Anthropic Revolutionizes with Claude Opus 4.7, Surpassing GPT-5.4

A New Milestone for Anthropic with Claude Opus 4.7

On April 16, 2026, Anthropic unveiled its most advanced artificial intelligence model to date, Claude Opus 4.7. This model stands out for its exceptional performance on the SWE-bench Pro benchmark, where it achieved a score of 64.3%, surpassing competitors like GPT-5.4 with 57.7% and Gemini 3.1 Pro with 54.2%. Claude Mythos, on the other hand, reached a score of 77.8% on the same benchmark, underscoring its position as the internal leader at Anthropic. Claude Opus 4.7 also serves as a testing platform for the Glasswing project, which explores AI applications in cybersecurity.

Claude Opus 4.7 focuses on complex tasks and agentic use, although it is not as powerful as Claude Mythos Preview, Anthropic's internal version. The model is capable of handling complex programming projects, strictly following instructions, and verifying its own results to minimize errors. Image analysis has been enhanced, allowing for better accuracy and creativity in professional contexts.

Benchmark Results

Opus 4.7 has demonstrated its superiority over its predecessor Opus 4.6 in various tests. On the SWE-bench Pro, it scored 64.3%, compared to 53.4% for Opus 4.6. For bug resolution on SWE-bench Verified, Opus 4.7 achieved 87.6%, while Opus 4.6 scored 80.8% and Mythos reached 93.9%. On Finance Agent v1.1, Opus 4.7 scored 64.4%, surpassing Opus 4.6 with 60.1%, GPT-5.4 with 61.5%, and Gemini 3.1 Pro with 59.7%.

Opus 4.7 also improves memory management through a file system, allowing the model to retain notes over the long term and use them to chain tasks with less need for context re-establishment.

Pricing and Efficiency

The cost of using Claude Opus 4.7 remains at $5 per million input tokens and $25 for output. However, the model employs a new token slicing system, allowing up to 1.35 times more tokens to be associated with the same content. While this may lengthen responses, Anthropic recommends optimizing prompts for effective use.

Opus 4.7 "thinks" longer at high effort levels, which can extend responses and potentially increase consumption, even at a constant rate.

Cybersecurity and the Glasswing Project

As part of the Glasswing project, Claude Opus 4.7 tests cybersecurity safeguards, although its capabilities in this area are intentionally limited compared to Claude Mythos Preview. Anthropic clarified that Opus 4.7's cybersecurity capabilities are deliberately less advanced than those of Mythos Preview, with mechanisms in place to detect and block high-risk requests.

Anthropic emphasizes that Opus 4.7 is not a model dedicated to cybersecurity and remains behind Mythos Preview for these uses. Feedback from this experience is expected to broaden access to Mythos models in the future.

Claude Opus 4.7 is now accessible via Claude interfaces, the API, as well as on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. This launch comes as OpenAI recently introduced GPT-5.4-Cyber, focused on cyber defense.