What jumped out at me is not that Anthropic is teaching Claude Code to be cheaper. It’s that it’s finally being explicit about the fact that agentic coding is a budgeting problem, not just a capability problem. That feels obvious once you say it, but most of the industry still talks about “better coding agents” as if the only axis that matters is benchmark quality. It isn’t. If the tool chews through context, rereads the world, and keeps spraying output back into the session, the user experience is going to feel expensive long before the model runs out of ability.
The article makes a pretty convincing case that Anthropic is trying to engineer around that by making cost behavior more legible and more steerable. The parts that sound most useful to me are the ones that reduce accidental waste: clearing stale context, compacting older turns, being stricter about when file reads get dragged into the prompt, and making output quieter when you don’t actually need a wall of text. That’s not glamorous. But it’s exactly the kind of stuff that decides whether an agent is something you keep open all day or something you abandon after a few annoying sessions.
I’m more skeptical about the confidence with which the article treats these controls as if they’ll line up neatly with real usage. In theory, /model, /effort, /clear, /compact, quiet output, and subagents sound like a tidy toolbox. In practice, people will misuse them, forget they exist, or disable the wrong thing and then wonder why the session got weird. The idea that Anthropic can make Claude Code cheaper by default is believable. The idea that it can make users consistently choose the right tradeoff every time feels a lot more optimistic.
The subagent angle is the bit I’d actually want to try. Splitting a task so the main session doesn’t have to carry everything forward is just good mechanical sympathy for these systems. It also feels like the cleanest way to keep agentic coding from turning into one giant, bloated prompt soup. That said, I suspect this is where the product can get confusing fast. Subagents only help if the boundaries are sharp. If they start acting like little disposable main threads with vague responsibilities, you haven’t really reduced cost; you’ve just hidden it.
What I like here is that Anthropic seems to be admitting something the marketing often soft-pedals: long-lived context is not a free lunch. Sometimes you should spend tokens to preserve state. Sometimes you should ruthlessly throw state away. The interesting part is not that Claude Code can do both. It’s whether the defaults help developers choose the right mode without having to think like a prompt accountant every five minutes.
Reference: なぜ「Claude Code」で想定外のコストになるのか？ 「トークン経済」を防ぐポイントまとめ：AI活用コストを最適化するメカニズムをAnthropicが解説 - ＠IT
