fees

As end users, pay a monthly fee to use a AI model It is the norm to access more complete and powerful models. However, developers who rely on an AI model to power their tool or application pay based on the tokens of input and output that are consumed (the minimum unit of text that a model processes when we use it, so that we understand each other). Which has announced GitHub Copilot has more to it than it seems, as it will now begin charging end users through a monthly plan based on the number of tokens. And this has set off alarm bells in the sector, because it could be a move that any other company could easily end up imitating. And all in a context in which Chinese startups prices continue to drop sharply in their models. Copilot can no longer maintain its business model. GitHub has announced that starting June 1 it will stop accepting requests for its current premium plans and will begin billing for AI credits instead. Each monthly plan will include a number of credits equivalent to the price of the subscription: anyone who pays $10 per month for Copilot Pro will receive $10 in credits. From there, consumption is measured in tokens, including input, output, and cache tokens. It is a play similar to when we use a image generation model either video: a use that depends on credits and that we recharge depending on the use. The reason for the change, according to the companyis that until now a quick consultation and an autonomous programming session of several hours cost the user the same. GitHub claims to have long absorbed that cost difference, but acknowledges that the model is no longer viable. QWhat exactly changes. The base prices of the plans are not touched: Copilot Pro is still at $10. Business in 19. Enterprise in 39. But: what you buy with them is no longer the same. Previously, the limit was a number of requests. Now, each interaction with the model consumes credits at a rate that depends on the chosen model and the volume of tokens. According to the rates published by the company itself, the most advanced OpenAI models can cost up to $30 per million output tokens. On the other hand, an agentic session, where the assistant executes tasks autonomously, can easily multiply the expense of a week of normal consultations. Ed Zitron, well-known critic and technology expert, counted that, according to internal documents to which he had access, Copilot’s weekly costs had almost doubled since January, coinciding with the boom in agentic assistants. Nor is it just Copilot. According to account The Information, Anthropic has begun charging its large enterprise customers the actual cost of computing Claudeabandoning any discount. Anthropic itself briefly tested the elimination of Claude Code of its $20 per month Pro plan. Large AI companies have been taking losses on their subscription models to attract users for some time, and are now trying to pass on the real costs to those who consume the most. China does the opposite. While the West adjusts prices upwards, several of the main Chinese technology companies have adopted a completely different strategy: turning tokens into a cheap commodity, almost like a telecom distributing mobile data. DeepSeek announced this week a 90% reduction in the price of cached accesses to its API (when the model reuses already processed context), bringing the minimum entry cost to about $0.14 per million tokens. For your most advanced model, DeepSeek-V4-Prothe figure becomes 32 times cheaper per conversation than the equivalent in GPT-5.5 from OpenAI, according to company data. Alibaba, for its part, has just separated its AI business and renamed it Token Hub Business Group, making clear what its strategic commitment is. According to share According to Reuters, Chinese models cost on average one-sixth the price per token of those from OpenAI, Anthropic, and the like. Why it can work, and why it has a limit. China’s advantage in inference (the moment at which a model responds to a request) rests on cheaper electricity, software efficiency that it has had no choice but to forcefully develop by chip restrictions from Washington, and a super competitive domestic race that forces prices to constantly drop. Token consumption in China has gone from 100 billion a day at the end of 2025 to 140 billion in March 2026, according to estimates collected by Reuters. However, as the media points out, this strategy has an underlying problem: the tokens are not interchangeable. One million tokens from Anthropic’s most advanced system are worth much more than the same volume processed by an inferior model. Companies that delegate complex tasks to AI agents will end up paying for quality, not just volume. And there, the Chinese models continue to lag behind the most advanced Western ones. Cover image | Alexander Mils and Roman Synkevych In Xataka | Anthropic decided to resist pressure from the Pentagon. Since then all other technologies have folded