OpenAI Rolls Out GPT-5.6 Sol, Terra, Luna in Limited Preview

GPT-5.6 family launches under a controlled preview

OpenAI announced a limited preview of its new GPT-5.6 series, comprising Sol, Terra and Luna models, available to roughly 20 vetted organizations after coordination with the U.S. government. Sol tackles heavyweight tasks such as advanced coding and security research, while Terra targets high‑volume business workflows, and Luna delivers rapid, low‑cost assistance for routine summarisation.

Pricing tiers and use‑case focus

Sol is priced at $5.00 per million input tokens and $30.00 per million output tokens, matching GPT‑5.5 rates but promising faster long‑run coding runs. Terra costs $2.50/$15 per million tokens, and Luna sits at $1/$6 per million tokens, making it the most economical option.

“We are starting with a trusted partner pool at the request of the U.S. government,” OpenAI stated in its blog post.

Performance metrics show Sol achieving a ↑ 91.91% score on TerminalBench 2.1, eclipsing GPT‑5.5’s 83.4% and Claude Mythos 5’s 88%. Luna, despite its lower cost, reaches near‑GPT‑5.5 levels on standard benchmarks.

Safety architecture and compliance

All three models carry a “High” risk classification for cyber and bio‑chemical capabilities, triggering real‑time misuse screening and activation‑based checks for Sol and Terra. Luna receives baseline monitoring but lacks the activation classifier layer. OpenAI reports a 94.8% recall on biology tests and 81.6% on cybersecurity assessments, acknowledging occasional false positives.

Enterprise implications

Customers must adapt to token‑caching protocols that charge 1.25× for initial writes with a 90% discount on reads, guaranteeing a 30‑minute cache life. Sol will also run on Cerebras hardware from July, promising up to 750 tokens per second for latency‑critical applications.

Regulatory backdrop

The rollout aligns with a June 2 executive order mandating federal oversight of frontier AI, a process slated for completion by July 2. OpenAI’s public stance warns against permanent government gate‑keeping, arguing it hampers broader innovation. Bloomberg notes the move mirrors recent export controls placed on Anthropic’s Claude models after jailbreak discoveries.

Intel provided by Dr. Aris Thorne (Artificial Intelligence Researcher).