The cheapest AI work is the work that can wait

Every few months a headline arrives about the machinery under your AI subscription. A memory maker lists on Nasdaq. A packaging plant gets announced. Somebody quadruples something. The implied question is whether you should be doing anything differently, and the honest answer is that the news itself changes nothing you will pay this year. Physical capacity gets ordered long before it exists, and the money in those headlines buys buildings first.

There is a lever in the same story that does change your bill, and it is sitting on the pricing page of every model vendor you already use. The same model, doing the same job, is sold at three prices depending on how quickly you need the answer back. Half price if you can wait a day [4][7][8]. Standard if you want it now [5]. Double if you want it now and prioritised [7]. Most solo operators pay the middle price for everything, including work that nobody is sitting and watching. This guide is about sorting your work by urgency and pushing as much of it as possible onto the cheap rung. It is not for anyone buying GPUs or negotiating cloud contracts, and it is not for latency-critical consumer products where every request has a human staring at a spinner.

Capacity is announced years before it exists

The reason a queue exists at all is that the hardware under it cannot be conjured. Take the most-cited example of the current cycle. In July 2026 SK hynix sold 177.9 million American depositary shares at $149 each and raised $26.5 billion, and the proceeds were earmarked for three things: a new fab in South Korea, a new packaging facility, and EUV scanners [2]. Note what is missing from that list. None of it is chips. It is buildings and machines that will one day make chips.

The company’s own quarterly statement puts dates on the same money. It describes executing mid-to-long-term investment plans in phases, including a recently announced P&T7 advanced packaging facility and an M17 NAND base, with the Yongin Phase 1 cleanroom opening in early 2027 [1]. A cleanroom opening is not production. It is the point at which equipment starts arriving and qualification begins. Meanwhile the company had already signed long-term agreements with around 10 customers [1], which is the part worth reading twice: a meaningful share of the capacity being built has been spoken for by large buyers before the roof is on.

That is the durable shape of this industry, and it will still be the shape when today’s specific companies and facility codenames are forgotten. Demand for inference moves at the speed of a product launch. Supply moves at the speed of the schedule the company itself published: capital raised in July 2026 [2], a cleanroom opening in early 2027, and mass production accelerating after that [1]. The two clocks run at different speeds, and everything downstream of the gap between them, including your invoice, is a mechanism for rationing it.

The shortage reaches you as a price ladder, not an outage

You will almost never be told there is no capacity. You will be offered a discount to go away and come back later, which is the same information delivered politely.

The bottom rung is batch. Anthropic charges 50% of standard prices for batched requests, so Claude Opus 5 runs at $2.50 per million input tokens and $12.50 per million output instead of $5 and $25 [3][4]. OpenAI’s price list puts batch at exactly half the standard rate for every model on it, input and output alike [7]. Google’s Batch API is likewise 50% of the standard cost and is described as being for large-scale, non-urgent tasks such as data pre-processing or running evaluations, where an immediate response is not required [8]. Three vendors, three independent pricing decisions, one identical number. That is not a promotion. That is what deferrable demand is worth to someone managing a queue.

The middle rung is what you are on now. Anthropic’s service tiers page describes standard as the default, prioritised alongside all other requests with best-effort availability [5]. Best-effort is doing real work in that sentence.

The top rung is urgency you pay a premium for. OpenAI’s pricing page lists a fast mode, formerly called priority processing, at double the standard price [7]. Anthropic’s priority tier prioritises requests over all other requests to minimise server-overloaded errors even at peak, and targets 99.5% uptime [5]. Half, standard, double. On OpenAI’s own price list that is a spread of four times between the cheapest and the most expensive way to run an identical prompt through an identical model [7], and none of it is about quality.

Sort your jobs by who is actually waiting

The sort is easier than it sounds because the test is not technical. For each recurring AI job you run, ask whether a human is sitting there when it runs. If the answer is no, it belongs in batch.

Work that almost always qualifies: classifying or tagging a backlog, summarising yesterday’s transcripts, drafting first passes of content that a person will edit tomorrow morning, enriching a spreadsheet of leads, re-processing an archive after you improve a prompt, and any evaluation run where you are comparing outputs rather than shipping them. Work that does not qualify: anything inside a chat window, anything a client triggers, anything where the output feeds the next step of a process a person is walking through right now.

The waiting is less bad than the word “batch” suggests. Anthropic states that most batches finish in under 1 hour, with a hard 24-hour processing window after which the batch expires, and results retained for 29 days [4]. Google targets 24 hours and notes that in the majority of cases it is much quicker [8]. So the real trade is usually minutes, not a day, and the discount is a flat 50% either way. Size limits are generous at solo-operator scale: an Anthropic batch holds up to 100,000 requests or 256 MB, whichever comes first [4].

