Quick Verdict
On June 26, 2026, OpenAI began a limited preview of GPT-5.6 — its next model generation, split into three tiers: Sol (the flagship), Terra (a balanced model OpenAI says performs competitively with GPT-5.5 at roughly half the cost), and Luna (the fastest, lowest-cost tier). The launch comes with what OpenAI calls its most robust safety stack to date, plus benchmark claims that put Sol ahead of GPT-5.5 on agentic coding workflows and competitive with Anthropic's newest models on cybersecurity tasks while using a fraction of the output tokens.
The catch is access, not capability. At the request of the White House, OpenAI is limiting the preview to "a small group of trusted partners whose participation has been shared with the government" — the first time the U.S. government has preemptively asked an American AI company to restrict a model's launch before it ships publicly. There is no ChatGPT, Codex, or public API access to GPT-5.6 right now for ordinary users or developers, so this review is built entirely from OpenAI's own published specs and system card plus independent reporting, not hands-on testing. We'll revisit with first-hand notes once broader availability actually opens up.
Our early take: 7.2/10 — the specs and safety work look genuinely strong, but the rating is capped by an unusual fact for a review: right now, you simply cannot sign up and use it.
What Is GPT-5.6 Sol?
GPT-5.6 is OpenAI's next model family, unveiled June 26, 2026. Rather than ship one flagship model, OpenAI split the release into three durable capability tiers that can now update on their own schedule: Sol (most capable), Terra (a balanced, everyday-use model OpenAI positions as competitive with GPT-5.5 at about half the price), and Luna (the fastest and cheapest of the three). In this naming system, the version number tracks the generation, while Sol, Terra, and Luna track the capability tier going forward.
Sol is built around three specific workloads: agentic coding, long-horizon biology and genomics analysis, and cybersecurity — vulnerability research, exploit development, and defensive security work. OpenAI introduced two new reasoning settings alongside it: a max reasoning effort that gives Sol more time to work through a single problem, and an ultra mode that coordinates multiple subagents on one complex task rather than relying on a single reasoning pass — a capability that materially increases token usage in exchange for tackling harder, longer jobs.
Why You Can't Use It Yet: The Government-Restricted Rollout
This is the part of the GPT-5.6 story drawing as much coverage as the model itself. OpenAI says GPT-5.6 will eventually reach ChatGPT, Codex, and the API "in the coming weeks," but for now it's available only through the API and Codex to a hand-picked group of partners — and that restriction wasn't OpenAI's idea.
- The White House asked for it. Per TechCrunch, the administration's Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to limit GPT-5.6's rollout to partners "whose participation has been shared with the government," citing security concerns as the administration builds a framework for testing and evaluating frontier models.
- It's the first request of its kind. This marks the first time the U.S. government has preemptively asked an American AI company to restrict a model's launch before it goes out publicly, rather than reacting after release.
- It traces back to a June 2 executive order. President Trump signed an order directing certain AI companies to voluntarily submit their most advanced models for government review up to 30 days before release. Dean Ball, a former White House AI advisor now joining OpenAI, has described the resulting arrangement as a "de facto involuntary licensing regime" for frontier AI.
- Anthropic hit this first, harder. After Anthropic released Claude Fable 5 — its most powerful public model — the administration ordered the company to cut off access for foreign nationals, and Anthropic ended up pulling the model down entirely. A separate Anthropic model, Claude Mythos 5, was also effectively banned this same month; OpenAI's own benchmark comparisons reference "Mythos Preview" as the closest competing model.
- OpenAI is complying, but objecting publicly. "We don't believe this kind of government access process should become the long-term default," OpenAI wrote in its launch post. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." The company frames the restriction as a "short-term step" while it works with the administration on a formal cybersecurity executive-order framework and a "repeatable process for future model releases."
Worth noting: OpenAI also says it deliberately built GPT-5.6's safeguards into the model's core behavior rather than as a separate filter that reroutes flagged prompts to an older, more cautious model — a design choice it frames as an attempt to avoid the false-positive backlash that hit Anthropic's Fable 5, whose classifier-based downrouting reportedly frustrated users on legitimate requests.
