On Thursday the ninth of July, GPT-5.6 went public. OpenAI opened the family, three models named Luna, Terra and Sol from the lighter one to the most capable, to every firm and every individual. A fortnight earlier they were not on sale in the ordinary way. At the request of the United States government, the company had held the release to a small group of what it called trusted partners while agencies ran further testing. The Department of Commerce then cleared a broad launch, and the models went to everyone.
What happened
OpenAI was direct about how it viewed the arrangement. It said it does not believe this kind of government access process should become the long-term default, and that holding the tools back keeps them from the users, developers, enterprises and defenders who need them. The politics are for others. The point for a law firm is the timing. For about two weeks the strongest tier of general AI sat behind a barrier. Now it is on the open market.
This followed the same shape of event at Anthropic a week or so before, when its Mythos 5 model was cleared for around a hundred American organisations under a specific arrangement and then had the licence requirement that fenced it off removed. So inside a single month the two leading American labs each took their most capable model, placed it briefly behind a government barrier, and then released it. The barrier is down and the models are the best either company has shipped.
Why a smaller firm should notice
The tier of model that was locked away a fortnight ago is now something a two-partner firm can subscribe to before lunch. The distance between what the largest firms can reach and what you can reach, on the raw tool, has closed to almost nothing. The frontier is no longer rationed by size or budget. It is a purchase, and a modest one.
That cuts against the instinct that the best tools belong to the biggest players. A sole practitioner and a thousand-lawyer firm now type into the same systems. What separates them is not access. It is the training, the checks, the written policy and the judgment about which task the tool touches and which it does not.
The capability is not the point
Be careful about what a stronger model gives you, because the marketing runs ahead of the truth. A better model writes more fluently and holds a longer document in mind, which saves real time in drafting and in turning a messy bundle into a working chronology. What it does not do is become reliable in the way a source of law is reliable. It still produces confident text that is sometimes wrong, and the better it writes, the more convincing its errors look and the harder they are to catch at the end of a long week. A weak tool that stumbles warns you. A strong tool that glides does not.
The courts have kept proving the point. Across this year judges have dealt with a run of cases where lawyers filed citations that no reporter ever published, because a model produced them and nobody checked, and solicitors have been referred to the regulator over it. A more capable model does not lower the standard of care. It raises the quality of the thing you have to check.
What to do this week
Do not upgrade on impulse. If someone wants to move to GPT-5.6, or to the newly released Mythos model, or to any newer version of what you run today, treat it as a decision rather than a click. A model is not a like-for-like swap. A different one can sit on different servers, in a different country, under different terms about whether your prompts are kept or used to train the system. The capability went up. Whether your client data is safe in the new tool is a separate question the upgrade does not answer for you.
Before the switch, write down three things and keep them on file. Write the name of the model and who provides it. Write what its terms say about where your prompts go and whether they train the system. Write the name of the person who signs off that it is fit for client work. If you cannot fill in that middle line, the tool is not ready for a live matter, whatever it can do. Then keep your verification step exactly where it is, and read AI-assisted citations, quotes and figures with the same care you would give a strong junior, not less.
The one instruction worth carrying from the week is short. When the model changes, stop, and put its name, its data terms and its sign-off owner on a single page before a client matter touches it.
The release and the government clearance that preceded it are reported in Nextgov's account of the GPT-5.6 public release, published on 8 July 2026.
If you would like your firm measured against these expectations before a regulator does the measuring, that is what our compliance check does: see how it works.