Anthropic released a new generation of medium-sized models, Claude Sonnet 5, which focus not only on enhancing the performance of the dialogue, but on moving the capacity of the Agent, which can carry out its mandate on its own, to lower-price zones. As OpenAI and Google continue to strengthen their products in recent months, model competition is shifting to implementation capacity under cost, stability and less manual intervention.
Starting July as default model
Anthropic indicates that, as of Tuesday, Claude Sonnet 5 will become the default model for free and Pro versions and will be open to all subscription levels. The company positioned it as a product nearing the high-level model, Opus 4.8, but at a lower cost.
At the time of publication, Sonnet 5 prices were $2 per million input token and $10 per million output by 31 August. Thereafter, the input price will be increased to $3 per million and the output price will remain unchanged.
- August 31st Predate: Enter $2/million token
- From September: $3/million token
- Output price: $10/million token
From a price comparison, Sonnet 5 is lower than Anthropic's own Opus 4.8 and lower than OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, but still higher than Gemini 3.5 Flash.
Agent mission performance upgrade
Anthropic claims that Sonnet 5 was upgraded in terms of reasoning, tool calls, software programming and knowledge-based tasks compared to the Sonnet 4.6 released in February this year. The baseline test given by the company showed that the model scored 63.2 per cent on the Agent programming project, which was more than 58.1 per cent of Sonnet 4.6, close to 69.2 per cent of Opus 4.8.
In the knowledge work test, Sonnet 5 performed even slightly higher than Opus 4.8. The latter had been used more previously to deal with difficult tasks, such as complex judgments and in-depth studies.
According to Anthropic, this means that developers can make better trade-offs between cost and performance. Sonet 5 will be a cheaper option for a scenario that does not require the highest level of precision but wants to stabilize multi-step tasks.
Increased security
Besides performance, Anthropic also focused on security performance. The company asserts that Sonnet 5 is superior to the previous generation ' s model in the areas of improper assistance, deception, incendiary response to an attack and refusal of malicious requests. There has also been a decrease in the incidence of hallucinogenic and concubine responses.
However, Anthropic also mentioned that Sonnet 5 has not yet reached the levels of Opus 4.8 and Claude Mythos Preview in addressing the risk of mismatch behaviour. At the same time, the company states that the model is significantly less capable of carrying out hazardous network security tasks than the current Opus series and is therefore more suitable for large-scale deployment in the Agent scenario.
Some of the test users also gave positive feedback. Zapier Engineer stated that Sonnet 5 had been able to complete a multi-step automation mission that had previously been easy to stop; another AI product company, Loveable, stated that the model was more stable in refusing unsafe requests.