The AI model landscape has been moving fast these past few months. Between surprise deprecations, new model families dropping, and pricing shifting every quarter, just keeping up has become a real chore.
At Fabio AI Chatbot (a WordPress AI chatbot plugin), we just did a full cleanup of our AI provider connectors. Here's what changed, and why.
OpenAI
We removed gpt-5.4, gpt-5.5, and gpt-5.4-mini — these generations have been superseded by the GPT-5.6 family. We're keeping gpt-5.4-nano (still relevant for some use cases) and adding gpt-5.6-luna, OpenAI's new fast, cost-efficient model built for high-volume workloads — pretty much exactly the profile of a chatbot running continuously on a site.
Google Gemini
We removed gemini-3.1-flash-lite-preview (discontinued by Google) and gemini-3-flash-preview (outdated). We're keeping gemini-3.1-flash-lite and adding gemini-3.5-flash.
A heads-up for anyone still hitting the Gemini API directly: gemini-2.5-flash-lite is being deprecated, with shutdown expected around October 16, 2026. If you've got that model hardcoded anywhere, now's the time to migrate before it breaks in prod.
Mistral AI
Radical simplification: we're down to a single model, mistral-large-3. Everything else got removed — fewer choices, but fewer configs to maintain and test every time Mistral's API changes.
Anthropic — the new kid 🎉
And the real news: we've added Anthropic as a provider, with claude-haiku-4-5 and claude-sonnet-5. This provider had been sitting in the database for a while but was never exposed in the UI — that's fixed now.
Why we do this cleanup regularly
A production chatbot pointing at a discontinued model is exactly the kind of silent bug that hurts — it can just stop working overnight with no obvious cause. Pruning the list of available models every cycle is as much about reliability as it is about cost — the newer nano/lite tiers are often meaningfully cheaper for comparable quality on chat-type tasks.
What about you?
How do you handle this on your end? Do you pin an exact version, follow the provider's "latest" alias, or have some kind of automated tracking/migration setup? Curious how other devs deal with this constant model churn — let me know in the comments.
Top comments (0)