Plan for the price of AI to change

Somewhere in your business there is a number you got from a vendor pricing page and then stopped thinking about. It might be in a spreadsheet where AI costs about $40 a month. It might be inside a client-facing tool you charge a flat fee for. It might be a background job you set running last year that summarises everything and has never once appeared on your radar. You treated that number the way you treat the price of electricity: it drifts a few percent, and it is somebody else’s problem.

It is not electricity. DeepSeek introduced peak and off-peak rates effective August 16, 2026, taking peak output on its V4-Pro model from $0.87 to $3.96 per million tokens, and said it was revising pricing “to allocate resources more reasonably” [8]. Google’s Gemini pricing page says today that Gemini 3.8 Flash input costs $0.75 per million tokens through December 31, 2026 and $1.50 starting January 1, 2027 [4]. Anthropic cancelled a scheduled increase and left Claude Sonnet 5 at $2 per million input tokens instead of moving it to $3 [2]. None of those is a crisis. All of them are the same fact, which is that the price is a decision somebody makes, it usually has a date attached, and the date is usually printed. This guide is for a solo operator or small team whose AI spend is metered and who would notice it doubling. If you have a negotiated committed-use contract, your prices live in that contract and none of this applies to you.

Every number on a pricing page has a date attached

Anthropic, OpenAI and Google each publish the temporary nature of their prices in a different format, and once you know the formats you can read any pricing page in about a minute.

Google prints both prices and the switchover date in the same cell. Gemini 3.8 Flash output is listed at $3.75 per million tokens through December 31, 2026 and $7.50 starting January 1, 2027, with context caching moving from $0.075 to $0.15 and cache storage from $0.50 to $1.00 per million tokens per hour on the same date [4]. That is a scheduled doubling, announced months ahead, sitting in public on the page you would have copied the number from.

OpenAI prints a footnote. Under the flagship pricing table it says that “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026” [3]. The current rate is $4 per million input tokens and $20 per million output tokens for short context [3]. The words doing the work there are “promotional” and “at least”, and neither of them appears in the table itself.

Anthropic prints the history. Its pricing documentation records that Claude Sonnet 5’s $2 and $10 rates were announced as introductory pricing through August 31, 2026, that an increase to $3 and $15 was scheduled for September 1, 2026, and that the increase will not occur because those rates are now standard [2]. That is the same mechanism running the other way.

DeepSeek prints no date at all, only the clause: “Product prices may vary and DeepSeek reserves the right to adjust them”, followed by advice to top up based on actual usage and check the page regularly [1]. Absence of a date is not a promise. It is the widest version of the same statement.

So the habit is small and it costs nothing. Read the footnotes before the table. When you copy a price into a plan, copy the sentence next to it and the date you read it, and put both in the same cell as the number.

The direction of travel matters less than your dependence on it

DeepSeek raised peak rates by more than 4x [8] and remains the cheapest option on this page by a wide margin. Peak output on deepseek-v4-pro is $3.96 per million tokens [1], against $25 for Claude Opus 5 [2] and $50 for gpt-6-astra at standard short-context rates [3]. Off-peak it is $1.98 [1]. Nobody was priced out of anything. The competitive position barely moved.

What moved was the status of the assumption. Two years of falling prices taught a lot of people to treat inference cost as an input that only goes one way, which made it safe to write a single number into a model and never revisit it. The number was never the fragile part. The single number was.

This cuts both ways, which is the part worth internalising. Anthropic’s cancelled increase means a business that had already repriced its own product upward in anticipation spent a quarter being more expensive than it needed to be [2]. Vendors adjust in both directions and on their own schedule, and a plan with exactly one price in it is wrong in both cases.

Your bill is tokens times price, and the tokens are your half

The vendor sets one factor. You set the other, and the other one is usually larger and more neglected.

Output is where the money is. Claude Opus 5 charges $5 per million input tokens and $25 per million output [2]. gpt-6-astra charges $10 and $50 [3]. Gemini 3.8 Flash charges $0.75 and $3.75, and its output price explicitly includes thinking tokens [4]. A ratio of 5 to 1 means that anything which reasons at length, retries, or runs as an agent producing intermediate steps bills disproportionately on the expensive side of the ledger. Cutting a verbose system prompt saves you input tokens. Telling the model to answer in 200 words instead of 800 saves you five times as much per token.

Context length is priced differently by vendor and the difference is large enough to change which one you pick. OpenAI charges more for long context, so gpt-6-astra moves from $10 and $50 to $20 and $75 [3]. Anthropic bills the full 1M token window at standard rates for Claude 4.6 and later, so a 900,000-token request costs the same per token as a 9,000-token one [2]. Two vendors, two different bills for the same habit of pasting everything in.

There are also adders that never appear in the headline table. Asking Anthropic for US-only inference applies a 1.1x multiplier across every token category [2]. OpenAI’s regional data-residency endpoints carry a 10% uplift for models released on or after March 5, 2026 [3]. Gemini’s Google Search grounding gives 5,000 free search requests per month shared across all Gemini 3.x models and then charges $14 per 1,000 [4]. A compliance requirement and a search tool are both price changes you agreed to.

Four levers cut the bill before you change vendor

Switching vendor is the expensive response and most people reach for it first. There are four cheaper moves, and the vendor already offers most of them.

