On May 14, Axios captured in a single sentence what the AI market had been whispering for months: "all-you-can-eat AI subscriptions may not survive the agent era." The article reported a move by Anthropic — placing the use of external agent tooling behind a separate credit meter, even for paying customers — and the immediate counter from OpenAI, which on the same day made Codex available for free inside the ChatGPT app on iOS and Android. The market reading is clear: the flat-fee subscription model, the pillar of the chatbot era, is under structural pressure in the agent era.
The thesis is not new among operators, but the past few days gave it empirical weight that is hard to ignore. Anthropic and OpenAI, the segment leaders, are making opposing bets about how to price intensive AI usage — and what is at stake directly affects the technology budgets of Brazilian companies that are starting to run agents in production.
Anthropic's move: agents on a separate meter
Anthropic announced it will once again allow the use of external agent tools on paid Claude plans — but behind a dedicated credit meter, decoupled from the monthly subscription. According to Axios, the decision reflects an operational reality: agents consume volumes of tokens far above conversational use, and absorbing that cost inside a fixed monthly fee was becoming economically unworkable.
Context matters. Days earlier, on May 6, the company had doubled Claude Code usage limits for Pro, Max, Team, and Enterprise per-seat plans, and lifted Tier 1 API input limits by up to 1,500%, according to an official note on Anthropic's website. To support that expansion, the company struck a deal with SpaceX to use the Colossus 1 data center capacity — over 300 megawatts and roughly 220,000 NVIDIA GPUs. In short: Anthropic is expanding capacity, but clearly separating what fits inside the subscription from what needs to be charged separately.
OpenAI's move: Codex free on mobile as bait
On the same May 14, OpenAI went the opposite direction. Codex, the company's coding assistant, became available inside the ChatGPT app on iOS and Android — including for free-tier users, as reported by Engadget and TechCrunch. Sam Altman added on X that new business customers will get two months of free Codex usage.
It is worth understanding what the mobile app does and does not do. It acts as an intermediary: developers do not run code on their phones, but they can monitor tasks started in Codex desktop, approve commands, switch models, and kick off new runs. For now, the mobile app only connects to the macOS version of Codex desktop; Windows support has been promised without a date. The play is clear: OpenAI is trying to pull in the coding power users that Anthropic just unsettled with its new credit policy.
Why agents break the subscription economy
The difference between chat and agent, from a cost standpoint, is orders of magnitude. A chatbot conversation consumes tens to a few hundred thousand tokens. An agent that reads files, navigates codebases, runs tests, and iterates on results consumes millions of tokens per session, often in parallel. The "pay a monthly fee and use freely" model was designed for the first scenario, not the second.
There is also an aggressive Pareto effect at play: a small fraction of intensive users consume most of the resources. When average usage was moderate, that imbalance could be absorbed by the platform's margin. With agents, a single intensive user can consume the equivalent of hundreds of subscriptions on their own. Anthropic's decisions reflect that math: separating agent usage into its own meter is the cleanest way to align cost and price without burdening users who treat Claude as just a chat product.
The counterpoint: subscriptions still make sense for many
It is important not to fall for the easy narrative that "the subscription model is dead." For conversational use, text generation, document summarization, and individual productivity tasks, the flat subscription remains the simplest and most predictable format for users and companies alike. Both Anthropic and OpenAI maintain per-seat plans that comfortably cover those scenarios, and Anthropic itself doubled Claude Code limits precisely to avoid friction with the typical user.
What is happening is a bifurcation in pricing: subscriptions for interactive use, consumable credits for agentic use. It mirrors what happened in the cloud last decade, when annual licensing coexisted with elastic per-hour or per-call consumption. IDC projects in a recent report that by 2028 about 70% of software vendors will have refactored their pricing around consumption, outcomes, or organizational capability metrics — and AI is simply pulling that shift forward.
What Brazilian leaders should do
Brazilian companies entering the agent era now need to revisit budgeting assumptions inherited from the chatbot world. A few areas deserve immediate attention:
- Map current usage by profile: how many employees use AI as chat versus as agent. The cost gap between the two groups can be one or two orders of magnitude.
- Model TCO by use case, not by user. A single agent running on a continuous pipeline can justify a dedicated consumable plan, while fifty conversational analysts fit comfortably into per-seat subscriptions.
- Avoid lock-in to a single vendor, especially while pricing structures are in transition. Multi-model strategies become a financial risk hedge, not just a technical one.
- Negotiate contracts with revision clauses: with Anthropic and OpenAI tweaking policies every few weeks, rigid annual contracts may age poorly.
Conclusion
The fight between Anthropic and OpenAI over how to charge for intensive AI usage is not just a commercial detail — it is the design of the next phase of monetization in the sector. For Brazilian companies, the message is practical: treat agentic AI as a variable infrastructure cost, not as subscription software. Those who enter this phase with chatbot-era thinking will discover, on the invoice, that the all-you-can-eat era has a limit. Those who arrive prepared will capture the productivity gains from agents without being blindsided by the cost.
This article was published on May 15, 2026. Follow Entercast to stay ahead of the next developments.