The design choice that matters is what it does not send, because an alert that fires constantly gets filtered into a folder and then the real one is missed too.

Keep a tested fallback already configured

Adding a second provider's key costs nothing until you use it, because you are billed by the provider on usage. Having one configured turns a migration from a procurement exercise into a dropdown change.

The useful pattern is one key with each of two providers, for example OpenAI and Google Gemini for chat, or Sarvam and ElevenLabs for Indian language voice. Then once a quarter, point one agent at the fallback, run three real conversations, and point it back. That is enough to know the fallback works, that the key is still valid, and that the replies are acceptable.

Do this before you need it. A key added under pressure is a key that fails validation at the worst moment, usually because the provider account has no credit on it. Our note on choosing a model for a support bot covers how to judge whether a fallback is good enough to ship.

The migration itself, in order

When a notice lands that names a model you use, work through it in this order and it stays a twenty minute job.

First, confirm the deadline and write it in a calendar, not a note. Second, identify every agent on that model, including dormant ones. Third, choose the nearest replacement rather than the newest or most capable, because a migration is not the time to also change tier. Fourth, switch one agent and run three real conversations, checking length, tone and pronunciation. Fifth, switch the rest. Sixth, if replies drift longer or less direct, tighten the system prompt instead of reverting, because reverting has an expiry date.

Doing this a month early costs the same twenty minutes as doing it in a panic, with the difference that nobody is messaging your WhatsApp number and getting silence while you work.

Step by step

List every agent, model and voiceRecord provider, exact model id and voice engine for each agent, including dormant test agents that still use live keys.
Subscribe to deprecations pages onlyFollow each provider's deprecations or release notes page, not its blog. Retirements are published separately from company news.
Configure a second provider keyAdd a key with a second provider now. It costs nothing unused and turns a future migration into a dropdown change.
Test the fallback quarterlyPoint one agent at the fallback, run three real conversations, then switch back. This proves the key and the quality.
On a notice, diary the deadlinePut the retirement date in a calendar with a reminder two weeks before, rather than leaving it in a notes app.
Migrate to the nearest modelChoose the closest replacement rather than the newest. Changing capability tier during a migration hides the cause of any change.
Fix drift with the promptIf the new model writes longer or differently, tighten the system prompt instead of reverting to a model that is expiring.

Example conversation

We only use one AI provider. Is adding a second key really worth the admin?

An unused key costs nothing, because providers bill on usage. What it buys is the difference between a dropdown change and sourcing a new account during an outage, which is usually the longest part of the recovery.

Keep control of your keys with Convo360

Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.

Start 14-day free trial →See pricing

Frequently asked questions

How often do AI models get retired?

Roughly once a quarter per provider, based on 2026. Between June and October, OpenAI, Anthropic, Google and Sarvam each retired or restricted models that bots in India were actively using.

Why does a retired model break the bot completely?

The model id is part of every request. Once the provider stops serving that id the request is rejected rather than downgraded, so every conversation fails at once instead of getting slightly worse.

Is bringing my own keys more risky than a managed platform?

It exposes you to provider changes, but a managed platform hides which model answers your customers and charges a markup for that. Most businesses prefer control plus this playbook.

Which pages should I actually follow?

The deprecations page for OpenAI, Anthropic's API release notes, the Gemini API changelog, and your voice provider's changelog. Four narrow pages beat four company blogs, which rarely list retirements.

Can Convo360 tell me when a model is retired?

Yes. Convo360 reads the Meta, OpenAI, Anthropic and Google changelogs four times a day and emails the owner immediately for anything that removes a feature or changes pricing, with the rest in a weekly digest.