GPT-5.6 Sol: OpenAI's Government-Gated Preview (2026)

The most interesting thing about OpenAI's newest flagship model is not a benchmark score — it is that almost nobody can use it. On June 26, 2026, OpenAI previewed GPT‑5.6 Sol, its strongest model yet, but released it only to a small group of trusted partners whose participation had been shared with the U.S. government.¹ No ChatGPT. No open API sign-up. The launch is the clearest sign yet that frontier model releases now run through Washington.

In one line: GPT‑5.6 Sol is OpenAI's new flagship model, previewed on June 26, 2026 alongside two cheaper tiers, Terra and Luna. Unusually, it shipped only to a small group of government-vetted partners through the API and Codex — not ChatGPT — as OpenAI coordinates with the Administration under the new cyber executive order, EO 14409.¹²

TL;DR

• What launched: A limited preview of the GPT‑5.6 family — Sol (flagship), Terra (balanced), and Luna (fast and low-cost) — on June 26, 2026.¹
• Who can use it: During the preview, only "a small group of trusted partners," via the API and Codex. It is not in ChatGPT yet. General availability is promised "in the coming weeks."¹
• Why the gate: OpenAI previewed the models to the U.S. government ahead of launch and, "at their request," started with a restricted release while it works with the Administration on the cyber executive order framework.¹
• The policy behind it: Executive Order 14409, signed June 2, 2026, sets up a voluntary framework giving the government up to 30 days of access to a "covered frontier model" before it reaches other trusted partners.²
• What's new in the model: A max reasoning effort for deeper thinking, and an ultra mode that uses subagents to tackle complex work.¹
• Pricing: Sol is $5 / $30 per million input/output tokens; Terra $2.50 / $15; Luna $1 / $6.¹
• The catch on benchmarks: OpenAI named the tests Sol leads — Terminal‑Bench 2.1, GeneBench v1, ExploitBench — but withheld most scores until the broad release.¹

What You'll Learn

• What GPT‑5.6 Sol is, and how it differs from Terra and Luna
• What the new max reasoning effort and ultra subagent mode actually do
• Which benchmarks OpenAI showed off — and the scores it pointedly did not
• Why you cannot use GPT‑5.6 in ChatGPT yet
• How Executive Order 14409 reshapes how frontier models get released
• How this rollout compares with GPT‑5.5 just two months earlier
• What the gated preview means for developers planning around GPT‑5.6

What is GPT‑5.6 Sol?

GPT‑5.6 Sol is OpenAI's new flagship large language model, introduced in limited preview on June 26, 2026 as the top tier of a three-model family.¹ OpenAI describes it as "our strongest model yet," with improved agentic capabilities in coding, biology, and cybersecurity.¹

The name itself signals a change in how OpenAI brands models. In this new system, the number — 5.6 — identifies the model's generation, while the names Sol, Terra, and Luna identify "durable capability tiers that can advance on their own cadence."¹ In other words, Sol is meant to persist as the high-end tier across future generations, the way "Pro" or "Mini" labels have, rather than being a one-off codename.

The GPT‑5.6 family: Sol, Terra, and Luna

OpenAI shipped three models at once, each aimed at a different point on the intelligence-speed-cost curve.¹

⚠ Prices change frequently. The values above are for illustration only and may be out of date. Always verify current pricing directly with the provider before making cost decisions: Anthropic · OpenAI · Google Gemini · Google Vertex AI · AWS Bedrock · Azure OpenAI · Mistral · Cohere · Together AI · DeepSeek · Groq · Fireworks AI · Perplexity · xAI · Cursor · GitHub Copilot · Windsurf.

Sol is positioned for the most demanding work, such as complex coding and security research. Terra is the everyday workhorse: OpenAI says it has "competitive performance to GPT‑5.5 while being 2x cheaper," which makes it the natural default for large-scale production. Luna brings "strong capability at our lowest cost" for routine, latency-sensitive tasks.¹

GPT‑5.6 also reworks prompt caching. It adds explicit cache breakpoints and a 30-minute minimum cache life, and changes the billing math: for GPT‑5.6 and later models, cache writes are billed at 1.25x the uncached input rate, while cache reads keep the usual 90% discount.¹ Separately, OpenAI said it will launch Sol on Cerebras hardware "at up to 750 tokens per second" in July, initially for select customers.¹

What's actually new: max reasoning and ultra mode

Two features stand out beyond raw model quality.¹

The first is a new max reasoning effort setting, which gives Sol "the most time to reason deeply." This extends the now-familiar pattern of trading latency and token spend for harder thinking on difficult problems.

The second is more novel: a new ultra mode that "goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work."¹ That is a meaningful shift. Instead of one model working a problem end to end, ultra orchestrates multiple subagents in parallel — a pattern that power users have been building by hand with agent frameworks, now offered as a first-class mode. It is the clearest signal yet that OpenAI sees multi-agent orchestration, not just bigger single models, as the path to harder tasks.

The benchmarks OpenAI showed — and the scores it didn't

Here is where the preview gets unusual. OpenAI named the benchmarks where Sol leads but, in most cases, withheld the actual numbers, saying it "will share an expanded suite of evaluation results when we make the model broadly available."¹ That is the opposite of a normal launch, where the scores are the announcement.

