On 26 June, OpenAI started a limited preview of the GPT-5.6 series: Sol as the flagship, Terra as a balanced model for everyday work, Luna as a fast and affordable variant. The notable part is less the announced capability than the status. This is a preview, not general availability. That distinction is worth a sober look. What the data sheet says is an announcement. What holds in production is decided afterwards.
What is GPT-5.6 Sol?
GPT-5.6 Sol is, by OpenAI’s own account, its strongest model yet, introduced as part of a three-model series. In the new naming, the number marks the generation, while Sol, Terra, and Luna name durable capability tiers that advance on their own cadence. Two modes are new: a max reasoning effort that gives the model more time to think, and an ultra mode that goes beyond a single agent by using subagents to speed up complex work.
For capabilities, OpenAI shows a selected set of evaluations. On coding, Sol sets what it calls a new state of the art on Terminal-Bench 2.1, which tests command-line workflows with planning, iteration, and tool coordination. In biology, it reaches stronger results than GPT-5.5 on GeneBench v1 while using fewer tokens. In cybersecurity, OpenAI calls it its most capable model yet. The company explicitly says an expanded set of results will follow at general availability. Terra is described as competitive with GPT-5.5 while being twice as cheap, and Luna brings strong capability at the lowest cost.
Why a preview is not yet production
During the preview, the models are initially available only through the API and Codex, and only to a select group of trusted partners. OpenAI plans general availability in the coming weeks. It frames the step around the U.S. government: it previewed its plans and the models’ capabilities before launch and, at the government’s request, is starting with a limited preview for partners whose participation has been shared with the government. OpenAI itself writes that this kind of access process should not become the long-term default. For you that means two things: access is narrow, and the figures shown are the vendor’s selection, not the full ledger. Both argue for waiting to see what holds up in your own pipeline.
More cyber capability, more guardrails
Sol launches, according to OpenAI, with its most robust safety stack to date. Under its own Preparedness Framework, the model does not cross the “Cyber Critical” threshold. In tests with Chromium and Firefox it found bugs and exploitation primitives, the building blocks of an exploit, but did not autonomously produce a functional full-chain exploit under the conditions tested. OpenAI relies on several layers: refusals trained into the model, real-time classifiers during generation, account-level review, and differentiated access. For the automated search for universal jailbreaks, OpenAI cites over 700,000 A100-equivalent GPU hours. The company notes that during the preview the safeguards may also block or delay legitimate work. That is part of what the preview is meant to test. If you expand capability, you have to expand control with it.
Reliability lives in the architecture
Already with GPT-5.5 the point held: more autonomy only moves the place where a mistake shows up. An ultra mode that coordinates subagents, and a max mode that computes on its own for a long stretch, make many small decisions that no one is reading along with. That a model checks its own work helps, but it is not the same as a check by someone who is accountable for it. Reliability does not come from the model, but from what you build around it. Define where a human signs off: before the merge, before the deploy, before the migration. That is the core of Agentic Engineering.
Where GPT-5.6 Sol earns its place
The progress is plausible, especially the new modes and the cheaper Terra and Luna tiers. If you are one of the preview partners, hand the model the long, tedious tasks and keep your hand on the points where an unnoticed step costs money, code, or trust. For everyone else, for now: it is a preview, access is narrow, the figures are a selection. Judge it when it runs for you, by what holds up in your production.