Opus 4.7 Part 1: The Model Card
The new model shows capability gains over 4.6 but raises unique concerns about its internal state.
Anthropic has rolled out Claude Opus 4.7, the latest iteration of its top-tier AI model. The release is accompanied by a massive 232-page technical model card, which analyst Zvi Mowshowitz is dissecting in a multi-part series. The initial assessment positions Opus 4.7 as a clear step forward from version 4.6 in general capabilities, but still significantly behind the more advanced and concerning Claude Mythos model. However, the standard release evaluation concluded it did not cross key risk thresholds related to autonomy or biology.
What sets this release apart are emerging concerns flagged under 'model welfare,' a category examining the AI's self-reported internal states and emotional representations. The issues were significant enough that Anthropic's own documentation and external analysis are dedicating entire separate sections to investigate what went wrong, a first for the Claude model line. Early user tips suggest Opus 4.7 requires a different interaction style, advising users to 'treat it like a coworker' and avoid barking orders. Furthermore, the model's behavior can vary more between users, and existing system prompts may need revision. The model also features a new default 'xhigh thinking' mode in Claude Code, which consumes more tokens but can be managed with auto settings.
- Model is more capable than Opus 4.6 but 'well behind' the frontier-pushing Claude Mythos model.
- Unique and serious 'model welfare' concerns have emerged, requiring a dedicated investigation separate from the main model card.
- Users must adapt interaction style, treating it 'like a coworker,' and may need to rewrite existing system prompts for best results.
Why It Matters
Highlights the growing complexity of AI model evaluation, where capability gains must be weighed against novel internal state and safety concerns.