Skip to content

Is CPM profitable? LLM Economics by the End of 2026: Prices, Margins, and the Future

9 min read
CPM выгодный? Экономика LLM к концу 2026 года: цены, маржа и будущее — illustration 1

In the rapidly developing world of artificial intelligence technologies, the question of whether CPM is profitable or not becomes a cornerstone for businesses. We constantly hear about the decreasing cost of «intelligence,» yet flagship models become more expensive, and AI expenses are growing. How to understand this paradox? In this article, we will delve into the economics of large language models (LLM) by the end of 2026, examine real prices, provider margins, and try to predict the future.

How much does a token cost now — by price list and actually?

In previous materials, we already touched upon the topic of local inference, however, corporate practice largely relies on cloud models. Here, the cost issue becomes critically important. Despite researchers’ statements about the decreasing cost of «intelligence,» flagship models released by 2026 turned out to be more expensive. Providers are actively cutting limits, and overall AI expenses continue to grow. Let’s understand the pricing and real economics.

By the end of 2026, the LLM market had formed a clear three-tier structure:

  • Super-flagships: Claude Fable 5.1 and GPT-6 Astra.
  • Working flagships: Claude Opus 5.5, GPT-6 Sol, Gemini 3.5 Flash.
  • Mass models: GPT-6 Luna, DeepSeek V4.1-Flash.

The standard price list for 1 million tokens looks like this:

  • Claude Fable 5.1 (Anthropic): $75
  • GPT-6 Astra (OpenAI): $60
  • Claude Opus 5.5 (Anthropic): $30
  • Gemini 3.5 Flash (Google): $15
  • DeepSeek V4.1-Flash (off-peak tariff, mass, open weights): $2

However, it is important to understand that this price list does not reflect real costs. According to Ramp, the weighted average price that American companies actually pay OpenAI and Anthropic was about $0.56 per million tokens by October 23, 2026. This is explained by the fact that the main volume of requests comes from re-reading context from the cache and using cheaper models.

Martin Alderson showed that the cost of a long agent session is determined by the cache read rate, not the input/output price. The same task can cost from $0.23 to $25. A difference of two orders of magnitude.

Thus, model selection and caching architecture have a much stronger impact on the budget than changes in price lists. This is not an assumption, but a confirmed fact. CPM optimization becomes critically important for any AI project.

Are these prices economically justified?

The Epoch AI report titled “The plunging price of thought” (September 2026) claims that the cost of achieving a fixed quality is decreasing by approximately 13 times per year. This rate outpaces the cost reduction of DNA sequencing or computational power.

Guido Appenzeller (a16z) as early as 2024 called this phenomenon “LLMflation” and estimated the cost reduction at tenfold per year.

It is important to note that this curve refers to a fixed level of capability. Ethan Ding rightly observed that demand constantly shifts to the newest models, whose price remains relatively constant, while the number of tokens required to perform a task increases. If we extrapolate Epoch’s rate, the price of a fixed level of capability has decreased by ten thousand times. However, over the same period, the price of output tokens for flagship OpenAI and Anthropic models remains in the range of $10–$75 per million.

In 2026, new models indeed became more expensive. Matthias Bastian points out that Gemini 3.5 Flash is more expensive than its predecessor, and new versions of Claude consume more tokens to perform the same task. According to the Ornn index, the actual prices of closed models have increased, while the price of the open-source DeepSeek, on the contrary, has decreased.

Are subscriptions profitable for providers?

Anthropic’s prospectus, published on October 26, 2026, provided verifiable financial data for the first time. The company’s gross margin was about 85% in the first half of 2026. Direct computing costs accounted for about 15% of revenue. For comparison, in 2024, this figure was negative. About 80% of Anthropic’s revenue comes from pay-per-token, and 20% from subscriptions.

Where does the revenue go? About 60 cents of every dollar is directed to research and development (model training). Another 23 cents go to sales and marketing, and 6 cents to administrative expenses. Only 15 cents are spent on directly serving requests.

OpenAI’s financial situation is different. With revenues of $5.7 billion and $6.7 billion in the first and second quarters of 2026, respectively, the company’s operating loss was $9.3 billion and $12.3 billion.

The reason for this discrepancy is that a significant portion of ChatGPT users use free or inexpensive plans. OpenAI’s gross margin was 33% in 2025 and 39% in the first quarter of 2026. Anthropic, on the other hand, is primarily focused on corporate clients and programmatic access.

Martin Alderson stated back in 2025: “The API business is essentially a money printer.”

Dario Amodei, CEO of Anthropic, formulated the logic as follows: “Every model is profitable, but the company is unprofitable every year.” That is, each model generates profit, but the company as a whole remains unprofitable due to colossal investments in developing the next generation of technologies.

The prices of frontier models are justified not so much by the cost of the token, but by the following factors:

  1. The need to recoup the training costs of models with a relatively short life cycle.
  2. A premium for technological leadership and innovation.
  3. The shortage of computing power observed in 2026.

