How OpenAI and Anthropic's usage-based pricing forces Brazilian AI startups to rethink revenue

The pricing documentation from Anthropic looks, on the surface, like a dull table of prices per million tokens. But it describes the economic engine behind nearly every AI product built in Brazil today, and it reveals a mismatch many founders haven't yet put into their spreadsheets: the infrastructure layer charges for consumption, while much of the Brazilian startups that consume it charge a fixed subscription. That difference in billing regime is where the margin disappears.

What the pricing table actually says

Anthropic's structure doesn't have a plan, it has a meter. Each model has one price for input and another for output per million tokens (MTok). Claude Opus 5 costs US$ 5/MTok for input and US$ 25/MTok for output; Sonnet 5 costs US$ 2 and US$ 10; Haiku 4.5 costs US$ 1 and US$ 5. There's no monthly fee: you pay for what the machine processes.

What's more revealing is what the documentation stacks on top of that. Web search is billed at US$ 10 per thousand searches, on top of the tokens. Code execution runs free for up to 1,550 hours per month, and after that it's US$ 0.05 per hour per container. Every tool the model invokes (bash, text editor, tool use) injects extra tokens into the request, which are counted in the final bill. An important technical detail for modeling: the Claude 4.7 and later models use a new tokenizer that produces about 30% more tokens for the same text, meaning the same prompt got more expensive in practice, even with the per-token price unchanged.

In other words: the more the agent thinks, searches and acts, the more it costs. Cost scales with autonomy. And that's exactly what the current phase of the market ("agents") most encourages.

Why this breaks the fixed subscription

The typical Brazilian AI product today charges something like R$ 99 or R$ 299 per month per user, in a classic SaaS model. That makes sense when the marginal cost of serving a customer is close to zero, as it was in traditional software. That's not the case when the backend is an agent consuming tokens.

The problem is usage distribution. In a subscriber base, a handful of heavy users trigger the agent dozens of times a day, each trigger setting off searches, code execution and long reasoning chains. With a fixed price, that user can consume more in API costs than they pay in subscription fees, while the subscriber who barely uses the product subsidizes the loss. Gross margin stops being a stable number and becomes a variable that depends on the behavior of the people you control least.

The logic of the infrastructure is the opposite: cost scales linearly with usage. Mismatching the customer's billing regime with the vendor's cost regime is a recipe for a margin that gets worse exactly when the product succeeds and engagement rises.

The path Anthropic itself signals

It's worth looking at how Anthropic itself tries to tame this cost, because these are the same levers the Brazilian founder needs to master before promising a price to the customer:

• Prompt caching: rereading a cached prompt costs 0.1x the input price (10%). A large system prompt repeated on every call, when cached, changes a product's economics.
• Batch API: a 50% discount on input and output for asynchronous processing. Ideal for tasks that don't need a real-time response.
• Model selection: routing a simple task to Haiku (US$ 1/US$ 5) instead of Opus (US$ 5/US$ 25) is a 5x difference in cost.

Whoever doesn't instrument these controls doesn't have a product with margin, they have a bet on the AWS bill. And the strategic implication is direct: the price charged to the customer needs to reflect the cost structure behind it, whether by passing through consumption or by charging for delivered outcomes.

| Regime | How it charges | Margin risk | |---|---|---| | Fixed subscription | Fixed monthly fee per user | High: heavy usage erodes margin | | Usage-based (consumption) | Passes through tokens/actions with markup | Low: cost tracks revenue | | Outcome-based | Charges per completed task | Medium: depends on correctly pricing the deliverable |

The counterpoint: pure usage also has a trap

It would be easy to conclude that everyone should just migrate to consumption-based billing and be done with it. It's not that simple, and the best counterargument deserves to be faced.

Usage-based billing has a real commercial cost: it transfers anxiety to the customer. A price that varies every month is harder to sell, harder to approve in a corporate budget, and harder to forecast. Companies buy software precisely for predictability. That's why Anthropic itself offers the Claude Consumption Units model on the AWS and Azure marketplaces, an abstraction layer that packages consumption, and why postpaid billing with a single monthly invoice exists. The infrastructure knows that raw consumption is scary.

The balance point for the Brazilian founder is probably neither pure subscription nor pure consumption, but hybrid models: a fixed base that covers the minimum cost of serving, with incremental charges per agent action or per result above an allowance, much like the 1,550 free hours of code execution before the per-hour price kicks in. Outcome-based billing (per completed task, per resolved ticket, per processed document) is the more sophisticated version of this, but it requires the founder to know precisely how much it costs to deliver each unit, which only per-token cost instrumentation allows.

What changes in the conversation with investors

The implication for those raising capital in Brazil is the one that appears least in the headlines. SaaS investors learned to look at 70% to 80% gross margin as a sign of health. An AI product with an agent at its core doesn't naturally deliver that: COGS has a heavy variable component, inference cost, that didn't exist in traditional software.

This forces two changes in the thesis presented to the fund. First, gross margin needs to be explained, not hidden: how much of the cost is inference, how it evolves with volume, and which levers (cache, batch, smaller model) are under control. Second, retention and expansion metrics change meaning, because in a usage-based product, revenue expansion tracks consumption expansion, not the sale of more seats. A founder still pitching subscription ARR for a product that consumes tokens is describing a business the cost structure can't sustain.

This reading, here, is my own: the Brazilian market tends to go through this repricing with a lag compared to the American one, and whoever adjusts the model before the margin breaks (not after) arrives at the next round with numbers that tell a consistent story. Anthropic's pricing table isn't just the vendor's bill. It's the blueprint for the business model that the Brazilian product will have to copy.