@Murgur we are looking into that. We have Sonnet 4.5 overall as Thinking and non-thinking model.
@Artemonim Sonnet 4.5 (Not Claude 4.5) costs same as Sonnet 4 and therefore lower cost than Claude 4.1.
@TRTRTRTR @furki @DavidWasHere unfortunately this time around it’s not available as free model.
The question is more about the general work. My Claude is several times more expensive Usage than gpt-5, despite the fact that they hardly have the same difference in API.
I am curious how the new model behaves with context/tool usage and caching. Including how it affects price.
One more thing: Sonnet 4.5 can use Browser from Cursor (Beta setting for Early Access users)
Settings:
Early Sonnet 4.5 Review:
It seems to not act like an AI, in the way Sonnet 4 did. It acts almost human–it’s far more likely to do the change in the same process/way that I would do it myself. Straightforward–doesn’t waste time. Simultaneously, major improvements to code writing from scratch–does it properly, like a human, as well. To sonnet 4, it’s kind of like what gpt-5-codex is to gpt-5 if you get what I mean.
Skip the shiny Claude Sonnet 4.5 update. Stick with GPT-5-Codex. It may feel slower, but it delivers the real coding power, you’ll thank me later.
what’s about grok-code-fast-1?
Grok-Code-Fast-1 is solid, fast, efficient, and good for everyday dev work. A good choice if you want speed.
These benchmark results can be misleading, especially when published by the model’s own company, bias is inevitable. At the end of the day, it’s more about selling you than proving real-world coding performance.
We and everyone need to use it to have the most accurate view instead of looking through that chart.
Update after Further use of Sonnet 4.5:
It has a strong aversion to chat summarization; It’s very aware of it’s token usage, sometimes to the detriment of output quality–it tries to finish the task before summarization kicks in automatically. I’ve observed it going step by step through the todos (as instructed), but once it gets to around 80% of token usage, it says something like “Due to length constraints, let me focus on the most critical remaining tasks.” then rushes through the final tasks. I think Agent design needs to be improved for 4.5 specifically? Cursor-side, that is–please tweak the threshold such that it properly summarizes earlier around 80% token usage–or at least the agent understands the summarization, and doesn’t rush. Ultimately, requires more careful usage of manual /summarization.
In cursor, a single use of Sonnet-4.5 conversation only deducts the number of times multiple conversations are used after adding content to the message queue and receiving feedback from MCP!
After using message queue and feedback MCP to add information, sonnet-4.5 in cursor will now deduct more conversation times. Previously, gpt5 was counted as one conversation, but now it directly deducts several times the consumption.
hi @Tao_Zou are you using thinking mode with Sonnet 4.5? If yes, you can switch to non-thinking mode which uses less tokens and is cheaper.
My honest review: didn’t like it. Leaves code with errors and gives out wrong solutions to tasks. I still prefer gpt-5-high.
I think Claude 5 will be the best one yet. Waiting for it.
My Sonnet 4.5 (not MAX) is cutting off when context window ends instead of summerizing and going forward. Is that only on my end?
As they say, if you don’t use it as thinking, it costs the same as Sonnet 4. However, calculating the thinking model directly as 2x is extremely expensive
> Opus 4.1 
(Duy Nguyen)
