OpenAI announced GPT-5.6 on June 26, 2026, roughly two months after GPT-5.5, and the launch is unusual in two ways at once. The first is the product: GPT-5.6 ships as three models named Sol, Terra, and Luna (after the Sun, Earth, and Moon), introduces two new reasoning effort modes called max and ultra, and reports a state-of-the-art result on agentic coding. The second is the rollout: instead of the usual broad release, OpenAI is starting with a limited preview to roughly twenty government-approved partners after the White House formally asked the company to stagger access, citing the model’s offensive cybersecurity capability. Both stories matter, and they are tangled together, so this piece separates what GPT-5.6 is from how you can (and mostly cannot yet) get it.
A note on naming before anything else, because OpenAI changed the scheme with this release. The number (5.6) identifies the generation. The names (Sol, Terra, Luna) identify durable capability tiers that OpenAI says can each advance on their own cadence going forward. So Sol is the high-intelligence flagship, Terra is the balanced everyday model, and Luna is the fast, low-cost option. This replaces the older "Pro / Mini / nano" style suffixes, and it means future releases will likely keep these tier names while the generation number ticks up.
The short version: GPT-5.6 Sol is a genuine step up on long-horizon agentic work (coding, security research) and pushes the context window toward a reported 1.5 million tokens, but most people outside a small set of approved enterprises cannot use it yet, and the benchmark figures circulating are OpenAI’s own reported numbers from a preview, not independently reproduced results. Treat the capability claims as credible-but-unverified and the availability as genuinely restricted for now.
What GPT-5.6 actually is
GPT-5.6 is a model line, not a single model. There are three members at launch:
- GPT-5.6 Sol is the flagship. It is the most capable tier on coding, biology, and cybersecurity, and it is the model the new max and ultra reasoning modes attach to. This is the model the government-access conversation is about.
- GPT-5.6 Terra is the balanced tier, positioned for everyday work where you want strong quality without paying flagship rates or waiting on the deepest reasoning passes.
- GPT-5.6 Luna is the fast, affordable tier, positioned for high-volume work where latency and cost matter more than the absolute capability ceiling.
The Sun/Earth/Moon naming is the headline change in how OpenAI talks about its lineup. Under the new scheme the generation number and the tier name are independent: a future GPT-5.7 Sol or a refreshed Terra could ship without renaming the whole family. For anyone tracking model selection over time, this is the practical takeaway: learn the three tier names, because they are meant to persist.
The two new reasoning modes: max and ultra
GPT-5.6 Sol adds two reasoning effort options above the existing levels, and they are the most genuinely new capability in the release.
Max gives Sol the most time to reason deeply in a single pass. It is the "think harder, take longer" dial for problems where a one-shot deep deliberation beats a faster answer. This is an extension of the reasoning-effort controls that have existed since the GPT-5 family, pushed to a higher ceiling.
Ultra is the more architecturally interesting one. Instead of a single agent reasoning longer, ultra mode coordinates subagents to work on a complex task in parallel and then integrate the results. In other words, ultra goes beyond what one model instance can do alone by orchestrating several. This is OpenAI productizing the multi-agent pattern that, until now, developers had to assemble themselves with frameworks and orchestration code. For long, decomposable tasks (a large refactor, a multi-file investigation, a research synthesis), the subagent approach is the kind of thing that can change the shape of what a single API call accomplishes.
The caveat worth stating plainly: ultra mode is also the most expensive way to run the model, because subagents multiply the token spend. It is a capability to reach for on genuinely hard, parallelizable work, not a default.
Benchmarks and context window
The performance story centers on agentic, long-horizon work rather than on chat quality.
Coding. OpenAI reports that GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark that tests command-line workflows requiring planning, iteration, and tool coordination. The figures circulating put Sol in ultra mode around 91.9 percent and Sol on its own near 88.8 percent, compared with roughly 88 percent for Claude’s Mythos 5 line and around 83 percent for GPT-5.5. These are OpenAI-reported numbers from a preview, so the right way to read them is as a credible claim of leadership on this specific benchmark, pending independent reproduction.
Cybersecurity. OpenAI describes Sol as its most capable model yet for cybersecurity, including long-horizon tasks like vulnerability research and exploitation. This is the capability that triggered the government-access conversation, covered below.