The scheduling change that follows is small. Instead of running your nightly job as a loop of individual calls, you assemble the requests into one file, submit it, and pick up the results when you next sit down. If your workflow lives in a tool like n8n, Make or Zapier rather than in code, the equivalent move is to stop triggering per-item and start collecting items into a daily run.

Caching is the second half of the same discount

The other thing a capacity-constrained vendor wants is for you to stop resending material it has already read. It pays you for that too.

On Claude, a cache hit costs 0.1 times the base input price, so the same Opus 5 input drops from $5 to $0.50 per million tokens, with a write costing 1.25 times base for the 5-minute cache or 2 times base for the 1-hour cache [3]. OpenAI’s cached input tokens cost 90% less than standard input, with cache writes carrying a 25% premium [7]. The economics only work if the stable part of your prompt sits at the front and does not change between calls, which is a formatting habit more than an engineering project. Put your instructions, your style examples and your reference document first, and the variable item last.

There is a second payoff that is easy to miss. Anthropic’s rate limits are measured in requests per minute, input tokens per minute and output tokens per minute, and for most models only uncached input tokens count toward the input-token limit [6]. Caching therefore buys you throughput as well as money. The same 40-page reference document, cached, stops eating your minute-by-minute allowance every time you ask a question about it.

Guaranteed capacity is not something you can buy at your size

It is worth knowing where the ladder ends, so you do not spend a week trying to climb to a rung that has been removed. Anthropic’s service tiers page states plainly that priority tier capacity commitments are no longer available for purchase, and that organisations with an existing commitment can keep using it until their contract ends [5]. Those commitments were structured as input tokens per minute, output tokens per minute, a duration of 1, 3, 6 or 12 months, and a specific model version [5]. That is an enterprise procurement object, and it is closed.

So plan for best-effort, and make best-effort survivable. Rate limits at Anthropic use a token bucket that refills continuously rather than resetting at fixed intervals, and exceeding one returns a 429 with a retry-after header telling you how many seconds to wait [6]. Honour that header instead of guessing at a sleep interval. A separate 429 exists for crossing your organisation’s monthly spend cap, and it arrives without a retry-after header, because access resumes at 00:00 UTC on the first day of the next month, or sooner only if you move to a higher usage tier [6]. Those two errors look nearly identical and mean completely different things, and the second one will take a workflow down for days if you have not read the error code.

The practical defences are unglamorous. Retry on the header. Keep a second model configured that you could route to, on a different vendor, tested on your actual prompts at least once so you know it works. And keep your deferrable work in batch, which Anthropic describes as sitting outside your normal capacity [5], meaning the bulk job you moved for the discount also stops competing with the request a client is waiting on.

Read supply news for direction, not for dates

Once you have the ladder set up, hardware headlines become almost restful, because you know which layer they touch.

A funding round, a listing, a fab announcement or a capacity target changes the supply curve on the timetable the company itself publishes, and that timetable is measured in cleanrooms and qualification runs rather than in quarters [1]. It is a reason to expect the price of intelligence per unit of work to keep drifting downward over multi-year horizons, and a reason not to sign anything long and expensive on the assumption that today’s scarcity is permanent. It is not a reason to change your October invoice, and it is not a signal to act on this week.

The things that do change your month are duller and easier to check. Vendor pricing pages, which move without a press release. Model deprecation notices, which arrive by email and get filed unread. Tier and plan changes on your subscription. Put a recurring 15-minute check on the first working day of each month, open the pricing pages for the two vendors you actually use, and compare against what you wrote down last time. That habit catches more money than any amount of chip-supply reading.

What still goes wrong

Batch is not free of friction, and the friction lands in the wrong place for small operators. You lose the interactive loop, so a prompt bug that you would have caught on the second message instead reveals itself across the whole run an hour later, with the tokens already spent. Run a batch of 20 first, read the outputs yourself, and only then submit the real one. Batches also expire: Anthropic’s window is 24 hours and unfinished work at that point is gone [4], which is survivable for a backlog and not survivable for a report due at 09:00.

The discounts also tempt you into volume you did not need. Half price on work that produced nothing useful is still money spent, and the cheapest possible batch is the one you delete from the schedule because nobody read last month’s output either. Before you optimise the price of a recurring job, check that anyone is using it.

And the deeper constraint does not go away by being routed around. The queue exists because the physical layer is genuinely tight, and the vendors sitting on top of it can change the terms of the ladder. Anthropic already withdrew purchasable priority capacity from new customers [5], and OpenAI renamed priority processing to fast mode while keeping the 2x price [7]. Nothing obliges any of them to keep batch at exactly half. The habit worth building is not “use batch”; it is checking, every month, what the current price of patience actually is.