OpenAI rolled out three new models across the entire ChatGPT platform and API on July 9, 2026, ending a 12-day period where only government-vetted organizations had access. The move marks a sharp pivot in how frontier AI labs now operate.

GPT-5.6 Sol, Terra, and Luna represent three different performance tiers. Sol targets frontier work and agentic tasks. Terra balances capability with cost for everyday use. Luna is fast and cheap. The U.S. Department of Commerce gave its approval after additional testing and meetings with government agencies, removing the restriction that had limited access to roughly 20 vetted entities.

Three Models, Three Purposes

Sol is the powerhouse. It’s built for reasoning over long contexts and handling complex, multi-step agentic work. The tradeoff is speed and cost.

Terra sits in the middle. OpenAI positions it as GPT-5.5-competitive performance at 2x lower cost. For most teams running production systems, Terra likely becomes the default choice.

Luna is the speed play. Fast inference on Cerebras hardware. Useful when latency matters more than raw reasoning power.

A New Model of Government-AI Relations

The 12-day window was not a mistake or a delay. It was deliberate. OpenAI, Anthropic, and xAI all agreed to a voluntary framework with the White House Office of the National Cyber Director. New frontier models get tested by vetted government orgs first. Then, if approved, they go public.

This is new territory. No U.S. tech company had done this before. It’s also unclear how long it will last. But for now, it’s the protocol.

Pricing and Token Economics

Sol costs $5 per 1M input tokens, $30 per 1M output tokens. Terra is $2.50 in, $15 out. Luna is $1 in, $6 out. For context, GPT-4 Turbo was $10 and $30. GPT-5.6 undercuts it across the board, which is how frontier models typically work once they ship.

The real cost question is total token consumption. Cheaper per-token pricing often pushes users to run longer prompts and bigger batches. We’ll need to see actual usage patterns to know whether this is cheaper in practice.

For the first time since export control restrictions began in June, every major frontier AI lab has a publicly available model at the same time.