Claude Opus 4.7 Goes Live, but Anthropic’s Most Powerful AI Still is “Mythos”

The truly most powerful LLM continues to be kept under wraps, but active users are still being taken care of: Anthropic today made its latest language model Claude Opus 4.7 generally available. The company positions it as the direct successor to Opus 4.6 and highlights improvements primarily in software development, image processing, and instruction adherence. At the same time, Anthropic also acknowledges weaknesses in its announcement.

How well Claude Opus 4.7 actually performs in practice remains to be seen. Neither arena.ai nor Artificial Analysis yet have test results and benchmarks for it to compare it with other LLMs available on the market.

What Anthropic advertises as strengths

Software development and complex tasks

According to Anthropic, Opus 4.7 has been particularly improved for demanding programming tasks. The company claims users can now hand off difficult coding tasks that previously required close supervision to the model with greater confidence. Opus 4.7 is said to handle complex, lengthy tasks with greater care and consistency, and to verify its own results before outputting them.

Improved image processing

Another advertised advancement concerns visual perception. Opus 4.7 is said to be able to process images with a resolution of up to 2,576 pixels on the long side, which Anthropic states is more than three times that of earlier Claude models. This is intended to enable applications that rely on fine visual details, such as analyzing complex diagrams or working with high-resolution screenshots.

Instruction adherence

Anthropic emphasizes that Opus 4.7 follows instructions significantly more literally and completely than its predecessor. However, the company explicitly points out that this can be a double-edged sword: prompts written for older models may produce unexpected results with Opus 4.7, because earlier models interpreted instructions more loosely or skipped parts of them.

Memory and long-term tasks

The model is said to be better at using file-system-based notes across multiple sessions. According to Anthropic, it can retain important information from previous sessions and use it for follow-up tasks, so that less context needs to be re-entered.

Acknowledged weaknesses and limitations

Anthropic describes Opus 4.7’s safety profile as similar to its predecessor, but identifies specific areas where the model performs worse. For instance, Opus 4.7 is reportedly more prone to providing excessively detailed harm-reduction guidance on controlled substances. The internal alignment assessment rated the model as “largely well-aligned and trustworthy, though not fully ideal in its behavior.”

“Largely well-aligned and trustworthy, though not fully ideal in its behavior” (Anthropic’s internal alignment assessment of Opus 4.7)

Anthropic also makes clear that Opus 4.7 is not the most capable model in its own portfolio. The model Claude Mythos Preview surpasses Opus 4.7 in overall performance and, according to Anthropic, also shows the lowest rates of misconduct in evaluations. Mythos Preview remains limited in its availability for the time being (more on that here).

Cybersecurity: deliberate limitations

In the context of the company’s own Project Glasswing, which deals with AI risks in the field of cybersecurity, Anthropic states it has actively sought to reduce Opus 4.7’s cyber capabilities compared to Mythos Preview. The model ships with automatic safety mechanisms designed to detect and block requests involving prohibited or high-risk cybersecurity applications. Security professionals can apply for a verification program to use the model for legitimate purposes such as penetration testing.

Technical and commercial details

Migration note

Anthropic points out that switching from Opus 4.6 to Opus 4.7 may be associated with higher token consumption. An updated tokenizer processes the same text using an estimated 1.0 to 1.35 times more tokens than before. In addition, the model produces more output tokens at higher effort levels. The company recommends measuring actual additional consumption using real traffic before carrying out a full migration.