Anthropic’s new Claude 4 AI models can reason over many steps

During its inaugural developer conference on Thursday, Anthropic launched two new AI models that startup claims are among the best of the art, at least how they score on popular benchmarks.

According to the agency, Claud Opus 4 and Clock Sonnet 4, New Family Part of Anathropick, Claud 4, can analyze large data sets, perform long-hurrying tasks and take complex steps, according to the company. Both models were tune in to perform well in programming functions, ethnographic said that they made them fit for writing code and editing.

Both of the payer users and the company’s free chatbot applications will get access to Sonnet 4, but only the payment users will get access to the OPS 4 for the AMazon’s bedrock platform and Google’s Vertex API, the price of OPS 4 is $ 75 per million tokens per million tokens ($ 75) per million.

Tokens are the raw bit of data that works with the AI ​​models, the equivalent of about 750,000 words – about 163,000 words “war and peace” longer than the “war and peace”.

The company has reached the clod 4 models of anthropic as the company looks at increasing enough earnings. ReportThe garment, founded by former Open Researchers, earned more than $ 2.2 billion this year to earn $ 12 billion to earn income in 2027. Ethnographic Recently closed A $ 2.5 billion credit facility and raised Billions of dollars From Amazon and Other investor In anticipation Expenditure Associated with the development of the border model.

The rivals did not make it easy to maintain the polar position in the AI ​​race. Time New Flagship AI Model Earlier this year, competitors, including Claud Coding equipment, also ran out to overcome the company, including their own powerful models and Dave Tooling, along with Claud Sonnet 4.7, OpenAI and Google.

Anthropic is playing with Claud 4.

The anthropologist says that the more capable of the two models introduced today, Opus 4, can maintain “focused efforts” in many steps of a workflow, ethnographic. Meanwhile, the sonnet impro the coding and mathematical models designed as “drop-in replacement” for Sonnet 4-Counte 3.7 and further specifically follow the instructions, according to the company.

Anthropologists claim that Sonnet 3.7 has less potential than Sonnet 3.7 to be involved in “award hacking”. Rewards hacking, which is also known as specification gaming, is a behavior where models take shortcuts and lufles to complete.

Clearly, these improvements did not give the world the fruit of the world Best Model by each criterion. Eg, when Opus 4 beats Google Gemini 2.5 Pro And Openai’s O3 And GPT -4.1 The needle-beech has been verified, which is designed to evaluate a model coding skills, it cannot exceed a set of multimodal evaluation MMMU or GPQA diamond, PhD-level biology, physics-and chemistry-related questions.

Nevertheless, anthropic bifid-up is publishing OPAS 4 under strict protection with harmful material detectors and cybercastle defenses. The agency claimed that its internal examination showed that the capacity of the Opus 4 stem background “may” enhance the ability to “enhance enough” to achieve, produce or deploy chemicals, biological or nuclear weapons, to reach. The “ASL -3” model specification of anthropicThe

Both Opus 4 and Sonnet 4 are “hybrid” models, ethnographic, capable of investing and capable of extending to deep logic (the amount that AI “argument” and “thought” as people understand these ideas). As the logic mode is introduced, the models may take more time to consider possible solutions to the problem provided before answering.

As a reason for models, they will show a “user-friendly” summary of their thought process, anthropological. Why not show the whole thing? Partially to protect the ethnographic “competitive facilities”, the company acknowledged in a draft blog post provided to TechCrunch.

Oppus 4 and Sonnet 4 can use parallel in multiple equipment, such as search engines, and alternatives between logic and equipment to improve their answer value. They can more reliably handle the work “memories” information and store the information, which describes it as “transparent knowledge” as well as anthropological time.

To make models more programmer-friendly, anthropological upgrades are rolling upgrades in the aforementioned clad code. Claud code, which allows developers to run specific tasks through anthropological models directly from the terminal, now integrate with IDEs and provide an SD that allows the divs to connect it to third party applications.

The Claud Code announced at the beginning of this week enables SDK, enables to operate the clad code as a sub-process in supported operating systems, provides a way to create AI-powered coding assistants and tools that gain the capacity of clad models.

Anthropic has published Microsoft’s VS code, Jetbines and GitHub Clod Code extensions and connectors. GitHub connecting developers to respond to the reviewer’s response to reviewers, as well as trying to correct the errors – or otherwise correct – allow the code to be tagged.

AI models still strive to code quality software. Code-expression introduces the AI ​​protection weaknesses And DefectBecause Weakness In fields like programming logic to understand the logic. Nevertheless their promise to increase coding productivity – and developers – pushing Quickly accept themThe

Anthropologists, intensely aware of this, promising more frequent model updates.

“We are […] Transferring more frequently to model updates, providing an uninterrupted flow to the customers to bring a speedy power to the customers, “wrote the startup in the draft post.” We keep this approach to the cutting edge as well as refine and improve our models. “