Follows instructions "substantially better": Anthropic releases Opus 4.7
The AI model Claude Opus 4.7 is said to follow instructions better and, above all, literally. The predecessor is surpassed everywhere, but not the best AI model.
(Image: Stockinq / Shutterstock.com)
Anthropic has released its latest AI model, Claude Opus 4.7, which is reportedly only surpassed by the closed-source Claude Mythos Preview in various benchmarks. Compared to the predecessor released at the beginning of February, the update represents a “notable improvement,” promises the AI company. Users can now confidently hand over even their most difficult coding tasks to the technology, as stringent control is no longer necessary. The AI model handles “complex, long-running tasks with rigor and consistency,” follows instructions precisely, and develops methods to check results before output. It can also process images in higher resolution.
Caution with old prompts
As central improvements, Anthropic points out that Claude 4.7 is “substantially better” at following instructions. It is interesting that the AI model may now show unexpected results with prompts that previously worked: “Opus 4.7 takes the instructions literally.” Therefore, such instructions should be reviewed. The new AI model is also better at analyzing financial data, can create more professional presentations, and better interlink individual tasks. Additionally, the updated Opus can utilize file system-based memory more efficiently. This requires less context information over multiple sessions.
Regarding the release of the new AI model, Anthropic also points out that Opus 4.7 can process text better. However, this affects token consumption, meaning up to a third more tokens may be required depending on the input. Improved performance in other tasks can also lead to higher token usage than before. Users can control this, for example, “by prompting the model to be more concise.” In a guide, those responsible have compiled such and other tips for migration, which are intended to ensure that token consumption is ultimately reduced.
Videos by heise
The release of Opus 4.7 comes just over a week after the introduction of an AI model that is said to be so dangerous that it cannot be made public. Primarily because Claude Mythos Preview is said to be unparalleled in finding and exploiting security vulnerabilities in software, the model is exclusively made available to companies working on IT security. In a list of benchmark results, Anthropic now shows that Mythos beats Opus 4.7 in all areas, but sometimes only by a narrow margin. In reproducing security vulnerabilities, Opus 4.7 is even slightly worse than its predecessor.
(mho)
