OpenAI retires GPT-3.5: how often do you upgrade the AI in your systems?
OpenAI scheduled the shutdown of GPT-3.5 Turbo, the model many companies built their first chatbot on. In audits we keep finding processes that depend on old models: they work, but they're slower, more expensive and less capable than what's available today. Here's what we recommend.
October 23, 2026 is the date OpenAI scheduled to shut down GPT-3.5 Turbo in its API. It's the model many companies used for their first chatbot, their first email classifier or their first AI automation, back in 2023.
If any of your processes still use it, the question is no longer whether it's worth switching. It stops answering.
But the shutdown is only the symptom. The real question is a different one: how often do you check which AI model runs your systems?
How often new models come out
To put it in perspective:
- November 2022: ChatGPT launches, running GPT-3.5.
- March 2023: GPT-4. Four months later.
- May 2024: GPT-4o, faster and cheaper than GPT-4.
- July 2024: GPT-4o mini, a small model that cost less than a third of GPT-3.5 Turbo and was more capable.
- September–October 2026: Anthropic alone released three models (Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5) in just over two weeks. Independent trackers counted more than twenty new models from different labs in two weeks of September.
At first, a new model came out every few months. Today one comes out practically every week, and each generation is usually faster, cheaper or more capable than the last; sometimes all three.
What we find in audits
When we audit processes at companies that already use AI, the pattern repeats: a system built one or two years ago that "works" and that nobody has touched since. It still calls the same model it was born with.
And yes, it does the job. It answers, classifies, summarizes. But doing the job isn't the same as being in good shape. An old model costs you in four ways:
- Price. Today there are small models that cost a fraction of what 2023 models did and do the work better. Paying more for less is money that leaves every month.
- Speed. In a customer service chatbot, seconds of waiting are customers who leave. Current models usually answer faster.
- Capability. What two years ago required huge prompts and patches, a current model gets on the first try: reading documents, following long instructions, using tools, understanding context.
- Risk. Providers retire their old models on a set date. If nobody is watching, one day the process simply stops working. Just like what happened with GPT-3.5.
The question isn't "does it work?" It's "why stay behind when there's already something better and cheaper?"
Upgrading isn't switching blindly
Careful: the opposite isn't a good idea either. Switching to the newest model the day it comes out, because it's "the smartest," can break a process that was working fine or cost more than necessary. The most powerful model isn't always the right one for every task.
At Avanzia we do it like this:
- One model per task, not one for everything. A chatbot that answers frequently asked questions doesn't need the same model as a system that analyzes contracts. For customer conversations we use fast, affordable models; the big ones only where they're truly needed.
- Periodic review. We check which models run the systems we have in production and whether there's a better or cheaper option, or whether any of them has a retirement date.
- Test with real cases before switching. Before moving a process to another model, we run it with the client's real conversations and documents and compare it against the current model. More than once, not with a single test. We only switch if the result is the same or better.
- Measure afterwards. Cost per conversation, response time and quality, before and after the change.
Five questions to review your systems today
- Do you know which AI model each of your processes uses?
- When was the last time someone reviewed it?
- Has your provider already announced a retirement date for that model?
- Do you know how much it costs you per month and per conversation?
- Is someone responsible for upgrading it, or was it left "as it was"?
If any answer was "I don't know," it's worth reviewing before your provider decides for you.
Want us to review yours?
At Avanzia we audit your company's AI processes: which model each one uses, how much it costs, what retirement risk it has and which alternative makes sense. And if switching makes sense, we do it by testing with your real cases, without breaking what already works.
