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

From Thursday, anthropic, anthropic, starting claims in the point of view of at least in popular criteria, at least the best of the best in industry, began two new AI models.

Claude Opus 4 and Claude Sonnet 4, part of the new Claist 4th model family, can analyze large databases, according to the company, can analyze long-term assignments and sees complex measures. Both models were adjusted to perform well in programming tasks.

Users for both paying users and the company’s free chatbot applications will receive access to Opus 4 to pay only users.

Tokens are raw bits of information that AI models work. One million Token is equal to about 750,000 words – about 163,000 words are “war and reconciliation”.

Anthropic’s Claude 4 model comes as a significant increase in the company. It was reported to have somethingThe material set up by Ex-Openai researchers aims to earn $ 12 billion in 2027, $ 2.2 billion this year. Anthropical recently closed $ 2.5 billion loan rig and removed billions of dollars Amazon and Other investors Where he expects Rising costs associated with developing border models.

The opponents were not easy to protect the pole position in the AI ​​race. An anthropic launched a New Flagship AI model At the beginning of this year, Claude Sonnet 3.7, along with the agent coding, including Openai and Google, including rivals, competed to overturn the company with strong models and dev instruments.

Resistant with an anthropic cluster 4.

How skilled the two models applied today, Opus 4 can maintain a “focus” in many steps of a workflow. Meanwhile, Sonnet 4 – Sonnet 3.7 is designed for “drop-down replacement” – improves coding and math compared to previous models of anthropy, and according to the company, follows more accurate rules.

The Claude 4 family is less likely to deal with “premium hacking” less than Sonnet 3.7. The reward judge known as a special game is a behavior that models receive shortcuts and gaps to complete the tasks.

These improvements have not gave the world to be clear the best models by each criterion. For example, Opus 4 in case of beating Google Gemini 2.5 Pro and Openai O3 and GPT-4.1 In SWE Bench, which is designed to assess a model’s coding abilities, Multimodal Assessment MMMU or GPGA Diamond, Fizi-based Biology, Physics and Chemistry cannot check O3 O3.

Again, it is not under tighter guarantees, including anthropic, reus 4, including temporary harmful content detectors and cybersaluture protection. The company claims to be “significantly increased” to buy, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce, produce or buy nuclear weapons Anthropin’s “ASL-3” model specification.

“Think” and “Think” and “Think” and “Think”, “Think” and “Think” and “Think”, “Think” and “Think”, “Think”, “Think” and “Think”, “Think” and “Think”, “Think” and “Think”, “Think”, “Think” and “Think”, “Think”, “Think”, “Think”.

The reason for the models will show the “user-friendly” summary of their thoughts. Why don’t you show everything? Partial anthropic “Competitive Advantages” acknowledges the blogging writing project to the company to protect the company TechCrunch.

Opus 4 and Sonnet 4, like search engines, can use the alternative between justification and tools to use multiple tools in parallel and increase the quality of answers. In addition, the anthropic can produce and maintain the facts in “Memory” to manage the tasks more secure to reflect the “tacit knowledge” over time.

Models spread upgrades to the clod code mentioned above an anthopic to make more programmer friends. Claude code that allows developers to perform custom tasks directly through an anthropic models from a terminal, now combines with IDES and offers SDK, which allows you to combine with third party applications.

Announcement Claude Code SDK, Claude Code SDK, Clode Mode allows you to work as a subprocess as supported operating systems, which provides a way to prepare a way that works with AI, which has the capabilities of support.

Claude code expansions and connectors for an anthropic Microsoft vs code, JetBrains and Github. GitHub Connector allows developers to meet the Cloder code reviewer feedback, as well as to try to correct errors or try to change errors or change other ways.

AI models are still fighting the code quality program. Create Code AI tends to apply security vulnerabilities and faultFor example, weaknesses in areas such as ability to understand programming logic. Again, to increase coding productivity is to push promises and developers – for Accept them quickly.

Anthropically, it promises more frequent model updates on sharp informative, more often.

“We … more often, the flow of improvements that brought more customers to customers to better driving opportunities,” he wrote a start in his project. “This approach keeps you in the advanced place because we are the farthest and developing our models.”