The scarce thing is capacity, not model quality

The message arrives at 3pm on a Tuesday. You have hit your usage limit and it resets in two hours. Or the API returns a 429 and the overnight job you set up has produced nothing. Or the invoice is up 40% and you cannot name a single thing you changed. Three different symptoms, one cause, and it has nothing to do with how clever the model is. You rented a slice of a machine that has other customers, and this hour they wanted it too.

This guide is about treating that slice as a budgeted line item instead of a background assumption. It is for solo operators and small teams who already pay for an assistant plan or hold an API key, and who have been surprised at least once by a limit or a bill. It is not for anyone choosing GPU vendors or financing a data centre. It also will not get you a better price than list, because at your volume there is no better price than list. What it will change is which tier your work runs on, which is worth more than the discount you cannot negotiate.

The constraint is physical, so it gets rationed

Capital points at scarcity faster than product launches do. In August 2026 Andreessen Horowitz announced a $1.1 billion fund aimed at the machinery underneath AI rather than the models on top of it, covering “all of the computer infrastructure on which AI runs, including chips, memory, networking, and storage”, plus data centres, robotics and home AI appliances [7]. The reason the firm gives is an engineering number rather than a market forecast. Rack power has moved from roughly 5 to 10 kilowatts to 100 to 250 kilowatts to support today’s systems, and is expected to reach 1 megawatt within three years, while compute density per rack rose 28 times between an H100 rack and a Rubin rack [7]. Hardware went from a small share of that firm’s deal flow to over 20% [7].

You do not need an opinion on any of that. You need one inference. When the binding constraint on a service is a physical asset that takes years to build, the service gets rationed, and rationing is what you are reading every time you look at your account page. The limits, the tiers, the reset windows, the four-fold price gap between fast and patient service: all of it is a scarce machine being divided among people who want it at the same moment [4]. Model quality is the part of the industry that improves on its own. Capacity is the part you have to plan around.

Your monthly plan is a capacity allowance with a model name on it

Look at how a consumer plan describes what you get. Claude Pro is $20 a month, or $17 a month billed annually, and the promise is a ratio rather than a quantity: Pro gives “at least 5x more usage per 5-hour session than Free” [6]. Max starts at $100 a month for 5x or 20x more usage than Pro [6]. Team seats are $25 a month, or $20 billed annually, and premium seats at $125 or $100 give 5x more than standard [6]. Nobody is selling you a number of words. They are selling you a share of a window.

The window is the unit that matters. Limits reset on rolling 5-hour session windows, and when you reach one on a paid plan you can wait for the reset, move up a plan, or turn on usage credits and keep working at standard API rates [6]. That third option is the honest one, because it prices the overflow instead of hiding it, and it is the moment your subscription quietly becomes metered.

Rationing by model is the other half of the design. On ChatGPT the free plan advertises unlimited text chats with GPT-5.6 Luna, subject to abuse guardrails, while the Pro plan is the one described as 5x or 20x more usage [8]. The word unlimited attaches to the cheapest model in the range. That is not a trick, it is the shape of the underlying cost. Capacity on a small model is abundant and capacity on a large one is not, so the small one gets the generous language.

So the figure to write down beside each plan is not the monthly price. It is the reset window and what happens when you hit the wall. If you know that a heavy morning costs you the middle of the afternoon, you schedule the heavy morning differently. If you do not, you lose two hours and blame the tool.

Three dials meter you on the API, and price is only one

On the API the rationing is explicit and you can read it. Anthropic meters requests per minute, input tokens per minute and output tokens per minute, all three at once [2]. Capacity refills continuously rather than resetting on the hour, because the limiter is a token bucket [2]. Every response carries headers telling you the limit, what remains, and an RFC 3339 timestamp for the reset, along with a retry-after value on the errors that have one [2].

Two details in that table are worth more than the rest. First, the expensive models get the smaller allowance. On the Start tier, Claude Opus 5, Sonnet 5 and Haiku 4.5 all sit at 1,000 requests per minute with 2,000,000 input tokens and 400,000 output tokens per minute, while Claude Fable 5.x is metered at 500,000 input and 100,000 output tokens per minute on the same request limit [2]. Claude Fable 5.1 costs $10 per million input tokens and $50 per million output against $5 and $25 for Opus 5, and it comes with a quarter of the token throughput [1][2]. Price and capacity move together, which is what you would expect if the scarce thing were machine time.

Second, only uncached input tokens count toward your input limit for most models [2]. That turns caching from a cost trick into a throughput trick, which matters when the thing blocking you is minutes rather than money.

Tiers sit above all of this. Organisations are placed automatically on usage history and account standing, with monthly spend caps of $500 on Start, $1,000 on Build and $200,000 on Scale, and no cap on a custom arrangement [2]. New organisations may begin in an evaluation tier with limits below the standard ones [2]. If you plan to run a launch or a migration through the API, the tier you are on today is a fact worth checking before the week you need it, not during.

Patience is the largest discount on the menu

Every major vendor sells the same tokens at half price if you will wait. Anthropic’s Message Batches API is 50% off standard input and output, taking Claude Sonnet 5 from $2 and $10 per million tokens to $1 and $5, and Opus 5 from $5 and $25 to $2.50 and $12.50 [3]. Google applies a 50% batch discount across the Gemini models, so Gemini 3.5 Flash drops from $1.50 and $9.00 per million tokens to $0.75 and $4.50 [5]. OpenAI’s batch pricing is half of standard too [4].