The first is patience. Batch processing is 50% off at Anthropic, OpenAI and Google alike [2][3][4]. Anthropic’s Message Batches API says most batches complete within an hour, with a 24-hour maximum window and results retained for 29 days, and caps a batch at 100,000 requests or 256 MB [5]. Any work whose answer does not need to exist within a few seconds belongs here: overnight classification, weekly report generation, bulk extraction from a folder of documents. OpenAI lists its Flex tier at the same rates as batch on every model in its pricing table [3].

The second is repetition. Prompt caching prices a repeated prefix at a fraction of fresh input. On most Claude models a cache hit costs 0.1x the base input price, and on Claude Fable 5.1 it costs 0.025x, or $0.25 per million tokens against a $10 base, with cache writes at 1.25x for the 5-minute duration and 2x for the hour [2]. OpenAI charges $1 per million for cached input on gpt-6-astra against $10 fresh [3]. Gemini charges $0.075 per million for context caching plus $0.50 per million tokens per hour of storage [4]. If every request in your system starts with the same 6,000 tokens of instructions and reference material, you are paying full price for the same text thousands of times. Batch and caching discounts stack at Anthropic, though cache hits inside a batch are best-effort [5].

The third is the clock, where the vendor offers one. DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, and every other hour of the week is half price [1]. Most background workloads run at peak only because nobody wrote the scheduling line that would move them.

The fourth is the tier, both of model and of speed. gpt-5.6-luna is $0.20 and $1.20 per million against gpt-6-astra’s $10 and $50 [3]. Claude Haiku 4.5 is $1 and $5 against Claude Opus 5 at $5 and $25 [2]. Speed costs in the other direction: OpenAI’s Fast mode doubles standard rates, putting gpt-6-astra at $20 and $100 [3], and Anthropic’s Fast mode prices Claude Opus 5 at $10 and $50 rather than $5 and $25 [2]. Paying twice for latency on a job nobody is waiting for is the most common form of this mistake.

You will be forced to migrate anyway, so keep the move cheap

The price change you cannot decline is retirement, and it arrives on a schedule regardless of what happens to rates.

Anthropic commits to at least 60 days of notice before retiring a publicly released model, and has already retired Claude Opus 4.1 as of August 5, 2026 and Claude Sonnet 4 as of June 15, 2026 [6]. OpenAI commits to at least 6 months for generally available models and at least 3 months for specialised variants, and says preview models may be retired with much shorter notice, such as 2 weeks; its gpt-5 and o3 snapshots are scheduled for shutdown on December 11, 2026, and the Assistants API shut down on August 26, 2026 [7]. Parameters retire too: temperature, top_p and top_k now return a 400 error on Claude Opus 4.7 and later when set to non-default values [6].

Read those numbers as your planning horizon. On a preview model it can be about a fortnight [7]. That is not enough time to discover a hard-coded model name in a script you wrote nine months ago, so move model names, prices and endpoints into configuration now, while nothing is on fire. The test is whether changing which model you call is an edit to one line of settings or a deploy.

Interface portability is the other half, and it is better than it used to be. DeepSeek serves both an OpenAI-format base URL and an Anthropic-format one, so calling it is a base URL and a model name rather than a rewrite [1]. That does not make providers interchangeable, because prompts tuned on one model rarely land identically on another and you have to re-run whatever evaluation you trust. It does mean the mechanical part of switching is close to free, and the mechanical part is what people wrongly assume is the obstacle.

Do not resell a price you do not set

If you charge a client a flat monthly fee for something that costs you tokens, you have written an option on somebody else’s pricing page and given it away for nothing. A 4x move on a rate of $0.87 per million tokens is small in absolute terms [8]. The same move against a fixed fee you already quoted comes entirely out of your margin, and you cannot reprice mid-contract without a conversation you would rather not have.

Three fixes, in ascending order of honesty. Cap the usage, so the fee buys a stated quantity and overage is billed. Meter it, so the customer sees the same variable you see. Or price with headroom by computing your cost at twice the current rate and twice the current volume and quoting against that, which is what the calculator above is for on its second run.

The free tier deserves the same suspicion, because a free tier is a data policy as well as a price. Google’s pricing page states that free-tier content is used to improve its products and paid-tier content is not [4]. Moving off free because your volume grew changes what you pay and what happens to your inputs on the same day, and only one of those two changes will show up on a bill.

What still goes wrong

Every figure here is a snapshot cited to the vendor’s own page on the date in the header, and the whole point of the guide is that snapshots expire. Check the source before you commit money to any of these numbers, including the ones that look stable. The pages themselves tell you to: DeepSeek’s ends by recommending you check it regularly [1].

Cheapest per token is not cheapest per job, and the arithmetic that makes a small model look like a bargain quietly assumes it gets the answer right first time. A model at one twentieth the price that needs three attempts and then a human read-through is not cheaper, it is more expensive and slower and you have hidden the cost in your own afternoon. The only way to know which side of that line you are on is to run your real work through both and compare outputs, not prices.

Multi-provider routing is genuinely useful leverage and it is also a second system to maintain, a second set of quirks to learn, and a second evaluation suite to keep current. For a business of one or two people the honest recommendation is usually to stay on one vendor, keep the exit written down rather than built, and spend the saved effort on the four levers in this guide, which are available today without adding a dependency. Build the routing when a real bill justifies it, not in anticipation of one. The same discipline that protects you from a price move will, applied too early, cost you more than the price move would have.