Biology. OpenAI lists biology alongside coding and cybersecurity as an area of advancement. Specifics are thin in the preview, and biology capability is also part of why the safety framing around this release is heavier than usual.
Context window. Reports put the GPT-5.6 context window at roughly 1.5 million tokens, up from the roughly 1 million GPT-5.5 shipped with. As of the preview this is a reported figure rather than a fully documented API specification, so treat the exact number as provisional until the official API docs land at general availability.
The honest read on the whole benchmark picture: the gains are concentrated in agentic and security-adjacent work, which is exactly the area where independent verification matters most and is hardest to do quickly. The numbers are plausible and the direction is consistent with where OpenAI has been pushing, but "OpenAI reports" is doing real work in every sentence above.
Pricing
The reported API pricing follows the three-tier structure, per million tokens:
- Sol: $5 input / $30 output
- Terra: $2.50 input / $15 output
- Luna: $1 input / $6 output
Prompt caching reportedly gets explicit cache breakpoints and a 30-minute minimum cache lifetime, with cache writes at 1.25x the uncached input rate and cache reads discounted around 90 percent. As with everything else in a preview, treat these as the announced numbers rather than the final, locked rate card; pricing sometimes shifts between preview and general availability.
The government-gated release
This is the part of the GPT-5.6 story that is genuinely without recent precedent for a consumer-facing model.
On June 25, 2026, the Office of the National Cyber Director and the Office of Science and Technology Policy asked OpenAI to limit GPT-5.6’s rollout. The request followed a June 2 executive order that established a 30-day government cybersecurity evaluation window before broad release of frontier AI models. The evaluation for GPT-5.6 is expected to conclude around July 2, 2026.
The rationale centers on Sol’s offensive cybersecurity capability. The concern, as reported, is that a freely available model strong enough to materially assist with vulnerability research and exploitation should be assessed by the government before it reaches everyone, including potential adversaries. Sam Altman reportedly told employees in an internal Q&A on June 25 that GPT-5.6 would initially go only to a limited group of enterprise partners.
The practical consequences right now:
- Roughly twenty government-approved companies have early access while federal agencies run their evaluation.
- Standard ChatGPT and API access is not available for most developers and enterprises, including internationally (EU, UK, India, and APAC regions are reported to be excluded through standard channels) until the review completes.
- General availability is expected "in the coming weeks," with the July 2 evaluation window as the gating event.
OpenAI’s public posture is compliance-under-protest. The company stated it does not believe this kind of government access process should become the long-term default, arguing that restrictions also keep the tools out of the hands of the developers, enterprises, and defenders who need them. The framing is that this is a short-term concession tied to one release, not a new permanent model for how OpenAI ships.
For most readers, this is the operative fact: you probably cannot use GPT-5.6 today, and the timeline for when you can is tied to a government review rather than to a product schedule.
Codex and the developer surface
OpenAI is upgrading Codex alongside the model release, which is consistent with the agentic-coding emphasis of GPT-5.6 Sol. The preview materials reference Codex improvements (including better response times) but are light on specific technical detail about what changes in the harness itself. The reasonable expectation is that the agentic coding gains in Sol, plus ultra mode’s subagent orchestration, are most directly useful through Codex and similar agent surfaces rather than through plain chat, because that is where long-horizon, tool-coordinated work actually happens.
What is confirmed versus what is still rumor
Because this is a preview launched into an unusual access situation, it is worth being explicit about confidence levels.
Reasonably confirmed (from the announcement): the three-model Sol/Terra/Luna lineup, the new Sun/Earth/Moon naming scheme, the max and ultra reasoning modes (including ultra’s subagent approach), the limited-preview rollout, and the government-access request and its rationale.
Reported but not independently verified: the specific benchmark percentages on Terminal-Bench 2.1, the 1.5-million-token context window, the per-tier pricing, and the prompt-caching terms. These come from OpenAI’s preview and early reporting, not from independent testing or finalized API documentation.
Still open: the exact general-availability date (gated on the ~July 2 review), the final pricing, independent biology and cybersecurity eval results, and the precise list of which partners have early access.
If you are making a decision that depends on GPT-5.6 specifically, the safe assumption is that nothing is locked until OpenAI publishes the general-availability documentation after the government review concludes.
How GPT-5.6 compares
Against GPT-5.5 (April 2026), GPT-5.6 is an incremental generation focused on agentic coding, security, and the new reasoning modes, with a larger reported context window. It is not described as a full retrain the way GPT-5.5 was; it reads as a capability-and-tooling step on top of the GPT-5 line rather than a new architecture.
