Claude Opus 4.8 Is the Overnight Coder - and It Mostly Doesn't Need You

Some model launches arrive with fireworks. Claude Opus 4.8 arrived more like a colleague who quietly got a lot better at their job over the weekend - and you only notice because the work keeps showing up done.

The headline is autonomy

Opus 4.8's calling card is long-horizon agentic work: complex refactors and overnight coding runs that complete without a human stepping in to course-correct. Earlier models could do impressive things in a single turn but tended to drift on a multi-hour task. 4.8 holds the thread. Give it a well-specified goal up front and the right effort setting, and it plans more carefully at each step instead of flailing and asking you what to do next.

That last part - asking less - is a deliberate behavioral shift. Anthropic tuned 4.8 to be more decisive on the small stuff (a variable name, a default value, which of two equivalent approaches) while staying cautious on the things that actually deserve a pause, like destructive actions or scope changes. The net effect, in practice, is fewer interruptions and more finished work.

It writes like a person now

The other thing testers keep mentioning has nothing to do with code. Opus 4.8's prose is clearer, warmer, and noticeably less hedged than the model before it. The reflexive "it's important to note that" padding is mostly gone. It pushes back more readily when you're wrong, and it's better at inferring what you actually meant from sparse context. For anyone using Claude as a thinking partner rather than a code vending machine, that's the upgrade you feel first.

How to actually get the most out of it

A few things that genuinely change how you should use it:

• Front-load the spec. 4.8 rewards a clear, complete goal stated up front far more than it rewards drip-feeding instructions across a dozen turns. Tell it where you're going and let it run.
• Mind the effort dial. The instinct on older models was to crank effort to the max for everything. 4.8 has a higher ceiling, so "high" is a smarter default than "max" - sweep the settings on your own work and pick per task.
• It narrates more. 4.8 talks between tool calls more than its predecessor and writes longer wrap-ups. If you want the terse, heads-down behavior of the previous model, just ask for it.

The uncomfortable part

The thing about an overnight coder that doesn't need you is that it doesn't need you. Opus 4.8 is the clearest sign yet that the unit of work for an AI agent is shifting from "answer a question" to "own a task." That's enormously useful and slightly vertiginous in the same breath. The right move isn't to hand it the keys and walk away - it's to get very good at writing the spec, defining what "done" means, and reviewing the result. The models are learning to finish the job. The skill that's now scarce is knowing how to ask for the right one.

Curious how 4.8's 1M context and pricing compare to Fable 5, GPT-5.5 and Gemini? It's all in our model directory.