# Experimented with Claude Opus 5.5

> Source: <https://www.dotnetperls.com/2026_9_27_experimented-claude-opus-55>
> Published: 2026-09-27 07:00:00+00:00

Recently I have been doing some **spec-based Rust code generation**. This has given me some correctness and performance improvements, and also off-loads the code maintenance a bit, making my life easier. I decided to try to improve some code with **Claude Opus 5.5**.

The **spec**, which I had previously used on GLM 5.3 Flash, was interpreted slightly **differently** with Opus. I had to correct an ambiguity in the spec, and then Opus generated the desired code. I don't think the issue was Opus's fault; it seemed to generate **correct code**.

The **cost** was $0.09 for the final code generation. **Opus improved the code** slightly, using `u8` values in a lookup table instead of `u32`, and the result was slightly faster in benchmarks.

I then did some further work in **DeepSeek V4.1 Flash**, which came out to less than a cent per generation. Basically, if it is possible to use a model like DeepSeek V4.1 Flash or GLM 5.3 Flash, those will be the most cost-effective. For **complex things**, or important functions, **Claude Opus** 5.5 may be worthwhile.
