Anthropic has launched Claude Opus 4.7, a significant upgrade to its flagship artificial intelligence (AI) model amid growing user frustration, in a bid to solidify its position in the competitive agentic economy.
The release features a dramatic leap in computer vision and coding capabilities, even as the company continues to withhold its most powerful technology from the public.
The rollout follows weeks of community backlash. Users, including high-level engineers at firms like AMD Inc., had recently taken to GitHub and social media to complain that the previous iteration, Opus 4.6, had “regressed” or been “nerfed.”
Anthropic denied claims that it had intentionally scaled back performance to save on computing costs, instead positioning Opus 4.7 as a direct answer to those reliability concerns.
The most striking improvement in Opus 4.7 is sensory perception. Internal benchmarks and early developer reviews highlight a surge in visual acuity, jumping from 54.5% in the previous version to 98.5%. This allows the model to process high-resolution images and complex documents with far fewer reasoning errors.
In the coding arena, Anthropic is introducing “extra high” (xhigh) reasoning levels, a new setting that allows developers to prioritize deep logic over speed.
The model achieved an 87.6% score on the SWE-bench verified agentic coding test. Anthropic claims the model is more “tasteful,” producing higher-quality slide decks and professional interfaces.
Benchmarks show Opus 4.7 outperforming OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro in financial analysis and long-horizon autonomous tasks.
Despite the gains, Opus 4.7 is not Anthropic’s strongest offering. The company took the unusual step of publicly conceding that the new model still trails Mythos, a frontier system currently restricted to a handpicked group of cybersecurity firms under Project Glasswing.
Anthropic executives maintain that Mythos is “too good” at cybersecurity, capable of exposing vulnerabilities so rapidly that a broad release would pose a systemic risk. Companies like Cisco Systems Inc. and CrowdStrike Holdings Inc. are currently using the restricted model to patch holes that Mythos itself identified.
“What we learn from the real-world deployment of [Opus 4.7] safeguards will help us work towards our eventual goal of a broad release of Mythos-class models,” the company said in a statement, suggesting Opus 4.7 serves as a live testing ground for the guardrails needed to eventually leash its more powerful sibling.
While Opus 4.7 leads in agentic reliability, it is not a “clean sweep” for Anthropic. Competitors still hold the edge in specific niches; for instance, OpenAI’s GPT-5.4 maintains a lead in agentic search, scoring 89.3% against Opus 4.7’s 79.3%.
The release puts further pressure on OpenAI, whose upcoming model, codenamed Spud, remains shrouded in what the industry calls “vagueposting” by CEO Sam Altman.
For now, Anthropic is betting that professional-grade reliability and improved task budgets will keep developers in its ecosystem.
Claude Opus 4.7 is available immediately on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, with API pricing remaining steady at $5 per million input tokens and $25 per million output tokens.