GPT-5.6 Pricing: What Sol, Terra, and Luna Cost
Pricing is confirmed directly from OpenAI's launch post and matches TechCrunch's reporting — though note this applies to the preview's API and Codex access for approved partners, since there's no consumer ChatGPT tier pricing yet.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Positioning |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | Flagship — most capable, highest cost |
| GPT-5.6 Terra | $2.50 | $15.00 | Balanced — OpenAI says competitive with GPT-5.5 at roughly 2x lower cost |
| GPT-5.6 Luna | $1.00 | $6.00 | Fastest, lowest cost of the three |
GPT-5.6 also introduces more predictable prompt caching: explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the model's standard input rate, while cache reads keep the usual 90% discount on cached input. Separately, OpenAI says it will bring GPT-5.6 Sol to Cerebras hardware at up to 750 tokens per second starting in July, though that access will also start limited to select customers as capacity expands.
GPT-5.6 Sol vs. GPT-5.5 and Claude: What the Benchmarks Show
OpenAI's launch post describes most results qualitatively rather than publishing a full numbers table, and says an expanded evaluation suite will follow once the model is broadly available. The figures below combine OpenAI's own claims with numbers independently reported by AI-benchmark trackers explainx.ai and Kingy AI, both citing OpenAI's preview materials — treat them as a launch-week snapshot, not a fully independently verified result.
| Benchmark | GPT-5.6 Sol | GPT-5.5 | Claude Fable 5 |
|---|---|---|---|
| Terminal-Bench 2.1 (agentic command-line tasks) | 88.8 (91.9 in "ultra" mode) | 88.0 | 83.4 |
| ExploitBench (cybersecurity / exploit research) | Competitive with Mythos Preview using ~1/3 the output tokens | Not directly compared | — |
| GeneBench v1 (genomics / biology workflows) | Stronger than GPT-5.5, using fewer tokens | Baseline | — |
| Cyber Critical threshold (Preparedness Framework) | Not crossed — found bugs and exploitation primitives in Chromium/Firefox evals, no autonomous full-chain exploit | — | — |
The headline framing from OpenAI is less "biggest jump ever" and more "the same or better results, using meaningfully less compute" — which lines up with the Terra positioning (GPT-5.5-level performance at half the price) more than it suggests an enormous capability leap for Sol over GPT-5.5.
Read OpenAI's full preview announcement and system card
View the GPT-5.6 Sol announcement on OpenAI →Key Features
- Three capability tiers, not one model: Sol, Terra, and Luna let OpenAI ship a flagship, a balanced option, and a budget tier on separate update schedules instead of one undifferentiated release.
maxandultrareasoning modes:maxgives Sol more time to reason through a single problem;ultracoordinates multiple subagents on one complex task, at the cost of substantially higher token usage.- OpenAI's most robust safety stack to date: layered safeguards — refusal training built into the model itself, real-time misuse classifiers during generation, account-level review across flagged conversations, and differentiated access by risk — rather than one single filter.
- 700,000+ A100-equivalent GPU hours of automated red-teaming: OpenAI says it used its own models to search for "universal jailbreaks" that could work across many prompts, not just isolated edge cases, alongside ongoing third-party human red-teaming through the preview period.
- Improved prompt caching: explicit cache breakpoints and a 30-minute minimum cache life make repeated-prompt costs more predictable for agentic workflows that re-send similar context.
- Cerebras hardware deployment coming in July: OpenAI plans to run GPT-5.6 Sol on Cerebras chips at up to 750 tokens per second, starting with select customers as capacity expands.
Pros and Cons
| Pros | Cons |
|---|---|
| Strong, multi-source-confirmed benchmark gains in coding and cybersecurity over GPT-5.5 | No public ChatGPT, Codex, or API access yet — preview is limited to government-vetted partners only |
| Genuinely detailed, transparent safety stack with a published system card | OpenAI itself says the access restriction "shouldn't become the long-term default" — it's openly uncomfortable with the arrangement |
| Terra tier offers GPT-5.5-level performance at roughly half the cost | No independent hands-on testing exists yet from anyone outside the approved partner group |
| Safeguards built into core model behavior, not a separate downrouting filter | Some headline benchmark figures come from third-party trackers rather than OpenAI's own published numbers |
| Cerebras deployment promises very high inference speed once it ships | Pricing is confirmed but the real-world access timeline ("coming weeks") is vague and outside OpenAI's full control |
Who Should Pay Attention to GPT-5.6 Sol Right Now?
