Google has limited Gemini 2.5 access to active users
On 18 September 2026 Google limited access to its Gemini 2.5 models to accounts that had actively used them, to keep performance reliable for everyone. If your bot has been calling 2.5 continuously you keep access. A project that was paused, or a new key created now, may find 2.5 unavailable and should move to 3.8.
- Google restricted Gemini 2.5 access on 18 September 2026 to users who had actively used those models.
- This is unusual: access depends on your recent usage history, not only on a published retirement date.
- Continuously running bots are the safest. Dormant projects and brand new keys are the exposed ones.
- Gemini 3.8 Flash reached general availability on 2 September 2026, with 3.8 Flash TTS and Flash-Lite TTS following on 22 September.
- Test a fallback provider before you need it, because this restriction can apply without a request ever failing first.
What Google changed, and the wording that matters
On 18 September 2026 the Gemini API changelog published a Gemini 2.5 models access update. The stated reason was to ensure reliable performance for everyone, and the mechanism was to limit access to the 2.5 models to users who had actively used them.
Read that twice, because it is a different kind of announcement from the ones the industry is used to. A normal deprecation says: this model stops working on this date, for everybody. This one says: whether you still have access depends on whether you were already using it.
That makes the impact uneven and hard to predict from the outside. Two businesses on the same plan, running the same code, can end up in different positions purely because one of them paused their project for a few weeks and the other did not.
Who is actually exposed
The risk is concentrated in a few specific situations, and it is worth checking which one you are in.
- Lower risk: a bot that has been calling Gemini 2.5 every day without a break. Continuous use is exactly what the restriction preserves.
- Higher risk: a project that was paused, a seasonal campaign that runs twice a year, or a staging environment that sits idle between releases.
- Higher risk: a brand new API key or a new Google Cloud project, which by definition has no usage history on 2.5.
- Higher risk: moving a working bot to a different Google account, for example when a client takes over billing from an agency.
That last case catches people out most often. The bot works, the code is unchanged, and the only thing that moved was the key. If the new account has no 2.5 history, the model it has always used may simply not be available to it.
Check it directly rather than assuming
Do not infer your position from the announcement. Make a real call.
The quickest test is to send one short request to the exact 2.5 model id your bot uses, from the key your bot actually uses. Not a similar key, not a colleague's project. Access depends on the account's own history, so only that key answers the question.
If the call succeeds, you have access today. If it fails with a model availability or permission error, you have your answer and a decision to make. Either way the test takes a minute and removes the guesswork.
On Convo360 the provider and model are set per agent, so you can see at a glance which agents name a Gemini 2.5 model. Check dormant agents too. A seasonal campaign agent that nobody has opened since the last festive season is precisely the profile this restriction affects.
Where to go instead
Google has been shipping steadily, so the destination is not in doubt. Gemini 3.8 Flash reached general availability on 2 September 2026, and Gemini 3.8 Live followed on 15 September. For voice work, Gemini 3.8 Flash TTS and Flash-Lite TTS became generally available on 22 September 2026.
Flash variants are the right shape for a chat bot. They are built for speed and cost at scale rather than maximum reasoning depth, which is what a two sentence WhatsApp reply needs. Our note on Gemini API pricing for chatbots covers how the tiers compare in practice.
Moving from 2.5 to 3.8 is a model id change, not a rewrite. What does deserve a check is reply length and tone, since newer models often format differently, and on WhatsApp a reply that suddenly runs to four paragraphs is a visible regression even if the facts are right.
The wider lesson about depending on one provider
This announcement is a useful reminder that model availability is not only about dates. It can depend on your account history, your region, your plan, or a capacity decision made by the provider on a Tuesday.
The practical defence is cheap. Keep a second provider's key configured, and test it occasionally. On a bring-your-own-keys platform an unused key costs nothing, because the provider bills on usage, so the only cost is the few minutes of setting it up.
Then once a quarter, point one agent at the fallback, run three real conversations, and point it back. That proves the key still works, the account still has credit, and the replies are acceptable. A key added during an outage is a key that fails validation at the worst possible moment.
Knowing about the next one in time
The Gemini API changelog is where this was announced, and it is not the same page as Google's AI blog. Launches, research and company news live in one place; access changes and deprecations live in the changelog. Following the blog and not the changelog gives you plenty of reading and none of the warnings.
That pattern holds across providers. OpenAI publishes retirements on a deprecations page that its news feed does not carry. Anthropic publishes them in API release notes. If you only ever read one page per provider, read the technical one.
Convo360 reads the Gemini API changelog along with Meta's, OpenAI's and Anthropic's several times a day, and emails the owner immediately when an entry removes a feature or changes pricing, with everything else held for a weekly digest. This Gemini 2.5 restriction is exactly the kind of entry it exists to catch, because nothing fails until the day you happen to need the model.
Step by step
- List agents naming a 2.5 modelCheck every agent, including dormant and seasonal ones, for a Gemini 2.5 model id.
- Make one real call with your own keyAccess depends on that account's usage history, so test with the exact key your bot uses.
- Move chat agents to Gemini 3.8 FlashFlash variants suit short chat replies. Change the model id rather than rebuilding anything.
- Check reply length after switchingNewer models often format differently. Run three real conversations and tighten the prompt if replies grow.
- Configure a fallback providerAdd a second provider key now and test it quarterly, since an unused key costs nothing.
Example conversation
Switch AI providers in a dropdown on Convo360
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
What did Google change about Gemini 2.5?
On 18 September 2026 Google limited access to the Gemini 2.5 models to users who had actively used them, stating the aim was to keep performance reliable for everyone.
Will my running bot lose access?
A bot calling 2.5 continuously is the least exposed, since active use is what the restriction preserves. Paused projects, seasonal campaigns and new keys are the risk cases.
What should I move to?
Gemini 3.8 Flash, generally available since 2 September 2026, suits chat bots. For voice, 3.8 Flash TTS and Flash-Lite TTS became generally available on 22 September 2026.
Why did it break when we changed Google accounts?
Access is tied to the account's usage history rather than your code. A new account or key has no history on 2.5, so the same bot can lose a model it always used.
How do I hear about changes like this sooner?
Follow the Gemini API changelog rather than Google's AI blog. Access changes and deprecations appear in the changelog, while the blog mostly carries launches and research.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.