Against Claude’s Mythos 5 line, the reported Terminal-Bench 2.1 numbers put GPT-5.6 Sol narrowly ahead on that specific agentic-coding benchmark, with ultra mode opening a wider margin. One benchmark is not the whole story, and Anthropic’s models have their own strengths, so the honest framing is that GPT-5.6 Sol is competitive-to-leading on agentic coding as measured by OpenAI, not categorically dominant across all work.
The more important comparison for most teams is availability. GPT-5.5, Claude, and Gemini are all generally available today. GPT-5.6 is not. For any near-term project, that gap matters more than a few benchmark points.
What to do about it right now
For most readers the practical guidance is short. If you are not one of the roughly twenty approved partners, there is nothing to adopt yet; the useful move is to watch for the general-availability announcement expected after the early-July review and to keep your current model choice (GPT-5.5, Claude, or Gemini) in place until then. If you run agentic coding or security-research workloads and you are an enterprise that could qualify for early access, the path is through OpenAI’s enterprise and trusted-partner channels rather than the standard API. And if you are planning architecture that might use ultra mode’s subagent orchestration, it is worth designing for it conceptually now (decomposable, parallelizable tasks) so you are ready when access opens, while remembering that subagent runs carry multiplied token cost.
Frequently asked questions
What is GPT-5.6? GPT-5.6 is OpenAI’s frontier model line announced June 26, 2026, shipping as three tiers (Sol, Terra, Luna) with two new reasoning modes (max and ultra) and a reported context window near 1.5 million tokens. It launched as a limited preview rather than a broad release.
What is the difference between ChatGPT 5.6 and GPT-5.6? GPT-5.6 is the underlying model line. ChatGPT is the consumer product that runs OpenAI’s models. People often say "ChatGPT 5.6" to mean the GPT-5.6 models as they will eventually appear inside ChatGPT, but as of the preview the models are not broadly available in ChatGPT yet.
What do Sol, Terra, and Luna mean? They are durable capability tiers named after the Sun, Earth, and Moon. Sol is the high-intelligence flagship, Terra is the balanced everyday model, and Luna is the fast, low-cost option. The generation number (5.6) is separate from the tier name, so the names are meant to persist across future generations.
What are max and ultra reasoning modes? Max gives the model the most time to reason deeply in a single pass. Ultra goes further by coordinating subagents to work on a complex task in parallel and integrate the results, which is OpenAI productizing the multi-agent pattern. Ultra is the most capable and the most expensive option because subagents multiply token usage.
Can I use GPT-5.6 right now? Most people cannot. At launch, access is limited to roughly twenty government-approved partners while a federal cybersecurity evaluation runs. General availability is expected "in the coming weeks," gated on a review expected to conclude around July 2, 2026. Standard ChatGPT and API access, including in the EU, UK, India, and APAC, is reported to be unavailable through normal channels until then.
Why is the release restricted? The White House (through the Office of the National Cyber Director and the Office of Science and Technology Policy) asked OpenAI to stagger the rollout, citing GPT-5.6 Sol’s offensive cybersecurity capability and a June 2 executive order that set a 30-day government evaluation window for frontier models. OpenAI complied while publicly stating it does not want this to become the long-term default.
How much does GPT-5.6 cost? Reported API pricing per million tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. These are preview-stage numbers and could change before general availability.
Is GPT-5.6 better than Claude or Gemini? On the reported Terminal-Bench 2.1 agentic-coding benchmark, GPT-5.6 Sol edges ahead of Claude’s Mythos 5 line, with ultra mode widening the gap. That is one OpenAI-reported benchmark, not a verdict across all tasks, and Claude and Gemini are generally available today while GPT-5.6 is not. For near-term work, availability is the deciding factor.
How large is the context window? Reports put it near 1.5 million tokens, up from roughly 1 million in GPT-5.5. This is a reported figure pending official API documentation at general availability.
Should I wait for GPT-5.6 or use GPT-5.5 now? Unless you are an approved early-access enterprise, you do not have a choice to make yet, because GPT-5.6 is not broadly available. Keep using GPT-5.5 (or Claude or Gemini) for current work and reassess when OpenAI publishes general-availability details after the early-July government review.