What OpenAI did claim, qualitatively:

• Terminal‑Bench 2.1 (coding): Sol "sets a new state of the art." Terminal‑Bench is a hard agentic-coding benchmark built around real command-line tasks in Docker environments, verified by the final state of the container rather than the commands typed.¹³ On its public 2.1 leaderboard the top score sits near 88%, so a genuine state of the art is a real bar to clear — but OpenAI didn't publish Sol's number.³
• GeneBench v1 (biology): Sol "achieves stronger results than GPT‑5.5 while using fewer tokens" on long-horizon genomics and quantitative-biology analysis.¹
• ExploitBench (cybersecurity): Sol is "competitive with Mythos Preview using only ~1/3 of the output tokens," per OpenAI's own evaluation — a notable efficiency claim against a rival system that has matched OpenAI on cyber evaluations before.¹
• ExploitGym: a benchmark from UC Berkeley researchers built with OpenAI and other labs, on which Sol, Terra, and Luna all improve as they are given more reasoning.¹

Treat all of these as vendor claims until the full evaluation suite lands. The honest reading is that OpenAI wanted to convey a capability jump — especially in cyber — without publishing a leaderboard that competitors could immediately target.

Why you can't use GPT‑5.6 yet: the government-gated preview

The headline restriction is the story. During the preview, GPT‑5.6 is available only "through the API and Codex to a select group of trusted partners and organizations." It is not in consumer ChatGPT, and access is not open to the public.¹ (Secondary reporting put the partner count at roughly 20 organizations, though OpenAI itself only said "a small group.")⁴

OpenAI was unusually direct about why. It "previewed our plans and the models' capabilities" to the U.S. government ahead of the launch, and "at their request," began with a limited release to partners "whose participation has been shared with the government."¹ OpenAI even editorialized against the arrangement: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."¹

The reason a model gets this treatment is cyber capability. OpenAI says Sol is its "most capable model yet for cybersecurity," shifting the frontier on long-horizon tasks like vulnerability research. Crucially, it adds that Sol does not cross the "Cyber Critical" threshold under OpenAI's Preparedness Framework: in tests against Chromium and Firefox, the model found bugs and exploitation primitives but did not autonomously produce a working full-chain exploit under the conditions tested.¹⁵ To harden it, OpenAI says it spent over 700,000 A100-equivalent GPU hours on automated red-teaming aimed at universal jailbreaks, layered on top of human expert testing.¹

Executive Order 14409, explained

The legal backdrop is Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security," signed June 2, 2026.² Its Section 3, "Secure Frontier Model Deployment," directs the Treasury, the Department of War (through the Director of the NSA), and Homeland Security (through CISA) to do two things within 60 days:²

1. Build a classified benchmarking process to set the threshold at which a model is designated a "covered frontier model" — a determination made by the Director of the NSA.
2. Design a voluntary framework through which developers can ask the government whether a model qualifies, give the government access to a covered frontier model for up to 30 days before releasing it to other trusted partners, and collaborate on which trusted partners get early access.

Two details matter. First, the framework is explicitly voluntary: the order states that nothing in it authorizes "a mandatory governmental licensing, preclearance, or permitting requirement" for releasing AI models.² Second, the mechanism — government access first, then a hand-picked set of trusted partners, then everyone else — is almost exactly the shape of the GPT‑5.6 rollout. OpenAI's own framing, that it is working with the Administration "to develop the cyber Executive Order framework and a repeatable process for future model releases," ties the two together directly.¹

That makes GPT‑5.6 Sol, in practice, what appears to be the first major commercial model to ship through this kind of government-coordinated pre-release process — even though the formal EO 14409 framework was still being designed when Sol launched. How "voluntary" this stays in practice, when the most capable models are also the most cyber-relevant, is the open question hanging over the next year of releases. OpenAI's GPT‑5.5‑Cyber work and its reception in the EU already hinted that cyber capability and regulation were on a collision course.

GPT‑5.6 vs GPT‑5.5: how the rollout changed

Two months earlier, GPT‑5.5 launched under a much lighter touch. The contrast shows how fast the release playbook is changing.

GPT‑5.5 was itself somewhat cautious — OpenAI briefly withheld API access at launch, citing "different safeguards," then added it a day later, and made GPT‑5.5 Instant the free-tier default about two weeks after.⁶ But that gating measured in days. GPT‑5.6's gate has no public end date beyond "coming weeks," and entry is by government-shared invitation rather than a subscription. That is a different category of restriction.

What the gated preview means for developers

For now, the practical takeaways are modest but real:

• You probably can't build on Sol today. Unless your organization is among the trusted partners, GPT‑5.6 is not callable. Plan production work around GPT‑5.5 (or Terra/Luna once they broaden) and treat Sol as a near-future upgrade, not a current option.
• Terra is the number to watch. Its "competitive performance to GPT‑5.5 while being 2x cheaper" claim is the one most likely to change real budgets once the family is broadly available.¹ If it holds, a lot of GPT‑5.5 workloads get cheaper by swapping models.
• ultra mode changes how you'd architect agents. If OpenAI ships first-class subagent orchestration, some custom multi-agent scaffolding becomes redundant. Worth prototyping against once access opens.
• Re-check your caching math. The new 1.25x cache-write surcharge and 30-minute cache life change the economics of prompt caching for high-volume apps.¹
• Expect the gate to become normal. EO 14409 makes government pre-review a repeatable path. If you ship in regulated or security-sensitive markets, the release timeline of your underlying model may now depend partly on Washington.²

The Bottom Line

GPT‑5.6 Sol is a genuine capability step — a new flagship with deeper reasoning, subagent orchestration, and a cheaper Terra tier that could reset production costs. But the lasting headline is the release model, not the model. In an apparent first, OpenAI shipped a frontier system to the government first and the public last, under a voluntary cyber framework it is helping write in real time. Whether that becomes a one-off or the new normal is the thing to watch — and OpenAI, for its part, says it hopes it's the former.