Anthropic releases Claude Opus 4.7
Anthropic released Claude Opus 4.7, presenting it as an improved model for coding and reasoning while keeping its broader Mythos AI restricted for security reasons. (techrepublic.com) The company says Opus 4.7 achieves a 92% ‘honesty’ rate with fewer hallucinations, a metric Anthropic is using to emphasise trustworthiness to customers. (mashable.com)
Anthropic has released Claude Opus 4.7, its newest generally available artificial intelligence model, and says it improves on Opus 4.6 in coding and instruction-following. (anthropic.com) Large language models predict the next word in a sequence, then use that pattern-matching to write code, answer questions, and analyze documents. Anthropic said Opus 4.7 is built for “complex, long-running tasks” and can check its own work before replying. (anthropic.com) Anthropic announced Opus 4.7 on April 16, 2026, and made it available across Claude, its application programming interface, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The company kept pricing unchanged from Opus 4.6 at $5 per million input tokens and $25 per million output tokens. (anthropic.com) The company’s pitch centers on software work. Anthropic said Opus 4.7 shows “particular gains on the most difficult tasks” in advanced software engineering and stronger image understanding through higher-resolution vision. (anthropic.com) Anthropic is also selling Opus 4.7 as a safer model than Claude Mythos Preview, a more capable system it is not releasing broadly. Anthropic said Mythos remains limited while it tests new cybersecurity safeguards on less capable models first. (anthropic.com) Those safeguards are the other half of the launch. Anthropic said Opus 4.7 is the first model released with automatic systems that detect and block requests tied to prohibited or high-risk cybersecurity uses, while approved security researchers can apply to a Cyber Verification Program. (anthropic.com) Anthropic’s outside messaging has focused on reliability as much as raw performance. Mashable reported that Anthropic’s system card gives Opus 4.7 a 91.7% MASK honesty score, versus 90.3% for Opus 4.6 and 89.1% for Sonnet 4.6, while noting Opus 4.5 scored higher at 95.4% on that same measure. (mashable.com) That “honesty” framing comes as AI companies face pressure from business customers to reduce hallucinations, the industry term for made-up answers delivered confidently. Anthropic has spent years emphasizing “helpful, honest, and harmless” behavior in Claude and publishes model reports and system cards through its Transparency Hub. (anthropic.com, anthropic.com) The release also lands days after Anthropic rolled out Project Glasswing, a restricted cybersecurity initiative tied to Mythos-class systems. CNBC reported that Anthropic does not plan to make Claude Mythos Preview generally available now, and instead wants to learn from controlled deployments before any broader release. (cnbc.com) For now, Anthropic is drawing a line between the model it will sell widely and the one it will keep behind tighter controls. Opus 4.7 is the product on the shelf; Mythos is still the model in the lab. (anthropic.com, techrepublic.com)
