The Good Models Got Pulled
In our own internal tests Opus 4.7 is consistently underperforming GPT 5.4 on coding tasks. That is GPT 5.4. Not GPT 5.5. And the strange thing is that GPT 5.4 is an unbelievably good model, on par with Opus 4.5. Both nail coding tasks with ease. Both are no longer available on OpenAI's and Anthropic's own subscriptions. I can only speculate as to why.
The cost of running these models is still reasonable. The capability is still there. They just are not the model you get when you pay for the standard tier anymore. You get the newer thing, which on paper is better, and in practice often is not.
The only place still holding up is Copilot, which gives access to GPT 5.4 at 1x. That is a genuinely good deal. It is also going away in a few months unless you locked in an annual subscription.
Effectively, all major coding agent providers are kneecapping their customers at the same time. The result is that the next several months will feel like a step backwards for anyone relying on these tools day to day. Code quality will drop. Costs will go up - significantly! And the gap between what was possible six months ago and what is possible on a standard subscription today will keep widening.
If you fired developers because of AI, this is the part where you find out the floor was held up by models you can no longer buy. The tooling got worse, the bill got higher, and the people who used to catch the difference are not there anymore. I hear some companies are hiring back junior developers to write some initial code before they feed it to the agent just to save on API costs. That is a very expensive way to solve the problem that did not exist a year ago and it feels like a very distorted market response, bordering on the absurd.
The lesson is the same one as always. Do not build your operation around a capability that exists only at someone else's discretion. The model you depend on today may quietly disappear from the menu tomorrow, and the replacement may be worse at the exact thing you needed it for.
A practical note: ChatBotKit lets you pick the model per use case and swap it without rewiring your agents. When a provider pulls the model you relied on, you change a setting instead of rebuilding the stack.