Independent testing

Change Radar

What changed with these tools, sorted by what it costs you to miss it rather than by when it happened. Most entries say there is nothing to do, because most of the time there is not.

How current this is

All 3 sources were checked recently, so an empty list below genuinely means nothing has changed.

  • Anthropicwithin the hour
  • Googlewithin the hour
  • OpenAIwithin the hour

1 to do something about, 0 worth knowing, 0 that need nothing.

Something to do2026-10-01

claude-sonnet-4-5-20250929

The maker has said it is on the way out. It still works for now, and there is usually a date.

What changed
We announced the deprecation of the Claude Sonnet 4.5 model (claude-sonnet-4-5-20250929), with retirement on the Claude API scheduled for November 30, 2026. We recommend migrating to Claude Sonnet 5.5.
Who it affects
Anyone who has built anything on it, including a prompt they reuse every week.
What to do
Note the date somewhere you will see it again. This is the change people most often read, agree with and then forget until it stops working.

Where we saw it

How every change gets filed

These are our rules, written down, applied the same way every time. No model decides what matters here: one asked whether a change is important would say yes, fluently, and be wrong in the direction that gets attention.

  • The model has been switched off. Anything pointing at it has stopped working, or is about to.Something to do

    Find where it is named in what you have built and change it before the date, not after. If you only use it through an app, the app's makers have to do this and you may want to ask them when.

  • The maker has said it is on the way out. It still works for now, and there is usually a date.Something to do

    Note the date somewhere you will see it again. This is the change people most often read, agree with and then forget until it stops working.

  • Something it used to do it no longer does — reading images, say, or handling a file type.Something to do

    Try the thing you actually use it for. If it has gone, you will find out now rather than in the middle of a job.

  • It can hold less at once than it could. Documents that used to fit may not.Something to do

    Put your longest ordinary document through it once. A shrunken window does not announce itself: the middle of a long input quietly stops being considered.

  • It costs more per use than it did. The published rate for sending it text, or getting text back, has gone up.Worth knowing

    If you are on a subscription, nothing. If you pay per use, look at what you spent last month and work out what the same usage costs now.

  • It has come out of preview and is treated as settled, so it is less likely to change under you.Worth knowing

    Nothing urgent. It does mean the behaviour you test today is more likely to be the behaviour you get next month.

  • It can do something it could not before — read a picture, say, or handle a file type it used to refuse.Worth knowing

    Only worth your time if you built a workaround for exactly this. If you did not, there is nothing here.

  • It can hold more at once than it could, so longer documents and longer conversations now fit.Nothing to do

    Nothing. It is there if you need it.

  • The same thing is being called something else.Nothing to do

    Nothing, unless a name is written down somewhere that will outlive your memory of the change.

  • It costs less per use than it did. The published rate has come down, which happens more often than people expect.Nothing to do

    Nothing. It happens on its own.

  • A model that did not exist before now does.Nothing to do

    Nothing. If what you use does the job, a new one existing is no reason to move. We have not tested either of them anyway.