← All articles
AI economics

Is AI getting more expensive?

Every few weeks someone in my feed announces that AI is getting expensive. So I went through the price lists. Almost nothing moved. Then I looked at what sits behind the prices, and almost everything did.

$20ChatGPT Plus in January and todaythe menu around it grew$0$100$200$300$400$500ChatGPT$8 Go$20 Plus$100$200$500Claude$20$100$200Google AI$100$200was $250X / Grok$8$30$100$200was $300Cursorwas $20$30Monthly price, USD. Gold = tier added in 2026. Dashed = old price.

ChatGPT Plus cost $20 in January and it costs $20 now. The menu just grew around it: Go for $8 with ads, Pro tiers at $100 and $500. Google even cut its top AI Ultra plan from $250 to $200. The only clear sticker increase I found was Cursor, which moved new customers from $20 to $30.

The catch is in the small print

Prices held. The limits are where things move. Plus got its five-hour limit back in August. In September OpenAI announced it would halve the allowance on its $200 Pro plan, then backed down a day before the cut and doubled the limits on the $100 and $500 plans instead. Claude renamed its $200 plan from Max 20x to Ultra. Price and weekly limit stayed, the five-hour cap went away, and the price list now says plainly: 10× Pro per week. OpenAI was refreshingly honest about the pressure: the $200 plans "put the most strain" on its systems. The people running agents and code all day are the ones pushing the bill.

So yes, the critics have a point. Half of one.

What twenty dollars buys now

A year ago the best result on ARC-AGI-2, a reasoning test built to be hard for AI, was 45% (Gemini 3 Deep Think, November 2025). By September 2026 GPT-6 was at 95%. For real work I find METR more telling. It measures how long a task, in human working time, AI can finish with 50% success. GPT-4 in 2023 managed about four minutes. Claude Mythos Preview in April 2026, about seventeen hours. That's a full shift plus overtime. (METR says anything above 16 hours is less reliable, and Mythos isn't in any normal plan. Still.)

4 min → 17 htask length AI finisheswith 50% success, log scale1 min10 min1 h10 h20232024202520264 minGPT-4, 2023≈4.9 hOpus 4.5, 11/2025≈17 hMythos Preview, 4/2026
METR: length of a task (in human working time) that AI completes with 50% success. Log scale. Source: METR Time Horizon 1.1.

METR's yardstick doubles roughly every four months, which works out to about 7× in a year (my conversion). Benchmarks aren't real work, I know. But name one other thing you buy that got several times better in twelve months while the price tag stayed at $20.

Then why does it feel pricier?

Because we use it so much more. The same level of performance gets about 13× cheaper every year (Epoch AI). A typical question cost six cents on GPT-4 in 2023. On GPT-6 Luna today it's $0.00035, roughly 171× less by my count. Cheap units invite more use. Economists call it Jevons' paradox: when steam engines got better, Britain burned more coal, not less. Google processed 9.7 trillion tokens a month in May 2024. Two years later, 3.2 quadrillion.

≈330×more tokensin two yearsMAY 20249.7 trillionMAY 2025480 trillionMAY 20263.2 quadrillionTokens processed by Google per month. Bar length on a log scale.
Tokens processed by Google per month, log scale. Source: Google I/O 2026.

Teams show the same pattern. A developer moves to a newer model with tokens 3.75× cheaper, and the agent now writes five times as many. The monthly bill goes up by a third (my calculation). Nobody got cheated. They just get a lot more done. And a team of five on $20 plans still pays only about 26,500 CZK a year. That's a fraction of one new hire.

The real savings aren't in haggling over plans anyway. Claude Opus 5.5 costs $0.98 per task on low effort and $10.04 on max. Leaving everything on max is like running every machine at full speed for every part, washers included. A cheaper model that scores 4 points lower can often do the job for an eighth of the price. Cached input is 95% off, batch jobs half price. On the shop floor we send simple parts to the fast station and hard ones to the specialist. AI needs exactly the same discipline.

The twenty dollars didn't change. The helper who managed four minutes is now a colleague who can work a whole shift. If AI feels expensive, look at your process before you look at your plan.

Sources: METR, ARC Prize, Epoch AI, Artificial Analysis, official pricing pages, help centres and announcements (OpenAI, Anthropic, Google, Cursor) and the tech press (The Verge, TNW, BleepingComputer). My own calculations are marked in the text. Prices as of 14 November 2026; 1 USD = 22.101 CZK (CNB).

Want AI to pay off in your operation, not just on your invoice? Let's talk.

Get in touch LinkedIn