[DigitalToday intern reporter Seung-ah Yoo] Anthropic has unveiled its new artificial intelligence (AI) model, Claude Opus 4.8, and launched it worldwide.

On May 28 (local time), foreign media including IT outlet 9to5Mac reported the update comes about 6 weeks after Opus 4.7 was unveiled on April 16.

The core of this version is performance improvements and adjustments to how it responds. Anthropic said Claude Opus 4.8 was designed to serve as a more effective collaborator than the previous model. The company said it refined judgment based on Opus 4.7, responds more honestly about work progress and strengthened the ability to work independently for long periods.

Anthropic said early tests showed Claude Opus 4.8 signals uncertainty more often during tasks and reduced the frequency of unsupported claims. "The new model showed higher reliability and sharper judgment, and its tendency to make claims with insufficient factual basis also decreased," the company said.

In particular, Claude Opus 4.8 is a quarter as likely as Opus 4.7 to miss defects in code it generates itself. That amounts to an improvement aimed at raising error-detection accuracy in development environments with extensive code review and iterative revisions.

There were also changes in alignment assessments. Anthropic said Claude Opus 4.8 set a new best level on socially desirable trait indicators, and the rate of misaligned behavior fell sharply versus Opus 4.7. It said it improved controllability and safety while strengthening task performance.

It also released benchmark figures. Agentic coding scores rose to 69.2 percent from 64.3 percent, and tool-use-based multi-domain reasoning increased to 57.9 percent from 54.7 percent. Agentic computer use improved to 83.4 percent from 82.8 percent, and knowledge work scores rose to 1,890 from 1,753. Agentic financial analysis also climbed to 53.9 percent from 51.5 percent.

Its pricing policy remained unchanged. Anthropic said it will keep the same prices for Opus 4.7 and Opus 4.8. Instead, it adjusted the performance and cost structure of fast mode. Standard pricing is $5 per 1 million input tokens and $25 per 1 million output tokens. A high-speed mode for Opus 4.8 that runs 2.5 times faster is priced at $10 per 1 million input tokens and $50 per 1 million output tokens. The company said fast mode is about 2.5 times faster than before and costs have fallen to about one-third.

It also revamped its developer environment, Claude Code. Opus 4.8 runs by default at a high level of computational effort. Anthropic said the setting delivers higher performance while using a similar number of tokens to Opus 4.7 on coding tasks. It plans to raise Claude Code usage limits in line with additional performance and maximum performance settings.

It also added features aimed at large-scale work. A dynamic workflow offered as a research preview supports Claude Code in handling larger-scale tasks. Claude Cowork and claude.ai also added controls to adjust the computational effort level used for responses next to the model selector.

The Messages API was also changed. The structure was revised so it can receive system items inside message arrays, allowing developers to modify instructions mid-task without going through a user turn. At the same time, it can maintain prompt caching and routing functions.

Anthropic also disclosed follow-up plans. The company said it is preparing a low-cost model with similar performance, and plans to more broadly release a Mythos-class model within the next few weeks after applying strengthened cyber safeguards based on Project Glasswing.

As competition with OpenAI intensifies, this update is seen as a move focused on improving real-world performance. It highlights coding error detection, long-duration autonomous work and honesty in responses, in an atmosphere of accelerating competition to secure developers and enterprise customers.