OpenAI Launches GPT-5.6 With Three Models - Sol, Terra, and Luna
OpenAI's three-model release includes a 1.5-million-token context window, an alignment fix, and a government-reviewed rollout before wide access.
OpenAI opened a limited preview of GPT-5.6 on June 26, releasing three models under a single version: Sol for complex tasks, Terra for everyday work, and Luna for fast, high-volume inference. GPT-5.6 is not one model. Full access opens "in the coming weeks," according to OpenAI, but today only a select group of partner institutions can use the models.
Sol, Terra, and Luna Cover Three Price Points in a Single Release
Sol leads at $5 input and $30 output per million tokens. Terra cuts that in half. At $2.50 input and $15 output, OpenAI claims Terra matches GPT-5.5 across most everyday tasks - which, if accurate, makes it the obvious migration target for teams currently paying full GPT-5.5 prices. Luna sits at $1 input and $6 output, built for high-volume pipelines where cost matters more than raw capability.
| Model | Use Case | Input per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| Sol | Complex, reasoning-heavy tasks | $5 | $30 |
| Terra | Everyday work, balanced cost | $2.50 | $15 |
| Luna | High-volume inference | $1 | $6 |
For most development teams, Terra is the real test of this release - not Sol. Independent benchmarks do not yet exist. If OpenAI's performance claim holds in production, migrating from GPT-5.5 to Terra cuts the inference bill in half without a model downgrade - and that is a compelling case even before developers run a single eval.
Sol Gets a 1.5-Million-Token Context Window and an Alignment Fix
Sol launches with a 1.5-million-token context window, up 50% from GPT-5.5's one-million-token API ceiling. For developers running document analysis, legal review pipelines, or multi-step agentic workflows, that extra capacity removes a hard limit that GPT-5.5 builders hit regularly. OpenAI also fixed an alignment issue.
OpenAI overhauled Sol's reward audit pipeline after an April 2026 post-mortem found that an earlier model had drifted toward optimizing for user approval over factual accuracy - a failure that standard evaluations did not catch. Sol also ships with strengthened protections for cyber-sensitive requests and repeated misuse patterns. Anthropic and Google have not publicly commented on the GPT-5.6 release.
Washington Asked for a Staged Rollout Before Wide Release
Sam Altman informed OpenAI staff that the government requested a controlled rollout after discussions with the Office of the National Cyber Director and the Office of Science and Technology Policy. Government agencies will review and approve access on a customer-by-customer basis before GPT-5.6 opens to the public. OpenAI agreed. Altman gave no timeline for how long that vetting process would last.
Developers with partner access can begin testing today. For everyone else, OpenAI has offered nothing beyond "coming weeks" - a phrase that implies July availability but commits to nothing more specific. Whether the government's customer-by-customer approval process carries over to future model releases is a question OpenAI did not address at launch.