Anthropic unveiled its latest generation of “frontier,” or cutting-edge, AI models, Claude Opus 4 and Claude Sonnet 4, during its first conference for developers on Thursday in San Francisco. The AI startup, valued at over $61 billion, said in a blog post that the new, highly anticipated Opus model is “the world’s best coding model,” and “delivers sustained performance on long-running tasks that require focused effort and thousands of steps.” AI agents powered by the new models can analyze thousands of data sources and perform complex actions.
The new release underscores the fierce competition among companies racing to build the world’s most advanced AI models—especially in areas like software coding—and implement new techniques for speed and efficiency, as Google did this week with its experimental research model demo called Gemini Diffusion. On a benchmark comparing how well different large language models perform on software engineering tasks, Anthropic’s two models beat OpenAI’s latest models, while Google’s best model lagged behind.
Some early testers have already had access to the model to try it out in real-world tasks. In one example provided by the company, a general manager of AI at shopping rewards company Rakuten said Opus 4 “coded autonomously for nearly seven hours” after being deployed on a complex project.
Dianne Penn, a member of Anthropic’s technical staff, told Fortune that “this is actually a very large change and leap in terms of what these AI systems can do,” particularly as the models advance from serving as “copilots,” or assistants, to “agents,” or virtual collaborators that can work autonomously on behalf of the user.
Claude Opus 4 has some new capabilities, she added, including following instructions more precisely and improvement in its “memory” capabilities. Historically, these systems don’t remember everything they’ve done before, said Penn, but “we were deliberate to be able to unlock long-term task awareness.” The model uses a file system of sorts to keep track of progress, and then strategically checks on what’s stored in memory in order to take on additional next steps—just as a human changes its plans and strategies based on real-world situations.
Both models can alternate between reasoning and using tools like web search, and they can also use multiple tools at once—like searching the web and running a code test.
“We really see this is a race to the top,” said Michael Gerstenhaber, AI platform product lead at Anthropic. “We want to make sure that AI improves for everybody, that we are putting pressure on all the labs to increase that in a safe way.” That includes exhibiting the company’s own safety standards, he explained.
Claude 4 Opus is launching with stricter safety protocols than any previous Anthropic model. The company’s Responsible Scaling Policy (RSP) is a public commitment that was originally released in September 2023 and maintained that Anthropic would not “train or deploy models capable of causing catastrophic harm unless we have implemented safety and security measures that will keep risks below acceptable levels.” Anthropic was founded in 2021 by former OpenAI employees who were concerned that OpenAI was prioritizing speed and scale over safety and governance.
In October 2024, the company updated its RSP with a “more flexible and nuanced approach to assessing and managing AI risks while maintaining our commitment not to train or deploy models unless we have implemented adequate safeguards.”
Until now, Anthropic’s models have all been classified under an AI Safety Level 2 (ASL-2) under the company’s Responsible Scaling Policy, which “provide[s] a baseline level of safe deployment and model security for AI models.” While an Anthropic spokesperson said the company hasn’t ruled out that its new Claude Opus 4 could meet the ASL-2 threshold, it is proactively launching the model under the stricter ASL-3 safety standard—requiring enhanced protections against model theft and misuse, including stronger defenses to prevent the release of harmful information or access to the model’s internal “weights.”
Models that are categorized in Anthropic’s third safety level meet more dangerous capability thresholds, according to the company’s responsible scaling policy, and are powerful enough to pose significant risks such as aiding in the development of weapons or automating AI R&D. Anthropic confirmed that Opus 4 does not require the highest level of protections, categorized as ASL-4.
“We anticipated that we might do this when we launched our last model, Claude 3.7 Sonnet,” said the Anthropic spokesperson. “In that case, we determined that the model did not require the protections of the ASL-3 Standard. But we acknowledged the very real possibility that given the pace of progress, near future models might warrant these enhanced measures.”
In the lead up to releasing Claude 4 Opus, she explained, Anthropic proactively decided to launch it under the ASL-3 Standard. “This approach allowed us to focus on developing, testing, and refining these protections before we needed. We’ve ruled out that the model requires ASL-4 safeguards based on our testing.” Anthropic did not say what triggered the decision to move to ASL-3.
Anthropic has also always released model, or system, cards with its launches, which provide detailed information on the models’ capabilities and safety evaluations. Penn told Fortune that Anthropic would be releasing a model card with its new launch of Opus 4 and Sonnet 4, and a spokesperson confirmed it would be released when the model launches today.
Recently, companies including OpenAI and Google have delayed releasing model cards. In April, OpenAI was criticized for releasing its GPT-4.1 model without a model card because the company said it was not a “frontier” model and did not require one. And in March, Google published its Gemini 2.5 Pro model card weeks after the model’s release, and an AI governance expert criticized it as “meager” and “worrisome.”
