OpenAI has quietly opened a narrow preview of GPT-5.6, limited so far to vetted partners inside Codex and the API. The family, unveiled June 26, now ships as three distinct products: Sol as the top-end model, Terra priced for the middle tier, Luna tuned for the fastest and cheapest runs. None of the three has reached ChatGPT yet.
That structure is the real news. Instead of one model with one price, OpenAI is now selling reasoning as a dial: pick Sol for the hardest problems, Luna for cheap and fast turnaround, Terra for whatever sits between the two. TestingCatalog reported on July 4 that current Codex builds swap the old preset buttons for a single slider governing how much reasoning effort the model applies to a task.
The slider lines up with a confirmed piece of the release: an added “max” toggle so Sol can spend longer working through hard problems, alongside a separate “ultra” mode that calls on subagents for the toughest jobs. A developer could move from a quick Luna pass to an “ultra” Sol run without switching models, only adjusting one control. The design echoes the reasoning selector Anthropic already ships inside its Claude Code desktop app, per TestingCatalog, though OpenAI’s version could still change before it launches.
Not every earlier feature survived. Earlier Codex builds referenced real-time voice; the current app no longer shows it, and OpenAI has not said whether the feature returns or was cut for good.
Naming a release Sol, Terra, and Luna rather than shipping a single version number signals something beyond marketing. Durable tier names let OpenAI advance Terra or Luna on their own schedules without relabeling the whole family every few months, similar to how Anthropic and Google have started separating model speed classes from model generations. It also gives OpenAI three price points to defend instead of one, at the exact moment competitors are cutting prices on their cheapest tiers.
The timing bears directly on Codex users. Anthropic’s Fable 5 returned to global availability on July 1 and will exit bundled subscription tiers by July 7, shifting to usage-based credits. Some agentic-coding teams will be pricing out alternatives in that same week, and a cheaper Luna tier plus a faster Sol option gives OpenAI an obvious pitch to make.
TestingCatalog’s reporting points to the same window for wider access, though that timeline depends on a voluntary security review tied to a recent US cybersecurity order. Approval, not a fixed release date, determines when GPT-5.6 exits preview. OpenAI has not confirmed a public ship date, and the pending review gives it cover if next week comes and goes without one.
Codex users and agent builders should treat GPT-5.6 as a preview, not a committed release, until OpenAI confirms general availability. Anyone paying for bundled Fable 5 access should test Codex’s new reasoning slider against current workflows before July 7, since Anthropic’s pricing shift lands on that date regardless of what OpenAI ships next.
Reported by TestingCatalog on July 4, 2026.
