I used Composer 2.5 everyday and it is a very fast and excellent
Model became very stupid after 20.06.2026 - I suspect you redirect to very cheap model. It even doesn’t understand in the same session that changes are not being committed and creates deprecation workarounds around initial code changes - absurd. That was never happened before. It doesn’t understand that I use UI library (previously it uses it and correctly without any mentioning) This model feels right like gpt 2024 model.
Completely nothing was changed from my side - no skills, system rules - nothing.
I rarely write complains and go to forums, but the degradation is so noticeable and now quality is so bad that I better write by hand.
2 point together affected me 1. I KNOW how good Composer can be and this is definitely a planned degradation because of whatever business reason and that HIDDEN change insults me as a customer. 2. Now it’s barely usable
Steps to Reproduce
Just use the model
Expected Behavior
Expect the same quality of model like before 20.06.2026
hey @Alex0 sorry this happened to you, would you mind sharing some request id found if you click on the 3 dots near the chat? so that the team will be able to look into it and see what was wrong!
Hi, it’s 58f2d9c7-2628-4263-9d6c-e53a003fc459, but I want to point that it is exactly about all sessions since Sunday. Also for the context I update IDE instantly after update is available, so I don’t know if harness has changed or if there is some shadow redirect after amount of continuous usage. The issue I see not about exact chat, but overall model behavior. Before the date I was amazed how good the Composer is, it was lighting fast with quality of Opus and I planned and implemented refactoring of my app last week! It was flawless. Now with the same model I cannot implement a simple feature without adding 100 restrictions like I am back to 2024-25
Hey, thanks for the report. We checked the ID you shared and also looked at your account overall.
On the main concern first. There is no silent redirect to a cheaper model. When you select Composer 2.5, your requests are actually served by Composer 2.5. Routing confirms that both before 20.06 and after. There is no shadow redirect after some usage volume. Also, the survey popup, if you answered it, does not affect anything. It is telemetry only, it does not change the model or its behavior.
What you are describing, same model but lower quality, is most often not a model switch. It is usually about the session context like context window size, how the chat history has built up, how the task is phrased, and which files made it into context. This is especially noticeable during long refactors. A couple things that usually help:
Start a new chat for a new task so you do not carry a bloated context
Explicitly pull the right files or docs into context, for example that UI library via an @ link, if you notice the model is forgetting it
If you catch a specific reply that is clearly bad, please send the request ID from that exact message. Use the 3 dots on that specific assistant reply in chat. Then the team can review that session directly. Unfortunately, a general it got worse trend cannot be diagnosed from a single ID.
Thanks for quick responses, I understand basics guidelines and used more simple language in recent session because model was extremely good at getting everything needed and of course each case depends on the task.
At least if something changed internally in system prompts I want to notify you Because Composer 2.5 is a new level and tokens cost is unreal low (not sure if you subsidize own model). Looking forward to new horizons!
Glad we cleared it up. About the system prompts: the routing and model serving on our side didn’t change for you. Composer 2.5 has been served as Composer 2.5 both before and after 20.06, with no hidden switch.
If you catch a specific bad reply again, please send the Request ID for that exact assistant message use the 3 dots on that message in chat, then the team can look at that exact session. With a general “it got worse” report and no specific ID, there’s unfortunately nothing we can diagnose.
The best place to leave model feedback is here: Share your Thoughts on Composer 2.5! the team checks it. Let me know if anything feels off again.
Because Composer 2.5 is a new level and tokens cost is unreal low (not sure if you subsidize own model). Looking forward to new horizons!