Claude Opus 4.8: Anthropic Bets on Honesty to Attract Users

With the launch of Claude Opus 4.8, Anthropic seeks to redefine the notion of honesty in the realm of generative AIs. The company is not just promoting a faster or more efficient coding model. It emphasizes an approach where honesty is paramount: fewer unfounded claims, more recognition of uncertainties, and better identification of errors in the provided responses. The idea is to create an assistant that prefers to say "I don't know" rather than provide an incorrect answer with the confidence of a LinkedIn intern.

Anthropic claims that Claude Opus 4.8 is about four times less likely than its predecessor to overlook errors in code without reporting them. In other words, the model does not just produce code; it also verifies its accuracy. This may seem less spectacular than a benchmark record, but it is likely more useful in daily practice.

The real challenge for AIs is not just to correct errors but to avoid presenting them as successes. A reported bug remains a bug, but an undetected bug can become technical debt. On this point, the honesty of Claude Opus 4.8 will need to be evaluated in real-world contexts, such as code reviews, system migrations, and document analyses, where each error can have significant consequences.

Anthropic also highlights a more rigorous and better-aligned model. However, alignment, just like trust, cannot simply be decreed in a release note. It must be earned when the model explains its limitations and does not turn every doubt into a pre-chewed certainty.

Another notable innovation concerns Claude Code. Opus 4.8 is capable of planning complex tasks, launching hundreds of sub-agents in parallel, and verifying their results before returning them. The goal is to tackle XXL projects, such as codebase migrations involving hundreds of thousands of lines. On paper, everything looks perfect. In practice, papers do not always compile.

This increased capability makes the question of honesty even more central. The more a model delegates, orchestrates, and automates, the less the user can monitor each step. If Claude becomes the conductor of dozens or hundreds of agents, it must also become a better quality controller. Otherwise, there is a risk of creating a distributed hallucination factory. Innovative, certainly. Reassuring, much less so.

In terms of availability, Claude Opus 4.8 is offered on Claude and via the API under the reference claude-opus-4-8, with unchanged pricing compared to Opus 4.7: $5 per million tokens for input and $25 for output. The fast mode promises a speed 2.5 times greater at a reduced cost.

In conclusion, Claude Opus 4.8 has not yet won the honesty trial by mere press release. It merely enters the courtroom with a better case. The final verdict will come from developers and all those who prefer an AI capable of saying "caution" rather than an AI that responds "mission accomplished" while hiding the debris under the rug.