GPT-5.6 Sol, Terra, Luna: A Builder's Read

OpenAI shipped a cheaper frontier on Friday afternoon, and you cannot get it. The preview of GPT-5.6 landed on June 26 as three models at once: Sol, the flagship; Terra, a balanced tier for high-volume work; and Luna, a fast and cheap everyday model. Sol costs about half of Anthropic's Claude Fable 5, posts a new state of the art on agentic coding, and adds a mode that fans work out to subagents. On OpenAI's own benchmarks it is the strongest model the company has shipped.

And for now it is locked behind a government access gate. The launch came less than a day after the news that the White House had asked OpenAI to stagger the release, and the preview is open only to a small group of trusted partners whose names were shared with the administration first.¹1The Verge, June 26 2026: OpenAI unveiled the preview "less than 24 hours after news broke that OpenAI would stagger its next model release at the request of the Trump administration." Customers are being approved case by case during the preview. That gap, between a model that is cheaper and better and a model you can actually deploy, is what should shape your planning, not the benchmark scores.

What actually shipped

Three tiers, one generation. OpenAI also changed how it names models: the number marks the generation, and Sol, Terra, and Luna are durable capability tiers that can each advance on their own cadence. Sol is the deepest, Terra trades a little capability for half the cost, and Luna is the cheap workhorse. OpenAI says Terra has competitive performance to GPT-5.5 at 2x cheaper, and Luna brings real capability at the lowest cost in the lineup.

Two new knobs come with Sol. A max reasoning effort gives it the most time to think on a hard problem. An ultra mode goes past a single agent and fans work out to subagents to parallelize complex jobs.²2The ultra mode reads as a direct nod to the subagent-orchestration pattern, and The Verge connected it to OpenClaw creator Peter Steinberger's work at OpenAI. If you have built a fan-out loop by hand, this is that pattern moving into the model's own serving layer. If you have been wiring up your own orchestration loops, the frontier labs are now shipping that shape as a first-class mode.

The pricing knife

The numbers move budgets, so start there. Per million tokens: Sol is $5 input and $30 output, Terra is $2.50 and $15, Luna is $1 and $6. Sol comes in at roughly half of Claude Fable 5's $10 and $50.³3Fable 5 sits at $10 input / $50 output per million tokens. Sol at $5 / $30 is not a rounding-error discount, it is a structural undercut on the flagship tier, with Terra and Luna opening up volume tiers underneath it. The caching scheme changed too, which reworks the effective input bill for any agent that re-sends a long system prompt every turn.side quest

For a high-volume agentic system, the tiering is the lever, not the headline flagship price. Route the easy 80 percent of calls to Luna or Terra and reserve Sol for the hard tail, and the blended cost drops well below an all-flagship stack. The question is always where the crossover sits for your traffic. Move the sliders below to see it.

The model that sets your bill is rarely the smartest one. It is the cheapest one that clears your quality bar on each class of query, with a frontier escalation for the cases that need it.

What it beats

OpenAI led with agentic capability across coding, biology, and cybersecurity. Sol sets a new state of the art on Terminal-Bench 2.1, which tests command-line workflows that need planning, iteration, and tool coordination, the closest public proxy for "can it actually drive a shell through a multi-step job." On GeneBench v1, a long-horizon genomics and quantitative-biology benchmark, it beats GPT-5.5 while using fewer tokens.⁴4Fewer tokens for a better score is the metric to watch on agentic work. On a long-horizon task, token efficiency compounds: it is lower cost and lower latency on the same job, not just a higher number on a chart.

The cyber numbers are the ones Washington cares about, and OpenAI clearly knew it. On ExploitBench, Sol is competitive with the Mythos Preview using only about a third of the output tokens. On ExploitGym, a benchmark built by UC Berkeley researchers with OpenAI and other labs, all three models improve sharply on cyber tasks as you raise reasoning effort.side quest

For a builder the practical signal under all of this is the defensive one. The safeguards are tuned to preserve code review, vulnerability research, patch development, and security education while making prohibited offensive work harder. OpenAI is explicit that during the preview those safeguards will sometimes block legitimate dual-use work, because telling defense and offense apart on a half-finished request is genuinely hard. That is part of what the preview is testing.

The catch

GPT-5.6 is not generally available, and that reorders the planning. During the preview it runs through the API and Codex to a select group of trusted partners and organizations, with general availability promised "in the coming weeks."⁵5OpenAI is also bringing Sol to Cerebras at up to 750 tokens per second in July, initially for select customers. Fast, and also gated. OpenAI previewed the models and their capabilities to the US government before launch, and at the government's request started with the limited, name-shared partner list.

OpenAI did not hide its discomfort. From the announcement: it does not believe "this kind of government access process should become the long-term default," because it "keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." The company framed the gate as a short-term step toward broad availability while it works with the administration on a cyber Executive Order framework and a repeatable process for future releases.side quest

The builder's read

Strip away the benchmark charts and the governance drama and you are left with one operational fact: the frontier is getting cheaper and harder to reach at the same time. Sol undercuts Fable 5 on price while sitting behind an access gate Fable 5 does not have. Terra and Luna open up genuinely cheap volume tiers that you cannot yet call.

The hedge against that is not a better vendor contract. It is owning enough of your own inference that a takedown, an export rule, or a partner-list decision in Washington cannot stall your roadmap. You route the easy work and the sensitive work to hardware you control, and you reserve the gated frontier for the genuine long tail, when you can get it. That is not a hypothetical posture. It is exactly the lab I am building in the open right now.

The labs will sort out the Executive Order framework and GPT-5.6 will reach general availability, probably soon. But the lesson of the last two weeks is durable. Frontier access is now a policy variable, not just a billing one. The teams that keep shipping through the next gate will be the ones who treated their own iron as part of the stack, not as a fallback.

Related reading on this site: From a DGX Spark to a Borrowed Node for why a small box on a desk de-risks a big borrowed one, Day One: Standing Up the Inference Platform for serving a reasoning model locally at speed when the frontier is not an option, and Waking the Research Engine for running real autonomous work on hardware you control.