Worth tracking closely if you are:
- An enterprise or government-adjacent organization already in conversations with OpenAI about partner access, since you may get preview access before the general public does.
- A developer evaluating GPT-5.5 spend — Terra's pricing alone (roughly half of GPT-5.5 at similar performance, per OpenAI) is worth planning around even before general availability.
- Anyone following AI policy — this is the first time the U.S. government has preemptively restricted a frontier model's release, and how OpenAI and the administration resolve the "repeatable process" they're negotiating will likely shape how every future frontier launch gets handled.
Not worth acting on yet if you:
- Need a model you can actually sign up for and use today — GPT-5.5 and Claude's currently available models remain your real options until GPT-5.6 broadens access.
- Want to evaluate real-world accuracy or reliability rather than vendor-published specs — that requires hands-on testing nobody outside the partner group can currently do.
- Are making a decision based specifically on the cybersecurity capability claims — those are the most safeguarded, least externally verifiable parts of this release.
Frequently Asked Questions
Can I use GPT-5.6 Sol right now?
Not unless you're part of the limited preview group. As of this writing, GPT-5.6 Sol, Terra, and Luna are available only through the API and Codex to a small set of partners whose participation has been shared with the U.S. government. OpenAI says broader availability through ChatGPT, Codex, and the API is coming "in the coming weeks," but hasn't given an exact date.
Why is GPT-5.6's rollout restricted?
At the request of the White House, OpenAI limited the preview to vetted partners while the administration evaluates the model's security implications. It's tied to a June 2026 executive order asking AI companies to voluntarily submit advanced models for government review up to 30 days before release — a process critics have called a "de facto involuntary licensing regime" for frontier AI.
What's the difference between Sol, Terra, and Luna?
They're three capability tiers within the same GPT-5.6 generation. Sol is the flagship and most capable; Terra is a balanced model OpenAI says performs competitively with GPT-5.5 at roughly half the cost; Luna is the fastest and cheapest of the three. Each tier can now be updated on its own schedule going forward.
How much does GPT-5.6 cost?
Per OpenAI's published preview pricing: Sol is $5 per million input tokens and $30 per million output tokens; Terra is $2.50 and $15; and Luna is $1 and $6. These are API and Codex preview rates for approved partners — consumer ChatGPT pricing hasn't been announced.
Did this happen to any other AI company?
Yes. Anthropic's Claude Fable 5 — its most powerful public model at the time — had access cut off for foreign nationals by the same administration shortly after release, and Anthropic ultimately took the model down entirely. A separate Anthropic model, Claude Mythos 5, was also effectively banned this same month.
Final Verdict
On paper, GPT-5.6 Sol looks like a real, if incremental, step up from GPT-5.5 — stronger agentic coding results, a more capable and unusually well-documented cybersecurity safety stack, and a Terra tier that could meaningfully cut costs for teams already paying for GPT-5.5-level performance. The safety engineering OpenAI describes — hundreds of thousands of GPU-hours of automated red-teaming, model-level refusal training instead of a brittle external filter — is the kind of detail that's easy to gesture at and harder to actually publish, and OpenAI published it.
But the honest center of this story isn't the model — it's that almost nobody can use it. A White House-requested restriction, tied to a new executive order on frontier-model review, has turned what should be a straightforward launch into the first real test case of how much control the U.S. government will exert over AI releases before they ship. OpenAI's own discomfort with the arrangement is on the record. Until that gets resolved and access actually opens up, this is a review of a spec sheet and a system card, not a product you can go try.
Early rating: 7.2/10 — strong fundamentals and a serious safety effort, held back from a higher score by the simple fact that, for now, this isn't a tool you can actually get your hands on.