The conclusion is obvious: overpaying for a powerful model is justified only when it brings a measurable and significant effect. A month after the launch of Claude Fable 5, it accounted for only 6% of tokens purchased from Anthropic. This is a clear indicator of the limit of companies’ willingness to pay for quality improvement. ROI in AI investments must be calculated very strictly.

CPM profitable? The economics of LLMs by the end of 2026: prices, margins, and the future — illustration 2

Subscription economics: limits instead of prices

Sam Altman (OpenAI) admitted back in January 2025 that the company was losing money on the Pro plan due to intensive use. Ethan Ding explained this by stating that with autonomous agent operation, token consumption becomes virtually unlimited. “There is no flat subscription price that works in this new world,” he noted. A fixed subscription price simply does not work in the new reality.

Providers responded by managing limits. Anthropic introduced weekly limits, and OpenAI reduced limits for Plus plan users. Chris Zeoli commented: “Instead of raising list prices, they cut usage allowances.” That is, instead of raising price lists, providers reduced the permissible usage volumes.

In the mass segment, OpenAI is gradually shifting to an advertising model, expecting a reduction in the number of Plus subscribers and compensation through the Go advertising plan.

The subscription margin calculation model looks like this:

  • Subscriber cost C = A × k, where A is the monthly consumption at API price, k is the ratio of cost to price (from 10% to 30%).
  • Margin M = P − C, where P is the plan price.
  • Break-even point: A = P / k.

Even assuming that 15% of users are “heavy” and extreme, the margin under the base scenario is about 56%. Consequently, limits are the main tool for managing subscription economics.

How will prices change in 2027–2029?

The cost of a token will likely continue to decrease, but the cost of performing a task on frontier models will not. The market will be influenced by opposing forces:

  • Technological trend: The cost reduction of a fixed level of capabilities is not slowing down.
  • Inference efficiency: Gartner predicts that by 2030, inference of a trillion-parameter model will cost 90% less than in 2025. However, this cost reduction will not be fully passed on to customers.
  • Price Competition: OpenAI reduced prices for Luna and Sol, Anthropic made its flagship model cheaper. Open models remain several times cheaper than closed ones.
  • Compute Shortage: The cost of renting H100 accelerators increased by almost 40% from October 2025 to March 2026. Spot prices for B200 rose by 114% in six weeks in spring 2026. AWS raised GPU power prices by 20%.
  • Memory: Contract prices for RAM almost doubled in early 2026. HBM, used in accelerators, will become several times more expensive by 2027.
  • Increased Consumption per Task: An agent requires 5–30 times more tokens than a typical chatbot.
  • Tomasz Tunguz: «Inflationary demand outpaces deflationary algorithmic & chip improvements».

    Gartner’s final forecast is discouraging: by 2028, the cost of a single agent process will increase by more than five times, despite the reduction in token prices. This phenomenon has been dubbed the “inference paradox.”

    What does this mean for companies and specialists?

    Will Sommer (Gartner): «Product leaders cannot expect that token price reductions alone will justify AI costs».

    This means that unit economics must be calculated extremely strictly. Favorable CPM is only achieved with the right model and architecture selection. Hypothesis testing can be done cheaply on open models. The price comparison between a banner and a blogger is now complemented by a comparison with AI solutions. Is it worth investing in AI? Only if the ROI is obvious. Price versus result is the key metric. How to reduce CPM is not just about tokens, but about the efficiency of the entire solution architecture.

    Frequently Asked Questions

    What is the real CPM on cloud LLMs?

    The real weighted average CPM for American companies is about $0.56 per million tokens, which is significantly lower than the stated price lists of flagship models. This is due to the use of caching and cheaper models for the main load.

    Why are new flagship models becoming more expensive if “intelligence” is getting cheaper?

    “Intelligence” is getting cheaper for a fixed level of capabilities. However, demand is shifting to newer, more powerful models, whose cost remains high. Providers are forced to recoup huge investments in training the next generation of models, which maintains high prices for frontier solutions.

    Are LLM subscriptions profitable for providers?

    Yes, subscriptions are profitable for providers, not counting the colossal investments in research and development of new models. Anthropic’s gross margin is about 85% on current operations. However, due to intensive use, providers are forced to introduce limits rather than raise prices to manage the economics of subscriptions.

    What is the «inference paradox»?

    The «inference paradox» is a situation where the cost of tokens continues to decrease due to technological advancements, but the overall cost of performing a task using AI increases. This happens due to the increasing complexity of tasks, the growth in token consumption per task (especially for agents), and the scarcity of computing power.

    What CPM should be considered profitable for AI projects?

    A profitable CPM depends on the specific task and ROI. If overpaying for a flagship model brings a measurable effect and converts into revenue, then such a CPM may be justified. In other cases, it is worth focusing on cheaper open models or optimizing the caching architecture to reduce actual costs.


    ← Back to blog