My love affair with Claude has cooled down considerably over the last couple of months. Claudisms became really annoying to decipher, and Sol 5.6 and Astra have just been outperforming Claude in so many ways. However, after playing around with Opus 5.5 a bit, I can officially say: Anthropic strikes back!
Extremely cost-efficient. I hit my OpenAI 5-hour usage limits all the time, but with Opus 5.5 I barely hit them, even on high effort. And it gets most tasks done well on medium effort already.
Much better writing output. They did a great job at reducing Claudisms, and it just writes in a more straightforward and easy-to-understand way. This is something I observed in my own usage, and Anthropic also published a side-by-side comparison of the output if you want to have a quick look.
Higher limits. Anthropic also raised users’ 5-hour limits by 20%.
AGENTS.md support. Claude finally supports AGENTS.md (instead of just CLAUDE.md).
Dan Shipper from Every also shared first findings from their team about Opus 5.5. It’s an interesting read, but the tl;dr is:
- Some prefer Opus 5.5 over Fable 5.1.
- At writing, it’s much better than prior Claude models, but still slightly behind Astra.
- It’s cheap!
- It’s great for building apps and interfaces.
If you want independent and well-structured evals, check out Mike’s Checks on Every.