Anthropic Revolutionizes AI with Claude Opus 4.8, Focused on Honesty

Anthropic, a pioneering company in the field of artificial intelligence, has recently launched its latest language model, Claude Opus 4.8. This model stands out for its emphasis on honesty, caution, and better adaptation to complex coding projects. Unlike previous versions, Opus 4.8 promises more honest AI responses, which could address a modern quest for digital integrity.

A More Honest Model

According to Anthropic, Claude Opus 4.8 is designed to reduce unfounded claims and inform users when it is uncertain about an answer. The company claims that this model is four times less likely than its predecessor to let errors slip through in the generated code. This improvement in honesty is a significant step towards more reliable interactions with AI.

Improvements in Honesty

Anthropic reports that Opus 4.8 is less likely to make unfounded claims. It is also more inclined to inform you when it is unsure about an answer. "This is confirmed in our evaluations, which show that Opus 4.8 is about four times less likely than its predecessor to let defects slip through in the code it has written," the company clarified.

In Claude Code, I found that Opus 4.7 represented a substantial improvement over 4.6. While 4.6 often misinterpreted instructions or provided erroneous results, Opus 4.7 regularly tells me when its initial approach to a problem has failed and adopts a different tactic. Recent projects have shown a much higher degree of understanding than with 4.6. Given the quality leap from 4.6 to 4.7, which was subjectively very noticeable over many sessions, I hope we will see the same in the transition from 4.7 to 4.8.

Enhanced Judgment

This seems to be the case, at least according to Tom Pritchard, an engineer at Spotify, who has already tested Opus 4.8. "Claude Opus 4.8 has significantly better judgment. In Claude Code, it asks the right questions, detects its own mistakes, questions a plan that isn't solid, and builds its confidence around complex explorations before making major changes. It's an excellent model for building," he stated in the blog post.

Effort and Performance

Claude Code has had the ability to define effort since at least 4.7. Effort is essentially a measure of the power the model dedicates to a problem, measured in tokens. In Opus 4.8, the default effort parameter, which is high, produces what the company describes as "the best overall balance between quality and user experience." In coding tasks, this parameter spends a similar number of tokens as the default level proposed in Claude Code Opus 4.7, but with better performance.

This effort capability is now integrated into Claude.ai and Cowork. With higher effort settings, Claude "will think more frequently and deeply." With a lower effort setting, Claude responds more quickly, and users will find that their experiences with the AI are less constrained.

Dynamic Workflows

At launch, this feature was not fully defined, but it is intriguing. Launched as a research preview, Opus 4.8 can schedule work, execute hundreds of sub-agents in parallel within a single session, and verify results before reporting them. This feature is designed for very large-scale tasks. The example given by Anthropic was codebase migrations involving hundreds of thousands of lines. It seems that Claude can generate and manage the workflow as the task evolves. Rather than following a fixed plan, agents can change their priorities and tasks based on what they discover while working.

Anthropic stated that the sub-agents verify their results before reporting them to users. If Claude coordinates hundreds of sub-agents, it is essential that it detects uncertainty, faulty assumptions, and failed outcomes. This directly ties back to the honesty claims discussed at the beginning of the article. If Claude is to launch "thousands of agents," obtaining reliable and verified results is crucial, as there is no way that human oversight could keep up alone.

Pricing and Availability

Anthropic announced that Opus 4.8 is available everywhere since Thursday via Claude and the Claude API under the name claude-opus-4-8. In practice, especially if you are using Claude Code, you may find that you need to restart your session or wait a day or two for Claude Code to recognize it. When Anthropic transitioned from Opus 4.6 to 4.7, I kept asking Claude Code which model it was using, and it was only the next day that it stopped reporting Opus 4.6 and began reporting Opus 4.7.

Overall pricing has not changed since Opus 4.7. The token-based pricing remains at $5 per million input tokens and $25 per million output tokens. The company also indicated that the fast mode, which allows the model to operate at 2.5 times the speed of normal mode, will be "three times cheaper than for previous models."