Claude Opus 4.8: Anthropic Focuses on Honesty and Accuracy

Launch of Claude Opus 4.8: A Measured Advancement

Anthropic has recently introduced Claude Opus 4.8, a new version of its artificial intelligence model. In the official announcement, the company described this update as a modest yet tangible improvement over its predecessors. This honest approach is noteworthy in an industry often marked by hyperbolic announcements. Anthropic also expressed its commitment to developing models that offer similar capabilities at a reduced cost, thereby emphasizing their strategy to make AI more accessible.

Honesty is a central theme of this update. Anthropic highlighted the improvement in the honesty of Claude Opus 4.8, a crucial aspect for the reliability of AI models. AI models often tend to assert conclusions without sufficient evidence. However, early feedback from testers indicates that Claude Opus 4.8 is more cautious, more frequently signaling its uncertainties and avoiding unfounded claims. This improvement is reflected in the evaluations, where the model is about four times less likely to overlook errors in code without reporting them.

Performance and Technical Features

Claude Opus 4.8 stands out with the lowest error rate among the six evaluated models, which is a direct measure of the reduction in factual hallucinations. This performance was achieved primarily by choosing not to respond to uncertain questions rather than by increasing the number of correct responses.

In terms of pricing, Claude Opus 4.8 remains aligned with previous versions, costing $5 per million inputs and $25 per million outputs. The "fast mode" is offered at double the standard rate, representing a significant reduction compared to earlier models. However, this mode is reserved for organizations participating in the research preview, requiring contact with an account manager to access it.

The reliable knowledge cutoff date and the training data date are set for January 2026, the same as those for version 4.7. The context window remains at 1,000,000 tokens, with a maximum output of 128,000 tokens.

New Features and Improvements

The document "What's New in Claude Opus 4.8" highlights several notable improvements. Among these, the introduction of system messages during conversations is particularly interesting. This feature allows for updated instructions to be added during a conversation without rephrasing the entire system prompt. This preserves the cache hits from previous turns and reduces the input cost in agentic loops.

Another improvement concerns the minimum prompt cache length, now reduced to 1,024 tokens, down from 4,096 in the previous version. This reduction allows for more efficient resource management and optimization of the model's performance.

Experiments and Results

To test the capabilities of Claude Opus 4.8, experiments were conducted using the LLM CLI. The logs were exported in Markdown format, and the model was tasked with creating an HTML tool capable of rendering this Markdown with SVG code blocks displayed on the page. While the maximum model provided the best performance, it also incurred a notable cost of 43 cents for 25 input tokens and 17,167 output tokens.

These tests illustrate not only the technical capabilities of the model but also the associated costs of its use, an important factor to consider for potential users.