OpenAI announced improvements to GPT-5.6 Sol in ChatGPT alongside expanded access to GPT-5.6 Luna for free users.

Free-tier changes get covered as consumer news. They are also, quietly, one of the larger determinants of who can do certain kinds of research.

The tier is the ceiling for a lot of people

For researchers at well-funded labs, free-tier capability is irrelevant — API budgets exist. For a large share of the world's researchers, it is not:

• Students at institutions without API budgets.
• Researchers in countries where card-based payment for foreign services is impractical.
• Independent researchers and small groups without institutional billing.
• Anyone in a setting where procurement takes months.

For these groups the free tier is not a preview of the paid product. It is the product, and its capability sets the ceiling on what they can attempt.

Why this shapes what gets studied

If frontier-adjacent capability is available at zero cost, a student can test an idea over a weekend. If it costs a few hundred dollars to find out whether an idea is worth pursuing, the idea does not get tested — and the ideas that die are disproportionately the exploratory ones, since those are the ones with the lowest expected value per dollar.

That is a selection effect on the research literature, applied before anyone writes a paper. It is invisible in the published record, because you cannot see the experiments nobody ran.

What matters more than the headline

Three details determine whether a free tier is research-usable, and they are usually buried:

1. Rate limits. A capable model at ten messages a day cannot support an experiment.
2. Context window on the free tier. Often substantially smaller than the paid tier, which rules out document-scale work entirely.
3. API access versus chat-only. Chat-only access cannot be scripted, which excludes any systematic evaluation. This is the single biggest divider, and it rarely appears in announcements.

A generous free tier that is chat-only is a fundamentally different thing from a modest free tier with API access, for research purposes.

The uncomfortable part

Free tiers exist for commercial reasons and can contract as easily as they expand. Building research infrastructure — or a thesis — on a free tier is building on a commitment nobody made.

That argues for the same thing it always argues for: open-weight models with permissive licences remain the only access guarantee that does not depend on a company's continued goodwill. Free-tier expansion is genuinely good news for access, and it is not a substitute for that.

What institutions can do about it

For groups that cannot rely on a free tier remaining generous, a few things reduce exposure:

Budget for API access explicitly in grant applications. Compute lines are standard; inference lines often are not, and reviewers increasingly expect them.

Prefer open-weight models for anything that must be reproducible. A result obtained against an API that has since changed is difficult to reproduce and awkward to defend. Open weights fix the artefact.

Cache aggressively and share within the group. A shared, cached evaluation harness across a lab costs a fraction of individual usage, and most groups have never set one up.

Record the model version and date with every result. This costs nothing at the time and is the single most common omission that makes an evaluation impossible to reproduce a year later.

None of this replaces genuinely accessible frontier capability. It does reduce the number of research programmes that quietly depend on a pricing decision made by someone else.