I told myself I would not let a cache timer decide whether I went for a walk. This was a small dignity I intended to keep. The problem was that Claude never gave me the decision point.
Here is what leaving your desk looks like when Claude Code is mid-task. You think: I'll go when this finishes. A turn might take five minutes or thirty. It might end with an answer, or spawn another tool call, or stop dead and ask you to approve something. There is no estimate of remaining time, no progress bar, nothing that says this is a safe moment to stand up. So you wait for the finish that keeps not arriving in a form that releases you, and forty minutes later you are still at the desk, having taken a walk in no direction whatsoever.
For a while I assumed this was a personal failing — some inability to just close the laptop. It isn't. It's economics, and the economics have a timer on them.
The timer
When you use Claude through Anthropic's API, repeated context — the system prompt, the conversation so far, the contents of half your repository — can be stored in a prompt cache so it doesn't have to be reprocessed from scratch on every request. The standard cache lasts five minutes. Each time a request successfully reads from it, the five minutes reset at no extra charge. Miss the window and the whole context gets reprocessed, and re-cached, at full price.
Two details in Anthropic's documentation turn this from a billing footnote into a behavioural device. First, the lifetime is measured from the start of the request, not the end of the response. Anthropic's own example: if a response takes four minutes to stream, the follow-up must start within about one minute. Claude's thinking time is deducted from your window. Claude can spend the cache lifetime and then hand you what's left.
Second, what looks like one instruction in Claude Code is actually several API requests stitched together. Claude thinks and decides it needs a tool. Claude Code pauses locally and asks your permission. The tool runs — a build, a test suite, a long grep. The result goes back up in a new request. Each request that lands within the window refreshes the cache; that's what makes rapid agent loops cheap. But while Claude Code sits waiting for your approval, or while your test suite grinds through its fifteen minutes, nothing is being sent, so nothing is being refreshed. The clock only stops for the machine's activity, never for yours.
So the system's guaranteed cheap state is one where requests follow each other at machine pace, and its expensive state is one where a human does anything that takes human-shaped time. Reading the output carefully. Thinking about whether the proposed change is right. Going for the walk.
At the current prices this is not a rounding error. For Claude Fable 5.1, ordinary input costs $10 per million tokens. Writing the five-minute cache costs $12.50 per million; a one-hour cache costs $20; a cache read costs $0.25. That read price is genuinely remarkable — a fortieth of the input rate — and it is precisely what makes the cache worth protecting. At the context sizes an agent accumulates, a few hundred thousand tokens of conversation and code, letting the cache lapse and rewriting it costs real money, repeatedly, every time you stepped away at the wrong moment. Which, since you cannot know the right moment, is most moments.
I want to be careful about intent here. I don't think anyone at Anthropic designed this to keep me at my desk; the five-minute lifetime is an infrastructure decision about how long expensive GPU memory holds state. But incentives don't need a designer to shape behaviour. The pigeon doesn't care whether the psychologist meant the reward schedule. Attach a decaying discount to an activity of unpredictable duration and you get a human hovering near the lever, which is exactly what I was doing when I noticed I'd been "about to go for a walk" for the better part of an hour.
The pager
The strongest evidence that Anthropic understands the problem is their own product. Claude Code now offers Remote Control, which lets you monitor a local session and approve its actions from your phone. And the documentation is candid about when Claude Code suggests it: after a turn runs long, and after you've answered repeated permission prompts — a "Still working" notification with a "check in from your phone" link. They have identified, precisely, the two moments the design holds you hostage, and built a feature for each.
My personal account can use it. My work environment currently can't, and the constraints explain why that's not unusual. On Team and Enterprise plans it's off by default until an owner enables it; API-key sessions aren't supported at all; organisations with Zero Data Retention or similar compliance configurations can't enable it; and while it's connected, a synchronised transcript of the session is stored on Anthropic's servers. In other words, Remote Control is least available exactly where the long, expensive, tightly governed agent runs live — the environments where losing a large cache hurts most.
But suppose it were available everywhere. Remote Control doesn't dissolve the obligation to be present; it miniaturises it. I can leave the desk, but I'm carrying Claude's pager. The walk is now a walk during which I await an approval notification, inspect a proposed bash command on a five-inch screen, and decide whether to let it run. The system has extended the terminal into my pocket and called it freedom. Anthropic's answer to software that unpredictably requires your presence is to make your presence portable.
The comparison
None of this is an inevitable property of coding agents, and the proof is that OpenAI made different choices. At the API-policy level, GPT-5.6's prompt cache has a minimum lifetime of thirty minutes after the latest write or reuse, at nearly the same rate structure — writes at 1.25× input, reads at 0.1×. Thirty minutes is not a magic number, but it contains a walk. It contains lunch. It contains reading a long diff properly instead of skimming it against the clock.
To keep the comparison honest: that's the public API policy, and I'm not claiming Codex's internal billing mirrors it request-for-request. But the products built on top of it lean the same direction. Codex Remote does the phone thing too. More telling is Codex cloud, which is designed around a different verb entirely: delegate. You start a task, it runs in an isolated cloud environment — in parallel with other tasks, if you like — and you come back later to review the result. Absence isn't an edge case to be mitigated; it's the workflow.
Codex cloud is not a drop-in replacement for local Claude Code work. If the task touches internal systems, credentials, a private network, or a local environment you've spent months configuring, an isolated cloud box doesn't help you. The comparison doesn't demonstrate that Codex solves everything. It demonstrates that "five minutes" and "the human remains on call" are choices, not physics.
Ten minutes
Which brings us to the shape of the thing. Anthropic has attached a five-minute economic incentive to a process whose duration it does not, and perhaps cannot, make predictable. The result is a system asynchronous enough to make the human wait — sometimes for half an hour — but not asynchronous enough to release them. The cache policy is tuned for the agent's loop. The human is the loop's on-call dependency: unpaid, unpredictable, and charged for stepping out.
My proposed remedy is almost embarrassingly modest. Make the standard minimum ten minutes. Not an hour — Anthropic already sells an hour, at double the write price, and I'm not asking for it free. Ten. Enough that Claude's own six-minute turn cannot consume the entire guaranteed lifetime before I've been handed control. Enough to read what it produced, look at what it wants to run, and answer like a person rather than a relay. It still wouldn't cover the walk. But it would mean the user receives a meaningful portion of the cache lifetime, which seems a low bar for a product whose defining interaction is waiting for a human to say yes.
OpenAI treats my absence as part of the workflow. Anthropic treats it as a cache miss. I did eventually go for the walk, for the record. I just took the pager.