Sonnet 4.5 — Reduced Context After Price Increase?
Mohammad Hussain
After the credit cost increase to 2×, Sonnet 4.5 in Windsurf is no longer handling the same complex prompts it used to. The same prompts run perfectly on Cursor, which suggests the usable context window in Windsurf has been reduced.
If capabilities are being changed — especially while increasing cost — please be transparent and share the updated context limits. Users rely on this model for real work.
Can the team confirm whether the context size was reduced, and what the current limit is?
— Hussain
Z7 Labs
Just circling back on this (Nov 30th, 2025)
It looks like Windsurf fixed this issue! I've installed a few updates since, and Sonnet doesn't seem to act "dumb" or "lazy". Also Sonnet has been upgraded to a 1M context window. Much thanks Windsurf team. Plus they added Opus 4.5 which is awesome and very capable. 👏�
I
Ignacio Bustos
I also confirm that after the promotion 4.5 sonnet is worse. The same than gemini pro after their promo exit a month earlier.
Jorge Palacios
Yes, Windsurf got really dumb since 2 weeks ago. It was really great and all of the sudden it became like sonnet 3.5
Z7 Labs
You hit the nail on the head. It does feel like less of a context window, but it might not be. I asked the agent to fill up the context window as much as it can by reading the codebase. It still reported a ceiling of 200K. But it definitely feels like something changed.
I've been using Claude Code for a week or two, and came back to Windsurf to do a few tasks. I too usually use Sonnet 4.5 thinking model while in Windsurf. It was doing a pretty good job for more complex tasks. Now I found after the Windsurf update. I wanted an easy sh script to do a build for me. It took multiple prompts, when in the past it would have thoroughly checked the code and one-shotted it.
I'm thinking of canceling if this isn't solved.. I hope they fix it because I really loved working with Windsurf, and it has taught me a lot of valuable coding insight.
(edit: Now I just noticed the $30 plan offers "Automated zero data retention"... so us peasants on the $15 plan get their data retained? I would have gladly upgraded if Windsurf worked as it did before)
Tamás Szelei
Z7 Labs I'm really not sure Windsurf is to blame here. This has been happening every time in the last few iterations with both Anthropic and OpenAI. New model comes out, it works great and definitely performs better than the previous. People move over and get used to it; then the model and/or compute gets nerfed behind the scenes. You silently get less context or you are routed simply to a dumber model.
Z7 Labs
Tamás Szelei I mean silently getting less context, or silently getting routed to a dumber model are both frustrating and annoying and seems like this is something in Windsurf, no? If you sort feedback by Newest here, you'll see that this topic sticks out and has the most votes of new topics Since October 29th (about last two weeks). So something is off.
I do feel It's working fine for me now, I'm just not sure why or how. I experimented some by using Ctrl Enter (to trigger fast context). I did that on any new chats about the project. So it's possible that Windsurf updates are not preserving some type of "memory" about the project/code base users are working with after an update. Seems very plausible this could be the case. But that's something Windsurf might need to address with how they run their updates. 🤷
Thanks for the reply though. I think Windsurf is a powerful tool, and I hope it just gets better and better. I'm glad they have this feedback outlet, I hope they see this and pinpoint the problem. Also, wen ARM Linux ver? :-)
Tamás Szelei
Z7 Labs Maybe it's windsurf, I don't know. I'm just saying the unfortunate state of things with LLMs being fundamentally SaaS is that we have no way to observe what is happening exactly with our queries. Each party involved (both Windsurf and the model service provider) have a vested interest in serving a cheaper model or using less compute with the model.