April 17th message, last nightAnthropic Official Release Claude Opus 4.7 Flagship model, priced in line with previous generation Opus 4.6: $5 per million input token and $25 per million output, which developers can call through claude-opus-4-7。
It was described that Opus 4.7 was able to maintain rigour and consistency in complex missions that operated over a long period of time and had the capability to validate the output in advance of the return. In Cursor's internal assessment, the pass rate for Opus 4.7 on CursorBench was 70%, whereas Opus 4.6 was 58%。
With respect to multi-modular capabilities, Opus 4.7 supports high-resolution image input with maximum edges of 2576 pixels (approximately 3.75 million pixels), more than three times the previous Claude model, opening up space for fine visual scenes such as computer manipulation of intelligent bodies, complex graphic data extraction, etc。
In terms of security and alignment, Opus 4.7 enhanced the interception of high-risk network security requests by limiting relevant capabilities and placing protection layers in training. To this end, Anthropic has introduced a "Network Security Verification Program" which is limited to professionals applying for compliance。
Anthropic also acknowledges that, despite the progress of Opus 4.7 on honesty and anti-infliction attacks, there has been a slight decline in recommendations for the management of substances, whose overall alignment is still second to Mythos Preview。