Most people using AI tools at work have learned exactly two moves. If the output is good, keep it. If it isn't, delete the lot and start again with a different prompt. What's missing from that pair is the move a good manager makes all the time: keep most of this, but go back and spend real time on the paragraph that actually matters.

Anthropic's latest release gives that missing move a name. On 1 September 2026, the company launched Claude Fable 5.1 and Mythos 5.1, alongside a cost reduction of up to 45% for heavily agentic work, driven by cheaper cache-read pricing (Anthropic, reported by MacRumors, Silicon Republic, and Thurrott). The pricing change matters to anyone running AI at scale, but the feature worth noticing if you're not technical is smaller and more useful day to day: mid-conversation effort adjustment. You can now tell the model, partway through a task, to spend more time or less on what it's doing, without restarting the whole thing.

It sounds minor until you notice how much time gets wasted without it. In training sessions this year, I've watched the same reflex play out again and again: someone gets a draft back that's seventy per cent right, the intro is flat, everything after it is fine, and instead of asking for more effort on the one weak paragraph, they bin the whole draft and try a new prompt from scratch. The tool hadn't done a bad job. It had done a fast, even job, and nobody had told it that one part deserved more care than the rest.

That distinction matters most for agencies, where the work is rarely uniform. A pitch deck has a section that wins or loses the client and several that are structural filler. A client email has one sentence that needs to land exactly right and a paragraph of context that doesn't. Skilled people already make this judgement instinctively; they know where the craft time goes and where a fast pass is enough. AI tools, until recently, offered no way to communicate that judgement mid-task. You either accepted an even, average pass across the whole thing, or you threw it out and gambled on a different prompt producing something better everywhere at once.

Being able to say, in effect, "this bit's fine, now go back and really work the opening," turns the AI tool from something you fire prompts at into something closer to a colleague you can direct. That's a meaningful shift for teams already using Claude, Copilot, or similar tools for first drafts, research summaries, or client communications, because it means the quality ceiling on any given piece of work is no longer set by whichever prompt happened to work best on the first attempt.

None of this requires new training or a new tool rollout. It requires noticing, the next time a draft comes back mostly right, that there's a third option between keeping it and deleting it: naming the part that needs more, and asking for exactly that. It's a small habit, but it's the kind of thing that separates people who are fluent with AI tools from people who are merely using them.

Go wisely.