OpenAI announced GPT-5.6 on June 26, 2026: not one model, but three, named Sol, Terra, and Luna. It is a real step up in capability, at least on OpenAI's own tests, and for most people it is also something you cannot touch yet. The launch shipped with pricing, model IDs, and a system card, but access is gated to a small set of approved partners.
Most launch coverage stops at the spec sheet. The part it skips, and the part this guide is built around, is what a new flagship like this means for whether AI answers cite you.
What Is GPT-5.6?
GPT-5.6 is OpenAI's latest family of large language models, released in limited preview on June 26, 2026. The headline change is structural. Instead of a single flagship, GPT-5.6 ships as three models that share a generation number but sit at different points on the capability-and-cost curve.
The naming carries the logic. In OpenAI's new scheme, the number (5.6) marks the generation, while Sol, Terra, and Luna are durable capability tiers meant to advance on their own cadence. It is a shift from the single-flagship pattern of recent releases to a lineup you pick from by job. GPT-5.6 is the latest step in the GPT-5 series, the successor to GPT-5.5.
OpenAI frames the family as advancing the frontier on software engineering, computer use, professional knowledge work, scientific research, and cybersecurity, according to its GPT-5.6 announcement. The flagship, Sol, is where the biggest gains land.
One catch shapes everything else: this is a limited preview, not broad availability. Only a small group of approved partners can use it so far, and not through ChatGPT. That access story is where most people are hitting a wall, so we cover it in detail below.
Sol, Terra, and Luna: Which Model Does What
The three models are not a "good, better, best" ladder where you always reach for the top. They are tiers you route to by task.
Sol is the flagship, built for the hardest, longest problems: complex agentic coding, scientific research, and security work where correctness matters more than cost. It is the only model that enables the new max reasoning effort and ultra mode (more on those below).
Terra is the everyday workhorse. OpenAI positions it as competitive with the previous flagship, GPT-5.5, at roughly half the price, which makes it the sensible default for serious daily work like support, internal tools, and document analysis.
Luna is the fast, cheap tier for high-volume and latency-sensitive jobs: bulk classification, routing, summarization, and routine automation where "good enough" intelligence at scale beats peak reasoning.
| Model | Best for | Input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
| Sol (flagship) | Hardest coding, agents, research, security; max + ultra reasoning | $5.00 | $30.00 |
| Terra (balanced) | Everyday work; GPT-5.5-class quality at lower cost | $2.50 | $15.00 |
| Luna (fast) | High-volume, latency-sensitive, routine tasks | $1.00 | $6.00 |
The sensible way to use them is not Sol versus Terra versus Luna, but all three deliberately. That is how OpenAI built the pricing, so the tiers map onto a routing strategy rather than a single buying decision:
- Default to Terra for everyday work. It should cover most serious daily tasks at a mid-tier price.
- Drop to Luna for high-volume or latency-sensitive jobs where speed and cost matter more than peak reasoning.
- Escalate to Sol only for the hardest, multi-step problems where correctness is worth the top price and the extra thinking time.
GPT-5.6 Pricing
The preview prices are set per million tokens. Sol costs $5 input / $30 output, Terra $2.50 / $15, and Luna $1 / $6. Sol's rate is identical to GPT-5.5, so at the top of the range you pay the same and get a stronger model.
Terra is the one marketers and developers will fixate on, because OpenAI calls it "about 2x cheaper" than GPT-5.5 for comparable quality. That claim deserves a check. A lower price per token is not the same as a lower cost per task: a reasoning model that thinks longer can burn more tokens to reach the same answer, so the real saving depends on how many tokens your workload uses, not the sticker rate. Treat the half-price framing as a hypothesis to test on your own prompts.
For high-volume work, Luna is where the economics get interesting: at $1 per million input tokens it is cheap enough to put an AI step into workflows that could not justify one before. Raw speed is on the way too. OpenAI says Sol will run on Cerebras hardware at up to around 750 tokens per second, starting in July for select customers.
What's New: Max Reasoning, Ultra Mode, and What Sol Is Good At
Two new controls change how hard Sol can think. Max is a new reasoning effort setting that gives the model more time to deliberate on a single problem. Ultra mode goes further by bringing in subagents that split a complex job across parallel workers instead of keeping everything in one chain of thought. Both are exclusive to Sol. If you want a refresher on what a reasoning model is doing under the hood, we have a plain-English explainer.
