The cost of frontier AI work is rising

Imagine your team pays $200 a month for an AI plan. From today, a new subscriber on that plan gets half the usage for the same price. For new customers of OpenAI's Pro 200 plan, this is now true.

What OpenAI changed

On 29 September, OpenAI launched a $500 a month Pro plan with 25 times the usage of ChatGPT Plus. At the same time, it cut the $200 Pro plan from 20 times to 10 times Plus usage. The $200 plan reopens to new subscribers today with the lower allowance. Existing subscribers keep their old allowance until 29 October.

Per unit of usage, the $200 plan cost $10. It now costs $20. Every ChatGPT tier now has the same price per unit, so there is no longer a discount for heavy users.

OpenAI is not the only vendor to do this. On 1 June 2026, GitHub Copilot moved all plans from a fixed number of requests to billing by tokens used.

Token prices fall, but task costs rise

The price of a single token is falling fast. Epoch AI finds that the price to reach a fixed level of performance falls by a median of about 50 times per year.

But businesses do not buy tokens. They buy completed work. Reasoning models and agents use many more tokens to complete one task. Two recent studies measure the result:

• The Price of Progress found that the cost to run frontier models on benchmarks rose by 3 to 18 times per year.
• The Price of Intelligence, published this month, found that the price per completed task stopped falling in 2025 and 2026.

The move to public markets

Anthropic confidentially filed for an IPO on 1 June 2026, and OpenAI filed soon after. Reported investor projections show that Anthropic plans to increase its gross margin from about 50% to 77% by 2028. OpenAI projects $278 billion of negative free cash flow from 2026 to 2030.

Across the industry, Bain estimates that AI must earn $6 trillion a year by 2031 to justify current data centre investment.

Private investors accepted losses to get growth. Public shareholders measure margin every quarter. There are only two ways to increase margin: the cost to serve each task falls, or customers pay more. Flat subscriptions, where the price is fixed but the cost is not, will probably change first.

The dataintelligence view

We expect the cost of running AI at the frontier to rise by up to an order of magnitude over 2027 and 2028. We do not expect the list price per token to rise. We expect frontier work to use many more tokens, and flat-rate plans to move to usage-based billing.

What to do now

1. Measure the cost per task, not the cost per seat.
2. Use the smallest model that does each task well.
3. Monitor vendor plan changes each month.
4. Plan your 2027 budget for usage-based billing, not fixed subscriptions.

The organisations that know their AI unit costs now will have more options when prices change.