Anthropic releases Claude 3.7 Sonnet, his first “hybrid reasoning model” that can solve more intricate problems and exceeds previous models in areas such as mathematics and coding.
In addition to the up-to-date model, Anthropic also publishes a “limited view of research” of its “agency” coding tool called Claude Code. While Anthropic already supplies tools encoding artificial intelligence, such as Cursor, it is Claude’s code as “an active collaborator who can search and read the code, edit files, write and run tests, commit and push the code to GitHub and use the poem tools commands. “
The Claude 3.7 Sonet is available from Monday in the Claude application, and for programmers via API, Amazon Bedrock and Google Cloud Vertix AI. The model costs the same as its predecessor, 3.5 sonnet, 3 million input tokens and USD 15 for one million output tokens.
While OpenAI and others offer separate so -called reasoning models, anthropic product study Dianne Penn tells The Verge that the company wanted to simplify the experience of using the model. “We basically believe that reasoning is a feature of artificial intelligence, and not a completely separate thing,” he says, noting that Claude should not answer the question “what time?” Compared to responding to a more intricate prompt, for example: “Plan a two -week trip to Italy, considering the weather at the end of March.”
Penn says that Sonnet Claude 3.7 will reach clearly better in “agency coding”, financing and legal tasks. While Claude is still missing in real time, like other models, the date of cutting off knowledge in version 3.7 in October 2024 is more current. Anthropic also allows programmers to assist manage how the “thoughts” model via Scratchpad, and even dictates exactly how much time I answer. “Sometimes a programmer just has to say that this question should not last more than 200 milliseconds,” says the Vice President for the Product, Michael Gerstenhaber. “And this is a decision about the product.”
Inside Antropic, employees used a up-to-date model for building websites, interactive games, and even spending up to 45 minutes for coding of work by “building test sets and editing test cases and back”, according to Penn.
He says that the company also tests its models on their ability to go through the ancient school video game Pokémon by mapping the API model of the model to the controller scheme. Sonet Claude 3.5 could not get out of Pallet Town at the beginning of the game, while version 3.7 was able to beat many gym leaders.
As Elon Musk from Grok-3 showed last week, the AI model race moves incredibly brisk. For now, anthropic seems to be at the forefront again thanks to the benefits of Claude 3.7 Sonnet. Its release also suggests that instead of offering independent reasoning models, the industry is heading towards the future in which one model can do anything.