Anthropor revealed its last generation of “borders”, or advanced models, Claude OPUS 4 and Claude Sonnet 4, during its first conference on developers on Thursday in San Francisco. AI Startup, estimated at more than $ 61 billion, said in a blog that the new OPUS model is very expected is the “best coding model in the world”, and “offers sustainable performance in long -term tasks that require focused effort and thousands of steps.” Artificial intelligence agents supported by new models can analyze thousands of data sources and make complex procedures.
The new version emphasizes fierce competition between companies that are racing to build the most advanced artificial intelligence models in the world – especially in areas such as program coding – and implement new technologies for speed and efficiency, such as Google Has the display of the experimental research form called this week called The spread of Gemini. On a standard that compares the extent to which large language models perform in software engineering tasks, modeling in Antarbur defeated the latest OpenAi models, while the best Google model left behind.
Some of the first laboratories have been able to reach the model for his experience in realistic tasks. In one of the examples presented by the company, the AI’s general manager at Rakuten Honk Bonuses said that OPUS 4 “is independently encrypted for about seven hours” after it was published in a complex project.
Diane Ben, a member of the Antarbur technical staff, told Antarbur, luck “This is in fact a great change and leaf in terms of what these artificial intelligence systems can do,” especially with the progress of models from work as “Copilots”, or assistants, to “agents”, or virtual collaborators who can work independently on the user.
She added that Claude Obus 4 has some new capabilities, including the following guidelines more accurately and improving her “memory” capabilities. Historically, these systems do not remember everything they had done before, but Pennsylvania, but “we were deliberate to be able to open awareness of the task in the long run.” The model uses a type file system to track progress, and then achieves strategically about what is stored in memory in order to take additional additional steps-completely as a person changes his plans and strategies based on the positions of the real world.
Both models can alternate between thinking and use tools like Search Web, and they can also use multiple tools simultaneously – such as searching in the web and running a code test.
“We really see that this is a race to the top,” said Michael Grestenhaber, who leads the product of the Antarbur AI platform. “We want to make sure that artificial intelligence improves for everyone, and that we press all laboratories to increase this in a safe way.” He explained that this includes displaying the company’s safety standards.
CLAUDE 4 OPUS is launched with stricter safety protocols than any previous anthropier model. RSP’s accountability policy (RSP) is a general commitment that was originally released in September 2023 and stressed that Antarbur will not “will not” train or publish models capable of causing catastrophic damage unless we implement safety and security measures that will maintain the risks less than acceptable levels. “The anthropologist was established in 2021 by former Openai employees who were concerned about that. Openai was a priority for speed and size for safety and governance.
In October 2024, the company updated its RSP with “a more flexible and accurate approach to evaluating and managing the risk of artificial intelligence while maintaining our commitment to not training or publishing models unless we implement sufficient guarantees.”
To date, all anthropor models have been classified as part of the level of the ASSL-2 intelligence safety according to the company’s liability policy, which “(s) provides a basic level of safe publication and typical security of artificial intelligence models.” While an anthropologist said that the company did not rule out that the new Claude Obus could fulfill the ASSL-2, it is proactively launched the model under the most striking ASL-3 safety standard-which requires improved protection against the stealing of the model and misuse, including the strongest defenses to prevent the launch of harmful information or access to the “on weights”.
Classifying models at the third safety level of anthropor meet the sills of the most dangerous ability, according to the company’s responsive scaling policy, which are strong enough to form significant risks such as helping to develop weapons or AI R&D automation. Anthropor confirmed that OPUS 4 does not require the highest level of protection, classified as the ASL-4.
“We expected to do this when we launched our last model, Claude 3.7 Sonnet,” said Antarbur spokesman. “In this case, we decided that the model does not require the protection of the ASL-3 standard. But we have recognized the very real possibility that the pace of progress, the future models in the future may call for these reinforced measures.”
She explained in the period leading up to the release of Claude 4 Obus, the anthropologist decided proactively to launch it under the ASL-3 standard. “This approach allowed us to focus on developing this protection, testing and improving it before need. We have excluded that the model requires ASL-4 guarantees based on our test.” Anthropor did not mention what sparked the decision to move to the ASL-3.
Antarbur has always released a model or system, cards with its launch operations, which provide detailed information on the capabilities of models and safety assessments. Bin said luck This Antarbur will launch a typical card with its new launch of the OPUS 4 and Sonnet 4, and an official spokesman confirmed that it will be released when the model is launched today.
Recently, companies including Openai and Google have been delayed. In April, Openai was criticize For its GPT-4.1 version without a model card because the company said it was not a “border” model and did not require a model. In March, Google Gemini 2.5 Pro Model Card published weeks after the form of the form, and an artificial intelligence governance expert criticize It is “meager” and “worrying”.
This story was originally shown on Fortune.com
https://fortune.com/img-assets/wp-content/uploads/2025/05/GettyImages-2154161015_a08d78.jpg?resize=1200,600
Source link