OpenAI has just announced its highly anticipated GPT-5.6 series of Large Language Models, introducing a trio of celestial-themed models named Sol, Terra, and Luna. But if you were hoping to fire up ChatGPT today and put them through their paces, you are going to be waiting a little longer.
In a move that highlights the mounting tension between tech giants and Washington regulators, OpenAI is launching the new models in a strictly limited preview, heavily vetted by the US government.
The celestial trio broken down
OpenAI is shaking up its naming convention with this release, shifting away from confusing suffix letters to distinct capability tiers.
The flagship of the fleet is GPT-5.6 Sol, a heavy-hitting frontier model built for complex reasoning, extended coding sessions, and advanced cybersecurity research. It is the absolute peak of OpenAI’s current capabilities, designed to tackle problems that previous iterations would simply choke on.
Next in line is GPT-5.6 Terra, positioned as the balanced workhorse designed to handle high-volume enterprise tasks like customer support and document analysis. Terra aims to deliver performance that rivals the older GPT-5.5 flagship but at a much more palatable price point for daily operations.
Bringing up the rear is GPT-5.6 Luna, a fast and highly affordable model tailored for rapid, everyday automation tasks like text summarisation and quick drafting. Luna represents the new entry point for high-volume API applications where speed and cost take priority over raw reasoning power.
GPT-5.6 performance vs the competition
When it comes to raw benchmarks, OpenAI is throwing down the gauntlet to its rivals, particularly Anthropic and Google.
In the Terminal-Bench 2.1 software engineering benchmark, the flagship Sol model achieved a record-breaking score of 91.91% when utilizing its new max reasoning mode. That comfortably edges past Anthropic’s Claude Mythos and Fable 5 models, which have recently faced their own regulatory hurdles overseas.
Even the mid-tier Terra model shows impressive teeth, outpacing older frontier models while keeping token costs significantly lower. For professional workflows evaluated on the rigorous Agent’s Last Exam benchmark, Sol became the first model to clear the halfway mark for autonomous task completion.
The secret sauce this time around appears to be inference-time compute, giving the models more time and structured space to think through complex problems before spitting out an answer.
The price of thinking big
For Aussie developers looking to build on the new architecture, OpenAI has laid out a clear pricing structure per one million tokens.
The top-tier GPT-5.6 Sol will cost US$5.00 for input and US$30.00 for output tokens, matching the pricing structure of the previous generation’s flagship. Shifting down to the mid-tier Terra slashes those costs to US$2.50 for input and US$15.00 for output tokens.
If you are looking to run massive automation loops on a budget, Luna is the absolute bargain of the bunch at just US$1.00 per million input tokens and US$6.00 per million output tokens.
To help keep those API bills from spiralling out of control, OpenAI is also introducing a revamped prompt caching protocol with explicit cache breakpoints and a guaranteed 30-minute minimum cache lifetime.
Why the US Government stepped in
You might wonder why a software update requires government approval, but the capabilities of these new frontier models are raising serious eyebrows in Washington.
The Trump administration has grown increasingly anxious about the potential for advanced AI systems to be weaponised for cyber warfare or critical infrastructure attacks. This caution follows recent drama with Anthropic’s Fable 5 model, which was abruptly pulled from general access following national security concerns and export control directives.
OpenAI CEO Sam Altman admitted to staff in an internal meeting that the government will be approving access customer by customer during this initial phase. Altman noted on social media that while the staggered release was reasonable given the capabilities, it was not the optimal process the company wanted.
When will Aussies get their hands on it
Right now, the preview is locked down tighter than Fort Knox, restricted to a tiny handful of trusted US enterprise partners and government-approved organisations.
The models are completely absent from the standard ChatGPT interface during this testing window, meaning everyday users cannot access them yet. OpenAI states that it plans to expand the rollout globally and bring the GPT-5.6 family to ChatGPT, Codex, and the broader API in the coming weeks.
Australian enterprises and developers will have to keep an eye on their inbox for official invites as the rollout slowly expands beyond US borders.
For more information, head to OpenAI.
