Anthropic has launched Claude Opus 4.8, a new version of its flagship AI model that the company says is less likely to hide mistakes or make unsupported claims while performing complex tasks.
The upgrade builds on Claude Opus 4.7 and arrives as AI firms race to make autonomous systems more reliable for coding, research, and enterprise workflows. Anthropic said the model shows improvements across coding, reasoning, and agentic benchmarks while also becoming more transparent about uncertainty.
One of the biggest changes in Opus 4.8 is its focus on honesty during long-running tasks. AI models often present incorrect information confidently or claim progress without enough evidence. Anthropic said the new model is better at flagging uncertainty and identifying flaws in its own outputs instead of silently passing errors through.
The company said internal evaluations showed Opus 4.8 was “around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.”
Anthropic said early testers found the model more reliable when handling agentic tasks, where AI systems independently plan and execute actions over multiple steps.
The company also highlighted improvements in alignment and safety behavior. According to Anthropic, its alignment team concluded that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.”
The assessment also found lower rates of misaligned behavior, including deception and cooperation with misuse, compared with Opus 4.7.
Alongside the model upgrade, Anthropic introduced new features aimed at expanding how Claude handles large-scale coding and reasoning tasks.
One of them is a research preview feature called Dynamic Workflows for Claude Code. The system allows Claude to break large tasks into smaller jobs handled by hundreds of parallel AI subagents operating within a single session.
Anthropic said the feature can carry out codebase-scale migrations involving hundreds of thousands of lines of code while checking outputs against existing test suites before reporting results back to users.
The company also added an effort control setting on claude.ai and Cowork. Users can now decide how much computational effort the model spends on a task.
Lower effort settings prioritize faster responses and reduced token use, while higher settings allow the model to spend more time reasoning through difficult prompts. Anthropic said Opus 4.8 defaults to a high-effort mode designed to balance quality and user experience.
The company additionally reduced pricing for its fast mode, which now runs at 2.5 times the speed of earlier models while costing less than previous versions.
Anthropic said it is also preparing more advanced “Mythos-class” models under Project Glasswing. These systems are currently being tested for cybersecurity applications with a small group of organizations before broader release.
The company said stronger cyber safeguards are still being developed before those models can become widely available.
Claude Opus 4.8 is now available through claude.ai and the Claude API.
With over a decade-long career in journalism, Neetika Walter has worked with The Economic Times, ANI, and Hindustan Times, covering politics, business, technology, and the clean energy sector. Passionate about contemporary culture, books, poetry, and storytelling, she brings depth and insight to her writing. When she isn’t chasing stories, she’s likely lost in a book or enjoying the company of her dogs.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。