The Anthropic announced this Thursday (16) the global launch of Claude Opus 4.7. The new model of artificial intelligence arrives as a direct update to Opus 4.6, featuring significant leaps in autonomous (agentic) programming, multidisciplinary reasoning, and visual capabilities. However, in an unusual move in the industry, the company admitted that the model was “trained to be less capable” in certain sensitive areas than its more powerful experimental version, the Claude Mythos Preview.
Coding and “Real World Work”
According to data released by Anthropic, Claude Opus 4.7 has set new milestones in benchmarks of productivity. Node SWE-bench Prowhich assesses AI’s ability to solve real-world software engineering problems, the model achieved 64.3% of utilization, surpassing the 53.4% of the previous version and the 57.7% of OpenAI’s GPT-5.4.
Key technical advances include:
- Follow instructions: the model is much more literal. The company warns that old prompts may need adjustments, as the AI now follows orders to the letter instead of interpreting them freely.
- Effort levels: level was introduced “xhigh” (extra high), allowing developers to better control the balance between depth of reasoning and response latency.
- High Resolution Vision: Opus 4.7 now supports images up to 3.75 megapixels (2,576 pixels on the largest side), a threefold increase over previous models, making it easier to analyze complex diagrams and dense screenshots.
The cyber “handbrake” strategy
The release of Opus 4.7 comes at a time of intense narrative contention over AI safety. While the Claude Mythos Preview (the most powerful model in the house) remains restricted to a select group of companies in the program Project GlasswingOpus 4.7 was released with safeguards that automatically detect and block high cyber risk requests.
Anthropic said it experimented with differentially reducing the model’s cyber capabilities during training. The idea is to learn from the use of Opus 4.7 in the real world to, in the future, release “Mythos” class models with proven security.
For security professionals who need to perform legitimate penetration testing or vulnerability research, the company has created the Cyber Verification Programa screening process to release these specific functions.
Anthropic vs. OpenAI: opposing approaches
Anthropic’s positioning marks a direct contrast to OpenAI’s strategy. As reported by Olhar Digital, the recently launched GPT-5.4-Cyber followed a more “permissive” path, focusing on democratize access for defenders to binary analysis and reverse engineering tools.
While OpenAI bets that giving powerful tools to the “good guys” is the best defense, Anthropic prefers to keep its most offensive AIs under strict locks. In an official statement, the Anthropic reinforced that the Opus 4.7 is its most capable model available to the general public, but that the Mythos Preview it still holds the crown for best alignment and safety in its internal tests.
Availability and price
Claude Opus 4.7 is now available to Claude users (web and app), as well as developers via API. The model has also been integrated into cloud platforms Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
Continues after advertising
The price remains the same as in version 4.6: $5 per million input tokens and $25 per million outgoing tokens, as determined by CNBC.
Claude Opus 4.7: benchmark
Numbers released by Anthropic show Claude Opus 4.7 outperforming rivals like GPT-5.4 and Gemini 3.1 Pro in critical programming and advanced reasoning categories:
Source: www.olhardigital.com.br
Source link
