What a million tokens costs in 2026 and why a subscription often wins

What a token is and why there are so many

A token is a chunk of text the model works with: part of a word, a whole word or a punctuation mark. Code produces even more tokens because of indentation, brackets and variable names.

The key thing about per-token billing: you pay not only for the answer but for everything you send. The model has no memory of the conversation, so every request carries the whole history, instructions and attached files again.

How the bill grows in practice

Take a regular Claude Code or Cursor session. The agent reads five 400-line files and keeps project instructions and step history in context. One request easily weighs 30–60K tokens. A task takes 20–40 steps, so 0.6 to 2.4 million tokens per task.

Ten such tasks a day mean 6 to 24 million tokens per developer per day, hundreds of millions a month. That is why our home page shows an example: 500 million tokens a month on Western APIs already means thousands of dollars, even when the per-million price looks modest.

Where the extra tokens hide

• Resending the conversation history with every request
• Long system prompts and project rules
• Files the agent rereads at every step
• Retries after errors and timeouts
• Parallel agent branches and helpers

When per-token billing is cheaper

If you make a few requests a day, paying for actual usage is fairer. The same goes for experimenting with many different models, where pay-per-token aggregators are handy.

When a subscription wins

Once AI becomes a daily tool (editor sessions, agents, bots, batch jobs), usage becomes large and, worse, unpredictable. A fixed-price subscription solves both: the monthly cost is known upfront and does not depend on tokens.

At Tenzo Cloud you pay for streams, how many requests the model handles at once. One stream for yourself, four for a team or an agent. Tokens are not counted at all.

Example: one month of a small team

Take three developers who work with an agent in their editor every day. In our observation an active developer with an agent uses 5 to 20 million tokens a day, depending on project size and how often the agent rereads files.

Even at the low end that is 5 million × 3 people × 21 working days ≈ 315 million tokens a month. At the high end, over a billion. A threefold spread is the real problem: the budget cannot be planned and finance learns the total from the invoice.

With a stream subscription the same team takes, say, 4 streams for a month. That is enough because three people rarely send requests in the same second. The price is known upfront and does not change even if all three leave agents running overnight at month end.

How to compare a subscription and per-token billing fairly

1. Use real stats for 1–2 weeks, not a guess
2. Count both input and output tokens; many providers price them differently
3. Add headroom: with agents usage grows as the team gets used to them
4. Compare with a subscription sized for your peak-hour concurrency
5. Include non-money costs: VPNs, foreign cards, time spent watching the budget

What else affects the cost

Prompt caching at some providers lowers the price of the repeated part of the context, but it does not always apply and needs careful setup. Discounted batch processing only fits tasks where answers can wait for hours.

A subscription has trade-offs too: you pay for streams even on quiet days. So for occasional use per-token billing stays reasonable, while a subscription wins where AI is a daily tool.