OpenAI and Anthropic change how LLM costs are calculated: in 2026, the price of a task is what matters

The market for large language models is entering a new phase: a cheaper token no longer means a predictable budget. Almost simultaneously, OpenAI and Anthropic have shown that in 2026 businesses will have to count not only the price of a million tokens but also the full cost of completing a task. For companies building products on agentic scenarios, this changes the very logic of procurement, planning and unit economics.

The first signal came from Anthropic. The company moved agentic frameworks to usage-based billing, that is, payment for the tokens actually consumed instead of fixed subscriptions. In practice, this means that some external wrappers and services that could previously live on a flat-rate model are losing their financial footing. While the load was relatively predictable, a subscription looked convenient for both provider and customer. But in agentic systems compute consumption grows quickly: the model does not just answer a single request, it plans steps, makes several calls, uses tools, double-checks the result and can launch a long chain of actions.

In parallel, OpenAI changed its approach for corporate customers. In the Enterprise, Business and EDU plans, the company introduced more flexible pricing in which cost scales with usage volume rather than staying rigidly tied to the number of seats or seat licenses. For buyers, this is an important shift. Not long ago, a subscription could be treated as an almost fixed line item; now the model is closer to cloud services, and the payment depends more heavily on actual usage intensity. The more actively employees use generation, search, document analysis and agentic features, the more noticeably the bill changes.

This does not cancel another trend the market has watched for the past two years. From 2023 to 2025, APIs did get cheaper, and the cost of a million tokens for GPT-4-class models fell. That is why many teams got used to a simple rule: if the token price falls, adopting LLMs automatically becomes more profitable over time. In 2026, this rule no longer works without caveats.

The key metric now is not the price per token in itself but the cost of solving a specific task. If a system needs several passes, a long context, tool calls, extra checks and repeated generations to produce one useful result, the total bill can grow even when the API is nominally cheaper. This is especially visible in agentic products, where a scenario that looks like a single action to the user can break down internally into dozens of model operations.

There is a practical conclusion for teams here. LLM budgeting now has to be built around the cost of a completed action: how much one report, one document analysis, one assistant session or one successfully executed agentic workflow costs. This brings new product requirements: limiting unnecessary steps, controlling the depth of agentic reasoning, shortening context, caching, routing to cheaper models where acceptable, and strictly measuring which calls actually create value. For CTOs, CPOs and finance teams, this means moving from talk of "cheap AI" to ordinary operational economics, where what matters is not an attractive price in a table but the price of a specific business result.

The main point of this shift is that the LLM market has not stopped getting cheaper; it has stopped being naively simple. The compute crunch in 2026 is not only a question of available capacity but also of cost controllability. The winners will not be the companies that look at the lowest price per token, but those that can calculate the cost of the end result and design the system so that every additional token brings measurable benefit.