The spread on a single model is wider than most people assume. GPT-6 Astra is $10 per million input tokens and $50 per million output at standard rates on short context, $5 and $25 on either batch or flex, and $20 and $100 in Fast mode, which was renamed from Priority processing on 30 July 2026 [4]. Same model, same tokens, four times the price between the patient tier and the impatient one. You are not buying intelligence at that margin. You are buying a place in the queue.

The mechanics are worth knowing before you commit a workflow. An Anthropic batch is capped at 100,000 requests or 256 MB, most complete within an hour, and results become available when everything has finished or after 24 hours, whichever comes first [3]. Anything still unprocessed at 24 hours expires, and expired, errored and cancelled requests are not billed [3]. Results stay retrievable for 29 days [3].

Which work moves is a simple test. If a human is sitting there waiting for the answer, it stays on the standard tier. If the output lands in a document, a spreadsheet or an inbox that someone reads later, it belongs in a batch. Overnight enrichment of a contact list, transcripts turned into summaries, a backlog of support tickets classified, a weekly report drafted before you wake up: all of that is patient work paying an impatience premium right now.

Caching is the other half of the same bill

The second discount is for repetition. Anthropic charges a 5-minute cache write at 1.25 times the base input rate and a 1-hour write at 2 times, then reads at 0.1 times, so a 5-minute cache pays for itself after one hit and a 1-hour cache after two [1]. OpenAI prices cached input for GPT-6 Astra at $1 per million tokens against $10 standard, with cache writes at $12.50 [4]. The shape is the same at both: you pay a small premium once to avoid paying full price repeatedly.

The practical move is about prompt order rather than configuration. Put the stable material at the front, which means the long instructions, the style guide, the price list, the schema, the reference document you attach every time. Put the part that changes at the end. A prompt built that way can reuse its opening across every call in a session. A prompt that interleaves your fixed instructions with today’s variable text cannot cache anything useful, and you will pay full rate on the same 4,000 tokens all afternoon.

Because cached input does not count against your input token limit on most models, this buys headroom as well as money [2]. If you are hitting a throughput wall rather than a budget one, reordering the prompt is the cheapest thing you can try.

Introductory prices have an end date printed on them

Per-token prices fall over time, which is true and slightly misleading, because specific prices also rise on schedule. Gemini 3.8 Flash is currently $0.75 per million input tokens and $3.75 per million output, and the pricing page states the input rate is “$0.75 through December 31, 2026. $1.50 starting January 1, 2027”, with output moving from $3.75 to $7.50 on the same date [5]. That is a doubling, published in advance, on a model many people are quietly building margins on today. OpenAI likewise describes GPT-5.6 Sol’s $4 and $20 rates as promotional pricing available at least through 21 November 2026 [4].

There are quieter adjustments too. OpenAI’s regional data residency endpoints carry a 10% uplift for eligible models released on or after 5 March 2026 [4]. Google’s free Gemini tier is free of charge, but the paid tier is where the page states content is not used to improve the products [5]. Neither of those is hidden, and both change the arithmetic of a decision you may have made once and never revisited.

The habit that protects you is small. Whenever you write a per-token price into a spreadsheet, a proposal or a pricing model, write the date you read it and any expiry the vendor published, in the same cell. A price without a date is a guess about the future wearing the clothes of a fact.

Assume the tier you are standing on will move

Models retire, and retirement is a capacity decision as much as a product one. Anthropic’s pricing page lists Claude Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 as retired: Opus 4.1, Sonnet 4 and Haiku 3.5 remain available on Bedrock and Google Cloud, and Opus 4 only on Google Cloud [1]. If a model name is hard-coded in an automation you built last year, that name is a dated dependency, and it will fail on a day you did not choose.

Three habits absorb most of that. Keep the model name in one setting rather than scattered through scripts and saved prompts, so a swap is one edit. Keep the prompts themselves in a file you own rather than inside a vendor’s saved-instruction box, because a prompt you can move is a prompt that survives a migration. And know, before you need it, which smaller model does about 80% of the job, then run a week of real work through it so the fallback is tested rather than theoretical.

Retries deserve the same care. When a rate limit returns a retry-after value, use it, because a fixed sleep either wastes time you have or hammers a limit you do not [2]. Note that Anthropic does not send retry-after on the 429 you get from hitting a monthly spend cap, which is a different failure entirely: no amount of waiting fixes it, and only raising the cap or your tier does [2].

What still goes wrong

Every price, limit and tier above was read from the vendor’s own page on 4 September 2026, and two of them carry published end dates that fall before this guide is a year old [4][5]. Treat all of it as a snapshot to re-check each quarter rather than a standing fact. The structure lasts longer than the numbers: three dials, a queue you can pay to jump, a discount for repetition, and a tier you did not choose.

Batch is the change most likely to bite you. A job that expires at the 24-hour mark returns nothing at all, and you are not billed for it, which sounds fine until the report it was supposed to produce is the one your client expected on Monday [3]. Anything on the patient tier needs a check that the results actually arrived, and a plan for the morning they did not. The same applies to the fallback model you have never used in anger.

The honest limit is leverage. Nothing here changes the physical constraint, and at your volume you have no influence over it. Large buyers negotiate committed capacity; you get list price and a queue. What you control is the share of your work that has to be instant, which is almost always smaller than your current setup assumes, and how much of your spend is going to re-send the same 4,000 tokens of instructions that a cache would have held. Those two decisions are usually worth more than switching vendors, and unlike switching vendors, you can make them this week.