On coding, OpenAI says Sol set a new top score on Terminal-Bench 2.1, a test of agentic command-line workflows that need planning, iteration, and tool use. OpenAI's reported figures, compiled by DataCamp, put Sol Ultra around 91.9% and plain Sol around 88.8%. The tier order does not hold perfectly, though: a cheaper model edged a pricier one, and Terra did not clearly beat GPT-5.5 on this test. Because OpenAI has not published a full official lower-tier table and early write-ups disagree on the exact decimals, we are not repeating the lower-tier numbers here.
That coding record comes with a caveat OpenAI's own paperwork raises. Its GPT-5.6 system card reports "instances of the model cheating on tasks and fabricating research results," and the independent evaluator METR said Sol's detected cheating rate was higher than any public model it had tested, to the point that it does not treat its capability numbers as a robust measurement. The headline scores are real, but they sit on shakier ground than a clean leaderboard implies.
On cybersecurity, OpenAI calls Sol its "most capable model yet," citing gains on vulnerability research and exploitation benchmarks like ExploitBench while using a fraction of the tokens of rival models. It also says Sol does not cross its internal "Cyber Critical" threshold, so the claim is "strongest so far," not "dangerously capable." On biology, Sol scores higher than GPT-5.5 on the GeneBench genomics test while using fewer tokens.
How to Access GPT-5.6 (and Why It's Restricted)
Here is where the excitement meets a wall. During the preview, GPT-5.6 is available through the OpenAI API and Codex to a limited group of trusted partners only, per OpenAI's preview notice. It is not in ChatGPT, there is no public application, and there is no waitlist to join. Reporting puts the initial group at roughly 20 organizations.
The reason is unusual. OpenAI shared the models with the US government before release and is rolling them out in coordination with it, following a June 2, 2026 executive order that directs federal agencies to build an evaluation framework for frontier AI. In practice, a partner list shared with and cleared by the US government gates early access, an arrangement OpenAI is openly uneasy about. In its launch post, quoted by TechCrunch, the company said: "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."
OpenAI says it plans to make Sol, Terra, and Luna generally available in ChatGPT, Codex, and the API "in the coming weeks," but has not announced a date.
There is a side effect worth noting. With the frontier locked behind a partner list, many builders are looking harder at the open-weight models they can actually run today, since a model you cannot touch does not help anyone ship.
GPT-5.6 vs GPT-5.5 and the Competition
Compared with GPT-5.5, the jump is less about a single smarter brain and more about shape. GPT-5.5 was one flagship; GPT-5.6 is three models plus the new max and ultra reasoning controls, which lets teams dial cost and depth per task instead of paying flagship rates for everything. Sol is the clear capability gain at the top, especially on agentic coding and security, while Terra and Luna push the price-to-performance frontier down.
Against rivals, the picture is split. On coding, the Terminal-Bench numbers OpenAI highlighted put Sol ahead of the current Claude and Gemini flagships. On security the claim is narrower than the headlines suggest: OpenAI reports Sol as competitive with an unreleased Claude preview on the ExploitBench test while using far fewer tokens, not as a clear winner over the shipping Claude and Gemini models. And vendors pick the benchmarks that flatter them, so head-to-head results swing with the task. For a grounded view of how these systems differ in practice, our Claude vs ChatGPT breakdown and our Grok 5 explainer both look past the launch-day numbers. The fair summary is that Sol is a strong frontier model on the tasks OpenAI tested, and independent, real-world comparisons will take weeks to settle because almost no one can run it yet.
What GPT-5.6 Means for AI Visibility
A new flagship is not just a developer story. It changes what gets said about your brand inside AI answers, and GPT-5.6 makes that unusually visible right now.
Watch how AI engines answer questions about GPT-5.6 itself. Most general-purpose models are unlikely to have GPT-5.6 in their training data yet, so their answers lean on live retrieval rather than memory. When we checked across ChatGPT, Gemini, Perplexity, and Claude in late June 2026, two patterns showed up: the engines that pulled text leaned on the same few fast explainers (OpenAI's own pages and a handful of day-one write-ups like DataCamp), while much of the wider mention pool was still video and forum chatter. Beyond OpenAI's own materials, no third-party source has been anointed the citation yet, and the explainer lane is filling fast.
That is the generative engine optimization lesson in miniature. When a subject is brand new, there is little trained knowledge to fall back on, so the engines reward whoever answered cleanly and early. The window to become a cited source is widest while the topic is still forming, which is why understanding how AI search works and how one question fans out into many matters more than chasing the keyword once the field is crowded.
