Anthropic releases Claude Opus 4, but says it's less capable than Mythos

Anthropic has officially launched Claude Opus 4.7 on Wednesday, April 15.

The latest AI model has introduced major improvements in software engineering and instruction-following.

However, it deliberately stayed back at cybersecurity capabilities as part of the new safety safeguards.

The American artificial intelligence company stated that the model possesses “notable improvement” over Opus 4.6, specifically on complex coding tasks that earlier needed close human supervision.

Earl testers' reports confirm that the hardest programming work can easily be given to Opus 4.7 with confidence.

The Opus 4.7 model also includes improved visual capabilities, which can now handle high-definition pictures that measure up to 2,576 pixels across the long side – almost three times more powerful than the Claude models before it.

But, it is made clear that the newly released model is “less broadly capable” than Anthropic’s most capable model, Claude Mythos Preview, which is not released publicly due to cybersecurity concerns.

The Project Glasswing flags risks of AI in cybersecurity. The firm promised that testing for security systems would be performed on weaker models first. 

The Opus 4.7 model is one such machine, with cyber abilities consciously weakened to below those of the Mythos. In terms of its vulnerability replication, Opus fell from 73.8% to 73.1%.