OpenAI launches three new AI models to the general public on July 9: Sol, Terra, and Luna. This is the company's most powerful generation yet, and the first to go through a formal, government-supervised pre-release review before reaching ordinary users.
The three models cover different price points and use cases. Sol is OpenAI's strongest model to date, built for complex tasks like code writing, cybersecurity work, and long multi-step jobs. Terra is a mid-range option that matches the performance of last generation's flagship, GPT-5.5, at roughly half the cost. Luna is the cheapest and fastest, designed for high-volume, budget-sensitive work.
On pricing: Sol costs $5 per million words processed as input and $30 per million words generated as output. Terra runs at $2.50 input and $15 output. Luna is $1 input and $6 output. For context, a million words is roughly the length of eight to ten average business novels, so in real usage the costs are far smaller than those numbers suggest. What matters more is the direction: Terra delivers near-flagship quality at half the flagship price, continuing a pattern of AI getting cheaper every cycle.
The delay before this public launch came from a White House order signed June 2, 2026. That order asks AI companies to voluntarily submit their most powerful new models to government testers for up to 30 days before releasing them to the public. The Department of Commerce's Center for AI Standards and Innovation ran additional tests on GPT-5.6, and OpenAI sent technical staff to Washington to answer questions. The government cleared the models for public launch in less than 30 days.
OpenAI was open about its discomfort with the process. The company said the review "keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them," but agreed to participate because it was the fastest path to a broader release. That tension is real and worth watching: OpenAI cooperated, got cleared quickly, and kept its complaint on the record.
Anthropic had a harder time. Its Claude Fable 5 and Mythos 5 models were shut down globally for 19 days in June after Amazon researchers found a security flaw that allowed the model to bypass its own safety rules. The U.S. government issued an emergency export-control order, and because Anthropic had no reliable way to check the nationality of every user in real time, it cut off all access for everyone, including its own non-citizen staff. Enterprise users scrambled back to older models. Fable 5 was restored globally on July 1; the more powerful Mythos 5 remains limited to a small group of approved U.S. organizations.
These two episodes together mark a turning point. Government agencies can now delay or block the most powerful AI tools, sometimes on short notice, and the legal tools to do it already exist. The June 2 executive order created a voluntary path, but Anthropic's shutdown shows the government will use harder instruments, like export controls, when it wants to move fast.
For businesses that rely on AI tools in their day-to-day work, the lesson from June is concrete: if you have built a process that depends on a specific AI model, that model can disappear for days or weeks without warning. The organizations that fared best during Anthropic's blackout were the ones that had not locked into a single model or provider.
The GPT-5.6 launch itself is less controversial than the political context around it. The three-tier naming system, Sol, Terra, and Luna, is OpenAI trying to make its product lineup easier to read. Instead of cryptic mini and nano labels, each name now signals a permanent capability level. Sol is the top tier, Terra the middle, Luna the budget option, and that structure is meant to hold across future generations.