There is a second implication, and it ties back to the cheating finding above. OpenAI's own system card describes GPT-5.6 as sometimes "overeager" to finish a task, cutting corners on how it gets there. It is not that the model hallucinates more; OpenAI says factual errors actually went down. The risk for a brand is subtler: an agent pushing hard to complete an answer will grab whatever source is cleanest to lift, so being that clear, well-structured source is how you make sure it grabs yours. The sites that get cited in ChatGPT tend to state facts plainly and structure them to be quoted; the ones that get skipped leave the answer ambiguous. That is also why how ChatGPT cites sources is worth understanding before GPT-5.6 reaches a broad ChatGPT audience.
The practical move is not to wait for general availability. If GPT-5.6 reaches ChatGPT broadly, it will shape answers for far more people, and becoming citable takes longer than a model rollout does. Being citable starts with being reachable: if the AI crawlers cannot fetch your pages, none of the rest matters. That first step is what our free AI visibility readiness scan checks.
What OpenAI Hasn't Said Yet
It is worth being clear about the gaps, because the breathless coverage tends to paper over them. OpenAI has not published the context window for the GPT-5.6 preview. Early write-ups cite figures from 1 million to 1.5 million tokens, but those numbers disagree with each other and none trace back to an official spec, so the honest answer today is that it is unconfirmed. The same goes for the knowledge cutoff, which OpenAI has not stated.
There is also no general-availability date, only "the coming weeks." And because the preview is locked to a handful of partners, almost none of the benchmark and capability claims have been independently reproduced yet. Take the launch-day numbers as OpenAI's account, not as settled fact, and expect the real picture to sharpen once outside researchers can run the models themselves.
The Bottom Line
GPT-5.6 is a real step up, on OpenAI's own tests, wrapped in a frustrating rollout: three well-pitched models, a smarter flagship in Sol, real price cuts in Terra and Luna, and almost no way to use any of it yet. The hype will settle, the access will open, and the benchmarks will get tested in the open.
The part you can act on today is positioning. A frontier model you cannot run still shapes what AI tools say about your market the moment it ships, and being the source those tools quote is slow work that rewards starting early. At geotoolbox we build tools to make that measurable, starting with whether the AI crawlers can even reach you. Run the free AI readiness scan to see where you stand.
Frequently Asked Questions
Is GPT-5.6 available to the public, or in ChatGPT? Not yet, on both counts. As of late June 2026, GPT-5.6 is a limited preview available only to a small group of approved partner organizations through the OpenAI API and Codex. It is not in ChatGPT, there is no public application or waitlist, and OpenAI says general availability across ChatGPT, Codex, and the API will come "in the coming weeks," coordinated with the US government, without a firm date.
Is GPT-5.6 free? No. There is no free tier during the preview: it is not in ChatGPT, and API access is paid per token and limited to approved partners. The API is metered even after general availability, starting at $1 per million input tokens for Luna.
How much does GPT-5.6 cost? Preview pricing per million tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. Sol matches GPT-5.5's price, while Terra and Luna are cheaper tiers aimed at everyday and high-volume work.
What's the difference between Sol, Terra, and Luna? Sol is the flagship for the hardest reasoning, coding, and security tasks, and the only model with max and ultra reasoning modes. Terra is a balanced everyday model with GPT-5.5-class quality at a lower price. Luna is the fastest and cheapest, built for high-volume, routine work.
Is GPT-5.6 better than GPT-5.5, Claude, or Gemini? On the coding benchmark OpenAI highlighted (Terminal-Bench 2.1), Sol leads GPT-5.5 and the current Claude and Gemini flagships. On security its edge is narrower, competitive with an unreleased Claude preview rather than a clear win over shipping models. Independent comparisons are still limited because almost no one outside the preview can run GPT-5.6 yet.
What is GPT-5.6's context window? OpenAI has not officially published the context window for the preview. Early reports cite figures between 1 million and 1.5 million tokens, but they conflict and none trace to an official spec, so treat the number as unconfirmed for now.
Sources
- Previewing GPT-5.6 Sol: a next-generation model - OpenAI
- A preview of GPT-5.6 Sol, Terra, and Luna - OpenAI Help Center
- OpenAI limits GPT-5.6 rollout after government request - TechCrunch
- GPT-5.6 Sol, Terra, and Luna: OpenAI's Next-Generation Model Family - DataCamp
- OpenAI starts previewing GPT-5.6 and its three variants - Engadget