OpenAI made two announcements this week that look unrelated. One is generous. The other explains how it can afford to be.

The generous one: OpenAI is giving 100,000 academic researchers free access to its frontier models through 2027, Axios reported. The programme, ChatGPT for Academic Researchers, starts with 10,000 users this summer and scales from there.

Free frontier access for scientists

Researchers get GPT-5.6 Sol Pro, OpenAI’s top model, and can each invite up to four collaborators, the company said. Their data is not used to train models by default. Early institutions include the Institute for Advanced Study and the École normale supérieure.

OpenAI frames it as accelerating science, and it forms part of a commitment of more than $250m through 2027 to fund outside research. Greg Brockman called it “more shots on goal against humanity’s hardest problems.”

Not everyone reads it so warmly. Sceptics note the flywheel: hook researchers on a capped compute budget, learn from what they do, and keep the frontier compute in-house. Getting a generation of scientists to think inside ChatGPT is its own kind of moat.

The engine underneath

The second announcement is drier, and it is the reason the first is possible. OpenAI detailed how it slashed the cost of running its agents, The Deep View reported.

The key piece is the agent harness. It sits under Codex and ChatGPT Work and directs the model, the tools and the context like a conductor. It is open-source, unlike Anthropic’s Claude Code. It is also a token furnace. Early this year, some developers ran up $20,000 monthly bills as agents burned through compute.

So OpenAI optimised it. GPT-5.6 Sol now beats Claude Fable 5 on a leading coding benchmark while using 54% fewer output tokens, the company says. A lighter model, GPT-5.5 Luna, costs 80% less than Sol.

The model that cut its own bill

The neatest detail is recursive. Using Codex, GPT-5.6 rewrote and optimised OpenAI’s own production kernels, the low-level code that runs the models, The New Stack reported. OpenAI says that lifted token efficiency by more than 15%. The model helped cut its own cost.

The timing is not an accident. Enterprises have soured on “tokenmaxxing,” the habit of throwing raw compute at every task. Databricks says curbing AI cost is the top question it now hears from customers.

Two sides of one strategy

Put together, the two moves are a single play. Cheaper inference lets OpenAI give models away at the top of the funnel, to 100,000 scientists and to the other 990 million ChatGPT users it wants on agents. Generosity and cost control are the same strategy.

There is a risk in it. Free frontier tools deepen the field’s dependence on one company’s stack, at the exact moment OpenAI is learning to run that stack for less. The cheaper it gets to serve, the more the world it serves belongs to OpenAI. Anthropic, as one poster put it, gets the next move.

Get the TNW newsletter

Get the most important tech news in your inbox each week.