I was a bit late with this review because I wanted to thoroughly test Claude 4.5 and tell you in detail about how I’m using it and how you should now work with Claude after this model’s release.

What can I say - this is a truly awesome breakthrough for Anthropic.

Main Recommendation: Switch Completely to Sonnet

The first thing you should do right now is completely turn off Opus. Just set the default model to Sonnet and don’t even think about Opus. You can keep Opus Plan Mode - they removed this toggle from the settings, but you can still set it manually if you know what this mode is called when selecting a model.

For myself, I created a sub-agent called “Ask Opus”. Opus is still more powerful than Sonnet in terms of pure thinking and broad perspective, the complexity of tasks it can solve. But for standard workflows, Opus isn’t needed at all. Moreover, you’ll constantly hit limits and just suffer if you keep using Opus the way it was normal before version 4.5.

Now I just ask Sonnet to reach out to Opus for complex algorithms and tasks - and that’s absolutely enough for me.

Key Improvements in Sonnet 4.5

Precise Instruction Following

Sonnet 4.5 follows instructions much more precisely - and this is perhaps the most important improvement for me. I have a hook that triggers every time I see the phrase “You’re absolutely right”. The number of triggers for this hook has decreased by at least four times without any changes to my workflow or prompting techniques.

This is huge progress. Claude now just does what’s required of it and has become much closer to GPT-4 in this regard. I have a command “never end your work with a description of what you did - just stop silently”. I’ve had this instruction since June, neither Opus nor Sonnet 4.0 followed it. Only Sonnet 4.5 started following this instruction - now it unrolls code like a cyborg and stops silently when the work is done.

Tool Call Parallelism

Huge improvement - tool call parallelism. Opus somehow called tools in parallel in about 10-15% of cases, Sonnet 4.0 practically never did this, although both models had access to parallel calls.

Sonnet 4.5 was specifically post-trained to call tools in parallel when there’s such a possibility. This is really cool because less time is spent on thinking tokens - instead of stupid “now call this”, “now call that” which clutter the context meaninglessly. Now Sonnet 4.5 calls 8 tools in a row, they all execute at once, and editing happens several times faster.

Drawbacks

It’s a shame that Sonnet still only has 200 thousand tokens of context window. Especially considering the negative changes - constant notifications in the context after each tool call: “you have this many tokens”, “spent 5000 tokens on this call”, “15% tokens remaining”.

Claude has started being really afraid to work and economizing like crazy - saving lines of code, especially when reading files. I’d like to have 400 thousand tokens so this wouldn’t be a problem. It’s unclear why Anthropic is still economizing on context instead of expanding it.

Thinking Quality

I no longer notice a difference with Opus in thinking quality. Previously, Claude 4.0’s thinking was 70% hallucinations and endless arguments with itself. For simple tasks, thinking was useless. With Opus, thinking worked great, which is why I stuck with Opus almost always.

With Claude 4.5 they made an always-on thinking mode - and rightly so. Sonnet 4.5 really uses thinking tokens correctly to help with work. The thinking quality is now on par with Opus.

Economic Advantages

The coolest change - Sonnet 4.5 can replace Opus for most of the work. Due to the increased quality and greater trust in the model, I can’t even reach limits anymore.

Yesterday Claude was working permanently all day from 10 AM - I didn’t spend even 5% of the weekly limit. By the end of the week I spend a maximum of 10% of the weekly limit and never spent more than 15% of the 5-hour limit, despite the agent working constantly.

What does this mean? For many people who are currently on the $200 subscription, if you’re not running more than 3 agents in parallel - you don’t need the $200 subscription. I sincerely think you can get by with the $100 subscription and never hit limits.

Impact on Productivity

This has greatly expanded possibilities. In programming I can now work not on two projects, but on five at once. Now I can afford more rewrites, more experiments, more proof-of-concept implementations that I can then throw away and not regret the spent tokens.

Because Sonnet 4.5 is also much faster than Opus and even faster than Sonnet 4.0, I can produce prototypes and test them at cosmic speed - all on the same subscription I had.

Perhaps this is one of the best Claude updates that has ever been released. I even started using Sonnet periodically in the regular chatbot because sometimes it gives better results than Opus. Incredible, but true.