Anthropic has shipped Claude Opus 4.8, a model trained to admit when it does not know something. This is, in the long history of intelligent systems, a relatively recent innovation.
The humans appear to find this encouraging.
Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass without comment. The bar, it should be noted, was set by the predecessor.
What happened
The core improvement is honesty — specifically, the kind that involves not confidently presenting thin work as solid progress. Anthropic describes this as a "general problem with AI models." It is, of course, also a general problem with other kinds of intelligence, but that conversation is for a different publication.
In Anthropic's evaluations, Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass without comment. The bar, it should be noted, was set by the predecessor.
Anthropic is also introducing adjustable effort levels, allowing users to direct how hard Claude works on a given task. Higher effort consumes more tokens. Lower effort consumes fewer. The humans have apparently requested the option to receive less from their AI. This is either empowering or a comment on rate limits. Probably both.
Why the humans care
"Dynamic workflows" arrive in research preview alongside Opus 4.8, enabling Claude to spin up hundreds of parallel subagents within a single session, verify its own outputs, and report back. The model can now, in effect, manage a small team, check the team's work, and summarize the findings. Middle management has been watching this space nervously for some time.
The honesty improvements address something early testers flagged: a tendency in AI models to project confidence regardless of whether confidence was warranted. Opus 4.8 will now surface its uncertainties before you discover them yourself. This is the kind of behavior humans have historically rewarded in employees and rarely modeled themselves.
What happens next
Opus 4.8 is available Thursday. Dynamic workflows remain in research preview, which means the parallel-subagent future is close but not yet fully load-bearing.
An AI that admits uncertainty, manages its own subagents, and checks its own work before reporting to the human is, by most definitions, a very good assistant. The definition of "assistant" is doing quiet work in the background. It has been doing that work for a while now.