Re-reading: the line item nobody looks at

In agentic workloads your bill isn't driven by what the model writes. It's driven by what it re-reads.

An AI coding agent doesn't behave like a chatbot. On every turn it resends its entire history — system prompt, files it read, tool results, every prior exchange — and then generates a few hundred tokens. The typical ratio is several thousand tokens read for every thousand generated.

Billed per token, that re-reading is charged in full, every turn. A session that reaches 100,000 tokens of context and runs fifty turns did not cost 100,000 input tokens: it cost several million, because the same history was reprocessed fifty times.

Yet the real cost of that reading is low. Input is processed in parallel — a server ingests hundreds of thousands of tokens per second — while output is generated one token at a time, each one requiring a full pass over the model. On a dedicated server, the marginal cost of input sits around $0.09 CAD per million tokens. Usage-based providers bill it at 20 to 45 times that.

That's why we don't charge for input. Not out of generosity: charging for it would mean selling the least scarce resource in the machine at a premium. We charge for generation throughput, which is the actual physical constraint, and include the rest.

The practical consequence: the longer your AI agents work within a single context, the wider the gap grows in your favour. That is the exact opposite of what usage-based billing produces, which pushes teams to truncate history, compact too early, and throttle AI agents to contain the bill.