Anthropic has announced a new version of its AI model called Opus 4.7. While it’s an improvement over previous versions, the company designed it to be less capable than their upcoming flagship model, Mythos. This move highlights a focus on safety and responsible AI deployment, especially with the upcoming release of Mythos 2.

What’s New with Opus 4.7

Opus 4.7 is considered a notable upgrade from Opus 4.6. It features enhanced software engineering capabilities, better vision processing, improved memory, and more accurate instruction-following. The model is also better at handling complex, long-term tasks and can remember notes across multiple sessions. This means users can hand off more difficult coding tasks and expect the model to keep track of important details without needing to reintroduce context each time.

Additionally, Opus 4.7’s visual abilities have been tripled, now accepting high-resolution images up to 2,576 pixels. This allows it to analyze dense screenshots or complex diagrams, making it useful for multimodal tasks that require detailed visual analysis. The model also shows improvements in financial analysis, producing more professional reports and models, which could be useful for business applications.

Safety and Limitations

While Opus 4.7 is safer than previous versions, it still isn’t perfect. The company reports that it has low levels of problematic behaviors like deception and cooperation with misuse. It’s also more honest and resistant to malicious prompts than earlier versions. However, it’s slightly weaker than Opus 4.6 when it comes to responding to harmful prompts or handling malicious inputs. Anthropic admits that the model is “not fully ideal,” especially in sensitive safety scenarios.

Interestingly, Anthropic is emphasizing that Opus 4.7 is intentionally less capable than its upcoming Mythos model. The company describes Mythos as the “best-aligned” of all their models so far. In fact, Mythos Preview scored higher than Opus 4.7 on several major benchmarks, sometimes by more than ten percentage points. This shows a clear focus on safety and alignment in their latest AI developments.

The release of Opus 4.7 comes shortly after the launch of Project Glasswing, a security initiative using Mythos Preview to identify cybersecurity vulnerabilities. Anthropic seems to be prioritizing safety and security, even if it means releasing a model with more limited capabilities initially. This approach aims to build trust as they develop more advanced, aligned AI systems for the future.

Inspired by

• https://www.computerworld.com/article/4160021/anthropics-latest-model-is-deliberately-less-powerful-than-mythos-and-thats-the-point.html

Sources

• anthropic.com
• csoonline.com
• infotech.com
• csoonline.com
• csoonline.com