Anthropic has unveiled Claude Sonnet 5 as its most agent-like Sonnet model to date, and this phrasing is no accident: the leap forward is not just about the quality of the responses, but the ability to plan, use tools such as browsers and terminals, and carry out complex tasks independently. The message to the market is clear: useful AI is no longer simply about speaking well, but about being able to work effectively without getting bogged down.
This development is particularly significant because Sonnet 5 bridges part of the gap with the Opus range, which has historically been more powerful and expensive. Anthropic claims that the new model’s performance is close to that of Opus 4.8, but at a lower price, with a marked improvement in reasoning, coding, tool use and knowledge work. In a sector where every percentage point of performance seems to come at a high cost, this combination of efficiency and capability could become a formidable commercial asset.
The most interesting aspect, however, is not merely technical but strategic. Anthropic is seeking to establish a foothold in the AI agent market with a model that has a more accessible entry threshold but delivers high enough performance to be used in production by development teams, businesses and software platforms. It is a shrewd move, as it targets that segment of users who do not wish to pay top dollar for the top-of-the-range model, yet still expect solid results on lengthy, operational tasks.
On the security front, Sonnet 5 adopts a more cautious approach than the more aggressive models. Anthropic states that the system exhibits less undesirable behaviour than Sonnet 4.6, has improved resistance to prompt injection attacks, and is less prone to providing complacent or delusional responses. At the same time, the model remains less suited to the most dangerous cyber tasks than Opus 4.8 and Mythos 5, which is why the guardrails are enabled by default. It is a delicate balance, but it reveals a clear direction: greater autonomy, yes, but without loosening the reins too much precisely where the risk increases.
There is also an important indication regarding real-world adoption. Anthropic reports very positive feedback from its first partners: Sonnet 5 is described as a model capable of carrying out multi-step tasks, monitoring its own output and completing jobs where previous versions would have stalled halfway through. In enterprise contexts, this matters far more than a spectacular benchmark, because it means fewer interruptions, less manual supervision and greater operational continuity. In short: fewer demos, more real work.
The pricing also reflects a clear intention to expand. For the initial period, Claude Sonnet 5 is being offered at $2 per million input tokens and $10 per million output tokens, with the aim of making the transition essentially cost-neutral compared to the previous generation. It’s an aggressive yet well-calibrated move: lowering the price whilst raising the quality is the quickest way to set a new standard, especially as competition in the field of agentic AI is becoming fierce.
To find out more, follow the linkto Anthropic’s official article.
