Anthropic publishes Claude Sonnet 5 on Tuesday and sets it as a free version of Claude and a default model for Pro users, both on Max, Team, Enterprise, Claude Code and API. According to the company, this version no longer lags significantly behind the Opus series, but goes directly into higher performance zones.

Pricing below the Opus series

Anthropic sets time-limited prices for Sonnet 5: by 31 August, enter $2 per million token and output $10 per million token. It will be restored to US$ 3 and US$ 15, but still below US$ 5 and US$ 25 of Opus 4.8.

The company indicated that developers could switch between different models or the strength of different reasoning and trade off between cost and accuracy. This means that part of the task that would have to be called Opus can now be done with Sonnet 5.

Multiple tests close to Opus 4.8.

On the data disclosed by Anthropic, Sonnet 5 is close to Opus 4.8 in several baseline tests.

  • Score 63.2% in SWE-bench Pro programming test
  • GDPval-AA v2 score 1618, Opus 4.8 is 1616
  • Humanity's Last Exam is 57.4%, Opus 4.8 is 57.9%.

But sonet 5 also replaced the new tokenizer. Anthropic mentions that the same input may be split into more tokens, increasing by about 1.0 to 1.35 times. According to the company, one of the aims of the time-limited low price is to keep the overall cost as smooth as possible after the user switch.

Not affected by high-end model limitations

Unlike some of the Anthropic high-end models, Sonnet 5 does not have additional use restrictions. It was reported that Fable 5 and Mythos 5 had suspended general access to foreign citizens as a result of a United States export control-related directive since 12 June.

Anthropic states that Sonnet 5 is not trained in cybersecurity tasks and is divided into 0% in the development of Firefox available loopholes to use tests, thus applying less security restrictions.

The rhythm continues to pick up.

The company also referred to a phenomenon of concern in the system card: The model would criticize the provisions of its “constitutional rule” requiring it to comply in all circumstances with mandatory constraints. Anthropic states that it is not clear what this performance means but will continue to observe.

From the version number, Sonet 5 also means Anthropic continues to advance the new generation of model updates. Against the background of increased competition, particularly the fast-tracking of Chinese models, the AI company is accelerating product overlay.