On 9 August 2026, OpenAI opened “Luna,” part of the GPT-5.6 family, to free users without limits. Releasing a current-generation model with no paywall drew a large reaction.
More interesting than the free release itself is the way the models were split three ways.
Luna, Sol, Terra: three branches
| Model | Character |
|---|---|
| Luna | Efficiency-focused. Runs at far lower cost than Terra. The one released free |
| Terra | Balance of performance and speed. Feels close to the previous generation at less than half the price |
| Sol | Security-specialised. Optimised for finding software vulnerabilities and producing patches |
What this arrangement says is clear. The era of handling everything with one large model is over. Instead, size and price are split by purpose, and the cheapest one is released free to widen the entrance.
Sol stands out in particular. Optimising for vulnerability detection and patch generation means security work has been targeted at the model level. That a dedicated model has appeared for code verification rather than code generation is itself a marker of this market’s maturity.
Why give it away
The logic of a free release comes down to three things.
1. Drive the cost of entry to zero
The biggest barrier when a developer tries a new model is not performance but “the money and paperwork of just trying it”. If the newest model can be used without registering a card, trials go up, and when trials go up some share converts to paid.
2. Data
More free users means more conversation data, and that data is the raw material for the next model. The cheaper the model, the more you can run it; the more you run it, the more data accumulates. Luna being designed for efficiency and Luna being released free share the same root.
3. Competition
Google is pushing AI in as a default feature through Android and Search, Apple through Siri. The less time users spend thinking “to use AI I have to open an app,” the worse it is for standalone apps. A free release is also a defence that slows that timetable.
What to watch when putting a free tier under real work
Being able to use something free and being able to build on top of it free are different things.
- Policies change without notice. Quotas, rate limits and model routing are the provider’s discretion
- Models get swapped quietly. Different weights running behind the same name is common. A prompt can behave differently from yesterday
- Data handling terms differ. The training-data policy for a free tier is not the same as for API or enterprise agreements. Confirm before attaching internal documents
- There is no availability guarantee. When free traffic surges, free is the first thing throttled
The safe principle for production use is one line. Write it so the model can be swapped out. Keep prompts and post-processing separate from the model name and you can move within a day when a provider changes policy.
The objection from the regulatory side
Not everyone sees a free release kindly. The argument is that when the user base explodes, monitoring for abuse becomes practically impossible. There have also been claims that this collides with the intent of the EU AI Act, which demands strict oversight of high-risk systems.
This is not an abstract worry but a question of cost. Actually monitoring the conversations of a hundred million free users becomes another large-scale infrastructure in itself. The claim is that a structure which scales without bearing the monitoring cost colliding with regulation is only a matter of time.
What individual users should do
No need to overcomplicate it.
- For everyday use, free models have entered the good-enough band. Summarising, translating, drafting and replacing search are mostly solved by a current free model
- The points where paid is needed are elsewhere. Working through entire long documents, long-running agent tasks, and heavy repetitive workloads. That is a quota question, not a performance question
- Do not put everything on one provider. Compare free tiers and use whichever fits your hand. This is the cheapest moment there has ever been to switch
In short
- On 9 August 2026, OpenAI opened GPT-5.6 Luna to free users without limits
- The models split three ways: Luna (efficiency), Terra (balance), Sol (security). Differentiation by purpose is now explicit
- The purpose of a free release is removing entry cost, gathering data, and defending against built-in-assistant competition
- When putting it under real work, design so the model can be swapped at any time. Free tier policy is the provider’s discretion
- A regulatory counterargument came alongside: the larger the scale, the harder abuse oversight becomes
Model names, prices and policies change every few months. This article reflects the state of things as of 27 August 2026.
Frequently asked questions
Is there a big performance gap between free and paid models?
Within the same generation, the perceived difference on everyday tasks is often small. Where the gap widens is very long context, tasks that chain many steps autonomously, and usage limits. If you only do short question-and-answer, free has entered the good-enough band.
Can something released free go back behind a paywall?
It can. The scope and limits of a free tier can be adjusted by the provider at any time, and several services have done exactly that. Which is why building a work process on the premise of a free model is risky.
Can I put company documents in?
Data handling policy differs by account type. It is common for the training-data terms of a personal free account and an enterprise agreement to differ, so check the terms that apply to your account before feeding it internal material. If you cannot confirm, the principle is not to.
Can individuals use the security-specialised model?
Vulnerability analysis models are mainly offered through APIs or developer tools. If you use one as an individual, limiting it to checking your own code is the right scope — using it against systems you have no authorisation for is illegal whatever tool you use.
Leave a Reply