OpenAI Starts Previewing GPT-5.6 — Sol, Terra and Luna in a Limited Release (2026)

According to reporting by Engadget and to OpenAI's own preview announcement, OpenAI has begun a limited preview of GPT-5.6 — a new model series split into three named tiers, Sol, Terra and Luna — made available first to a small group of trusted partners and organizations through the OpenAI API and Codex, with broader availability planned for the coming weeks. The launch is notable not just for the models themselves but for the structure around them: a new naming scheme that decouples a model's generation from its capability tier, a tiered pricing ladder, and an unusual limited rollout that OpenAI says was shaped by a US government review process. This article walks through what was actually reported about GPT-5.6, how the three variants differ, where it fits in the broader 2026 model landscape, and the caveats worth keeping in mind. All product claims are attributed to OpenAI and to the cited reporting; figures are as reported.

What's new in GPT-5.6

Per OpenAI's announcement and Engadget's coverage, GPT-5.6 is a single new generation delivered as three distinct models rather than one monolithic release. The headline change is conceptual as much as technical: instead of a single flagship that everyone shares, OpenAI is shipping a family in which each member targets a different balance of capability, speed and cost. According to OpenAI, the most capable member, Sol, is described as its strongest model to date, with agentic improvements highlighted in areas such as coding, biology and cybersecurity. Engadget reported that Sol supports a "max" reasoning-effort setting, positioning it for the hardest problems where deeper deliberation is worth the extra compute and latency.

The second notable shift is the rollout itself. Rather than a broad public launch, OpenAI said GPT-5.6 is initially available only to a limited group of trusted partners and organizations, reachable through the API and Codex. According to Engadget, the participation of those partners was shared with the US government, and OpenAI plans to make the family generally available in ChatGPT, Codex and the API in the coming weeks. We attribute the specifics of who has access, and the timeline, to the reporting rather than presenting them as independently confirmed facts.

OpenAI also emphasized safety work tied to the most powerful tier. As reported, Sol ships with strengthened protections around high-risk activities and is positioned as a strong option for helping users find and fix software vulnerabilities in legitimate cybersecurity work. Engadget reported that the model went through a substantial red-teaming effort — described as roughly 700,000 GPU hours spent searching for universal jailbreaks — before the preview. That figure comes from the reporting; the underlying methodology and what exactly it measured are not something this article can independently verify, so treat it as a reported detail rather than a benchmark.

The three variants: Sol, Terra and Luna

The clearest way to understand GPT-5.6 is through its three tiers, each aimed at a different job. According to OpenAI, Sol is built for the hardest problems — complex coding, security research and other demanding agentic tasks — and is described as the company's strongest model yet, with that "max" reasoning effort available when needed. It sits at the top of the pricing ladder, reflecting its positioning as the premium, highest-capability option.

Terra is the balanced, everyday-business tier. OpenAI frames it for high-volume work like customer support, internal tools and document analysis, describing it as having competitive performance to the previous GPT-5.5 generation while costing roughly half as much. In practical terms, that makes Terra the natural default for teams that were already happy with GPT-5.5-class quality and primarily want to lower their per-token bill. Luna, the third tier, is positioned as the fast, lowest-cost option for routine work such as summarization, drafting and lightweight automation — the model you reach for when throughput and price matter more than maximum reasoning depth.

A key part of the announcement is the naming system itself. According to OpenAI, in this scheme the number — 5.6 — identifies the model's generation, while Sol, Terra and Luna identify durable capability tiers that can each advance on their own cadence. In other words, the names are meant to outlast a single release: a future generation could ship its own Sol, Terra and Luna, letting customers reason about which tier they want independently of which generation they are on. Whether that naming convention sticks and stays clear over multiple generations is something only time will tell, but the stated intent is to make the lineup easier to navigate than a sprawl of version numbers.

How GPT-5.6 fits into the 2026 model landscape

GPT-5.6 lands in a 2026 market where the major labs have largely converged on the idea of model families rather than single flagships, with tiers spanning a spectrum from cheap-and-fast to expensive-and-deliberative. OpenAI's move to give those tiers durable names — Sol, Terra, Luna — rather than suffixes like "mini," "nano" or "pro" is partly a branding decision and partly an attempt to make the tradeoffs legible to buyers who increasingly mix and match models inside a single product. The explicit framing of Terra as "GPT-5.5-class quality at half the cost" is the kind of price-performance message that has defined much of the year's competition, where each incremental release is judged as much on cost efficiency as on raw capability.

The other defining feature of this launch is regulatory. According to Engadget and TechCrunch, the limited rollout is connected to a US government review: an executive order on AI and cybersecurity is reported to require companies to submit sufficiently powerful models for voluntary review some weeks before public release. OpenAI made clear it views the arrangement as a one-off rather than a precedent, stating, as reported, that "we don't believe this kind of government access process should become the long-term default." That tension — between shipping frontier models quickly and satisfying pre-release government scrutiny — is likely to be a recurring theme for the rest of 2026, and GPT-5.6 is one of the first high-profile examples of it playing out in public. As with the other claims here, the details of the order and the review process are attributed to the reporting.

GPT-5.6 variants at a glance

The table below summarizes the three tiers and their listed API pricing. All figures are as reported by OpenAI and Engadget and have not been independently verified by comparee.ai:

What it means in practice

For developers and teams, the most immediately useful takeaway is the pricing ladder, because it maps cleanly onto how people actually deploy models. The roughly 5x spread between Luna's output price and Sol's output price is large enough that picking the right tier per task — rather than routing everything to the most capable model — becomes a meaningful cost lever. The likely pattern, based on how the tiers are described, is to default to Terra or Luna for the bulk of routine, high-volume work and reserve Sol for the genuinely hard cases where its stronger reasoning earns its premium. OpenAI's framing of Terra as GPT-5.5-quality at half the price is essentially an invitation for existing GPT-5.5 users to cut their bill without changing what they build.

It is worth being precise about what this preview does and does not establish. It tells us OpenAI's positioning, its naming intent and its listed prices — all useful and all attributable. It does not, on its own, prove how the models perform on independent benchmarks, because no neutral third-party evaluations were part of the announcement covered here. The "700,000 GPU hours" red-teaming figure and the "strongest model yet" description are claims from OpenAI and the reporting, not verified results. Until the family is generally available and outside testers can probe it, the honest stance is to treat GPT-5.6 as a clearly defined, attributed product launch whose real-world quality remains to be measured.

The bottom line

GPT-5.6's preview is as much about structure as about raw capability: OpenAI is shipping one generation as three durably named tiers — Sol, Terra and Luna — with a pricing ladder that runs from a $1/$6 economy option to a $5/$30 flagship, and it is doing so through an unusually controlled rollout shaped by a US government review it openly says should not become the norm. As reported by Engadget and OpenAI, the most concrete, verifiable details right now are the lineup, the naming logic and the prices; the capability claims are OpenAI's own and await independent testing once the models reach general availability in the coming weeks. For anyone building on these models, the practical move is to watch for that wider release, then match each tier to the task — Luna and Terra for volume, Sol for the hard problems — rather than over-indexing on the flagship by default.

Disclaimer: based on reporting by Engadget and OpenAI's own preview announcement (with corroborating coverage from TechCrunch), linked below; figures, prices and quotes are as reported and have not been independently verified by comparee.ai.