On July 1, Beijing time, American head model manufacturer Anthropic launched a new generation of mid-range main model Claude Sonnet 5. At the same time, the official confirmed that it has received a notification from the U.S. Department of Commerce to officially lift restrictions on the company's top flagship models Claude Fable 5 and Mythos 5, and will fully restore global access starting tomorrow.
However, Anthropic did not directly respond to the question raised by netizens whether the new model is open to everyone or only to authenticated users/citizens. Previously, officials revealed that Anthropic has been working closely with the U.S. government and deployed Mythos5 to some U.S. operations and defense critical infrastructure organizations on June 27.
The launch of Sonnet 5 marks the completion of Anthropic’s fifth-generation model system. In the product series, Sonnet 5 is between the lightweight Haiku 5 and the high-end flagship Opus 4.8 (and Fable 5). Test data shows that in terms of core Agent programming capabilities, Sonnet 5 recorded a score of 63.2%, 5.1 percentage points higher than the previous generation Sonnet 4.6 (58.1%), and the gap from the top flagship Opus 4.8 (69.2%) narrowed to less than 6%.
In the official technology positioning layout, the Sonnet 5 longboard focuses on complex code generation, in-depth analysis of long documents, multi-step automation, and the strong implementation of "computer control" (including independent takeover of browsers, terminals, and GUIs). The new model has the ability to independently formulate long-term plans and coordinate calls across tools. It can run independently for a long time without manual intervention, meeting the needs of the current developer community for executing long-term tasks.
Agent and landing application capabilities are also the focus of Anthropic's core competition with OpenAI. In order to compete with OpenAI's commercial data and revenue redline before it goes public, Anthropic adopted a new word segmenter (Tokenizer) in Sonnet 5, which improves the underlying accuracy while causing the same text input to consume 1.0 to 1.35 times more tokens than before.
In order to hedge against the rebound in total user cost (TCO) caused by the reconstruction of the tokenizer, Anthropic has launched a phased price promotion: during the promotion period before August 31, the basic pricing per million Tokens of the API is reduced to US$2 for input and US$10 for output; but from September 1, it will return to standard billing (US$3 for input and US$15 for output). The developer group believes that due to the expansion of the actual Token consumption base, after the promotion period, the actual call bill will inevitably be higher than the historical version.
Behind this business decision is the pre-IPO financial pressure faced by Anthropic. When Wall Street evaluates cutting-edge AI laboratories, it no longer only looks at the number of users (MAU), but will focus on auditing ARPU (average revenue per user) and gross profit margin. Compared with the top flagship Opus, which is expensive and has limited call volume, low-end Haiku has meager profits. The mid-range Sonnet system can handle most complex services, thereby increasing the company's core business flow rate.
It is worth noting that during the time when Fable 5 and Mythos 5 were previously restricted, the window period in the global AI market was accelerating towards Chinese manufacturers. During this period, the GLM-5.2 launched by Zhipu, a large Chinese model company, triggered widespread discussions in overseas open source and developer communities due to its cost-effective reasoning. Overseas developers are showing a trend of relying on diversified technology stacks, and Chinese large models have shown stronger substitution flexibility in the global wave of Agent evolution.