The company Anthropic has released Opus 5, the latest version of its most powerful artificial intelligence model. According to the announcement, the new model is cheaper and less restricted than the previous flagship, and on some tests it surpasses even the firm's larger model. In other words: smaller, faster, more permissive - and sold as a step forward.

The pace alone says a lot. Opus 5 comes out just two months after Opus 4.8, which became available on 28 May. In the meantime the company released three more models in the same series. When a tech firm announces a new „best” version every few weeks, it is worth asking whether every new release is real progress or just a marketing rhythm that must not stop.

The most interesting part is not the power, but the loosened restrictions. Anthropic says that with Opus 5 the safety mechanisms will kick in 85 percent less often than with the larger model. The company explains this by the fact that a less capable model gets a „lighter touch”. Translated: the more helpful the machine, the more it is watched over the shoulder - and the weaker it is, the more freely it is allowed to run.

Certain guardrails do remain. The model has built-in protections around tasks tied to cyber-security - for example, it must not look for weaknesses in a finished program file, but it may look for them in the source code, since the latter more often serves defence than attack. A thin line, drawn by the very company that produces and sells the product. Who checks the line?

Anthropic also introduced a new option called automatic fallback paths. When a query triggers the safety filter, the system, instead of an error message, automatically reroutes the query to a weaker model, so the user still gets an answer. Practical for the user, but also symptomatic: the industry is looking for ways to keep the machine from saying „no”, even when it initially should have.

Opus 5, like its predecessor, is not subject to the 30-day data-retention policy that applies to some other models - a detail that worried privacy-sensitive users. A small item in the announcement, but it is exactly such small items that decide where your conversations with the machine end up. Read them, because no one else will do it for you.