Back to News

Software Development · Sep 29, 2026

When AI model upgrades break things: maintaining AI integrations in 2026

On September 28, Anthropic released Claude Sonnet 5.5. The company says it is more than 30% faster and up to 30% cheaper than Sonnet 5, with better coding and agent performance. Six days earlier it released Claude Opus 5.5 with a 1 million token context window.

Both releases came with something the headlines skipped: breaking changes. Anthropic's own migration guidance lists five breaking changes for teams moving code from Sonnet 5 to Sonnet 5.5. For Opus 5.5, the platform release notes say extended thinking can no longer be turned off, and requests that force a specific tool choice now return an error.

If none of that means anything to you, that is fine. Here is what it means in practice: software that worked yesterday can stop working when someone changes one line to use the newer model.

AI integrations are a dependency like any other

When a business builds a custom tool that uses an AI model, whether it is a quote generator, an intake form that reads documents, or a support assistant, that tool depends on the model provider's API. The provider changes that API on its own schedule.

This is not unique to AI. Every integration with an outside service carries this risk. What is different with AI is the pace. In September alone, the major labs shipped more than half a dozen new models. Each one is tempting to adopt, because each one is faster, cheaper, or better.

Why "just switch to the new model" is risky

Even without breaking API changes, a new model behaves differently. It may format answers differently, follow instructions more literally, or be more cautious about certain requests. A prompt that was tuned for the old model may produce worse results on the new one, even if the new one is better overall.

I have seen automations where a model upgrade changed the way dates were written in the output, and a downstream step that expected a specific format quietly stopped working. Nothing crashed. Data just stopped flowing.

A simple upgrade process

You do not need a large engineering team to handle this well. You need a routine.

  1. Pin the model version. Your software should call a specific model, not "whatever is latest." That way upgrades happen when you choose, not when the provider does.
  2. Keep a test set. Save 20 to 50 real examples of inputs and the outputs you expect. Run them against any new model before switching.
  3. Read the migration notes. Providers publish them. They are short and they list exactly what will break.
  4. Upgrade one workflow at a time. Start with the least critical one.
  5. Watch the output for a week. Check that downstream steps are still getting what they expect.

When to upgrade at all

Not every new model is worth adopting immediately. Upgrade when there is a clear reason: lower cost at your volume, a capability you need, or because the provider is retiring the old version. Otherwise, the old model you have tested is often the better choice than the new one you have not.

What this means for your business

  • If you run custom software that uses AI, someone needs to own model upgrades. Ask who that is.
  • Breaking changes are normal and documented. The risk comes from upgrading without reading them.
  • A small saved test set turns a risky upgrade into a routine check.
  • Faster and cheaper models are worth adopting, on your schedule, after testing.

Keeping an AI workflow working through a year of model releases is where the real work shows. If you want a team that plans for maintenance from day one, that is how we work at Centervert.

Sources