The battle of AI models is no longer won by benchmarks. Price decides — and the job market is changing too

July 21, 2026 Daniel Cesak

AI article illustration for ai-jarvis.eu

The competition for the smartest AI model takes place in benchmarks. But the competition for the customer is increasingly taking place in the price list. OpenAI, Anthropic, Google, and others have been subsidizing the operation of artificial intelligence with investor billions for years to gain users — this era is drawing to a close and AI bills are rising. Along with them rises the question that will change not only the model market but also the labor market: is the best model worth twice the price when you can get a slightly worse one for a fraction?

The subsidized era is ending, price lists speak clearly

That cheap artificial intelligence is becoming a thing of the past is no longer just a claim by analysts, but is directly stated in official price lists. Anthropic lists an introductory price of 2 dollars per million input and 10 dollars per million output tokens for its Claude Sonnet 5 model — but only until August 31, 2026. Then the price jumps to 3 and 15 dollars, i.e. fifty percent higher. The top-tier Claude Fable 5 costs 10 and 50 dollars per million tokens, as shown in Anthropic's price list.

A similar pressure is visible with subscriptions. A full Claude Pro costs 20 dollars per month, a heavy user pays extra starting at 100 dollars (roughly 2,500 CZK) under the Max plan, and a corporate "premium" seat comes to up to 125 dollars per month. OpenAI charges 25 dollars per user per month for ChatGPT Business (20 with annual payment), states the OpenAI price list. The price hikes are not limited to America: as we wrote last week, even the Chinese open model Kimi K3 has become the most expensive model ever offered by a Chinese lab — 3 and 15 dollars per million tokens.

The logic is simple: without subsidies, the true cost of intelligence is revealed. And it is measured in tokens, energy, and chips.

Price math: what a million tokens costs today

When we put the price lists side by side, the differences seem almost unbelievable. Today you will pay the following for a million input and output tokens (input / output prices):

Claude Fable 5 (Anthropic): 10 / 50 USD
Claude Opus 4.8 (Anthropic): 5 / 25 USD
Claude Sonnet 5 (Anthropic): 2 / 10 USD, from September 3 / 15 USD
Gemini 3.1 Pro (Google): 2 / 12 USD
DeepSeek V4 Pro: 0.435 / 0.87 USD
Gemini Flash-Lite (Google): 0.10 / 0.40 USD
DeepSeek V4 Flash: 0.14 / 0.28 USD

In other words: for the cost of one answer from the most expensive Claude Fable 5, you can get roughly fifty-seven answers from DeepSeek V4 Pro. And between the cheapest Gemini Flash-Lite and the top-tier Gemini 3.1 Pro, there is a twentyfold price difference, confirms Google's price list. DeepSeek, moreover, lists a context window of one million tokens for its fourth generation and results that approach the American top tier in many benchmarks — similar success is being seen with other open Chinese models.

Is "nearly the best" enough? For most work, yes

Differences in benchmarks between flagship models are now measured in single percentage points. Differences in price are measured in multiples. That is precisely why corporate AI architecture is changing: instead of one best model, companies are deploying so-called model routing — common queries go to cheap models, and the expensive top tier is used only where it truly pays off. Anthropic also offers a 50% discount for batch processing and significantly cheaper cache reads, so disciplined operation costs a fraction of the list price.

For the customer, the question "which model is the best" is turning into "which model is the best for the money for my specific task." And exactly the same logic is now shifting to the labor market as well.

Work will not be taken from either side. It will be divided

It seems that AI will not take work away from all developers, designers, accountants, or lawyers. A more likely scenario is a division of labor based on the same price math as with models. AI may be somewhat more expensive than human labor — but precise, consistent, and available 24 hours a day. It does not sleep, does not get sick, does not lose focus after lunch. Human labor will be cheaper, but less perfect. The customer will then choose according to purpose: where faultlessness and machine reliability matter, they will pay for AI. Where human standard suffices, they will hire a person.

A new profession: the human as captain of AI agents

The third scenario is perhaps the most interesting: some people will become the "human factor" in the work of AI agents. It already holds true that the prompt author determines the quality of the result. A website of some quality can be put together by a layman today. But a website is not just about how it looks — it is about performance, security, accessibility, SEO, and code maintainability. When the assignment and cooperation with AI is handled by a developer with years of experience, the result tends to be not only substantially better, but often also cheaper: an experienced person gives a more precise task, burns fewer tokens on dead ends, and spends less time fixing errors that a layman would not even notice.

History repeats itself. Once, documents were typed by those who were great on a typewriter. Then came Word and printers — those who learned to work with a computer carried on; those who did not had a problem. With AI it will be similar, and the European Union has already embraced this: the Artificial Intelligence Act (AI Act) requires companies to ensure AI literacy of their people from February 2025. Learning to use AI properly is therefore not a hobby for enthusiasts, but an obligation.

The human factor will become a premium service

The fourth scenario assumes that some professions and services will require a human presence due to human nature itself. A customer in crisis wants to hear a person, not a voice from a speaker. A patient wants a diagnosis from a doctor, a client in court from a lawyer. And regulation is heading in the same direction: the AI Act directly requires human oversight measures for high-risk systems, and from August 2026 transparency rules take effect — the user must know they are communicating with a machine. Full application of the regulation falls on August 2, 2026, i.e. in less than two weeks; rules for high-risk areas will then take effect from December 2027 after an agreement on simplification.

What does this mean?

For users, practically nothing changes in terms of accessibility — ChatGPT, Claude, and Gemini are normally accessible in the Czech Republic and all handle Czech; DeepSeek and Kimi are used mainly via API. What will change, however, is the economics: expect a full subscription to mean roughly 500 CZK per month and demanding professions more likely in the thousands. For Czech companies, the message is clear: competitiveness will not be about having "the best AI," but about having people who can select the right model for the right price and give it the right task. The winner of the model battle will ultimately not be the smartest lab, but the best-configured combination of human and machine.