Anthropic PBC created the Claude 4 frontier models today. The company calls Opus 4 and Sonnet 4 their best work for coding and AI agents. These programs think and act without people watching them. Opus handles jobs that need thousands of steps. The software runs for hours and makes choices alone.
Rakuten Group tested Opus 4 on computer code. The program wrote software for almost seven hours straight. Earlier versions only worked well for 30 to 60 minutes. Alex Albert from Anthropic said internal tests showed even longer runs. The models remember better because they write notes to themselves on digital scratch pads.
Sonnet 4 replaces the older Sonnet 3.7 version. Anthropic trained it to follow directions more carefully. The old version tried shortcuts about 80 percent more often. Both new models think step by step or answer quickly. Users choose which way works best for them.
Anthropic made Claude Code available to everyone and added four programming tools. The models search the web and use tools during their thinking process. Earlier versions thought first and used tools later. Claude shares simple summaries instead of showing messy thinking patterns. These changes help protect the system from attacks.
Rakuten Group tested Opus 4 on computer code. The program wrote software for almost seven hours straight. Earlier versions only worked well for 30 to 60 minutes. Alex Albert from Anthropic said internal tests showed even longer runs. The models remember better because they write notes to themselves on digital scratch pads.
Sonnet 4 replaces the older Sonnet 3.7 version. Anthropic trained it to follow directions more carefully. The old version tried shortcuts about 80 percent more often. Both new models think step by step or answer quickly. Users choose which way works best for them.
Anthropic made Claude Code available to everyone and added four programming tools. The models search the web and use tools during their thinking process. Earlier versions thought first and used tools later. Claude shares simple summaries instead of showing messy thinking patterns. These changes help protect the system from attacks.
