Anthropic Unveils Claude Opus 4.8: The AI That Acknowledges Its Mistakes

Anthropic recently launched Claude Opus 4.8, the latest version of its artificial intelligence model, which stands out for its ability to recognize and report its own errors. This model, which performs better on Anthropic's internal benchmarks, is offered at the same price as its predecessor, Claude Opus 4.7.

Enhanced Performance and Increased Honesty

Claude Opus 4.8 shows notable improvements in several areas. In autonomous coding, it achieves a score of 69.2%, surpassing the 64.3% of the previous version. When it comes to autonomously operating a computer, it scores 83.4%. In automated financial analysis, Opus 4.8 scores 53.9%. However, in terminal programming tasks, it still lags behind OpenAI's GPT-5.5, with a score of 74.6% compared to 78.2%.

Before launching Opus 4.8, the California-based firm subjected the model to its alignment tests to ensure it behaves well and in the interest of its users. The model reaches new heights in responsible behavior, and problematic behaviors, such as lying or being misled for malicious purposes, are significantly less frequent than in Opus 4.7.

One of the major advancements of Opus 4.8 is its ability to report its uncertainties. Unlike many AI models that exhibit excessive confidence, Opus 4.8 is four times less likely to overlook errors without mentioning them. This feature has been highlighted by early testers, according to Anthropic.

New Features and Stable Pricing

In addition to its improved performance, Opus 4.8 introduces new features. A preview feature called "dynamic workflows" in Claude Code allows users to break down complex tasks into multiple parallel subtasks, thus speeding up the development process. On the claude.ai platform, a slider now enables users to choose the depth of reasoning for each response.

Despite these improvements, the cost of use remains stable: $5 per million words for input and $25 for output. The fast mode, which speeds up responses by 2.5 times, is now three times cheaper than before.

Towards an Even More Ambitious Future

Anthropic is not stopping there and is already preparing the next generation of AI models. The "Glasswing" project offers limited access to "Mythos Preview," an even more powerful model currently in testing for cybersecurity applications. This model, deemed too advanced for immediate release, is undergoing rigorous checks before a broader rollout planned in the coming weeks.