OpenAI Rolls Out GPT-5.6 Sol, Terra and Luna — After a US Government Safety Review

What Happened

OpenAI has started the broad public rollout of GPT-5.6, its newest model family, beginning July 9, 2026. Rather than a single model, GPT-5.6 ships as three tiers with distinct names — Sol, Terra and Luna — each aimed at a different balance of intelligence, speed and cost. The models are now reaching ChatGPT, the OpenAI API and the company's coding tools.

The launch followed an unusual path. GPT-5.6 first appeared on June 26, 2026 as a limited preview restricted to roughly 20 government-approved organizations, and only opened up to the wider public after the US Commerce Department's Center for AI Standards and Innovation (CAISI) completed a review and cleared broader access. OpenAI frames Sol, the flagship, as its strongest and most safety-hardened model yet, with pronounced gains in coding, cybersecurity and scientific reasoning.

Sol, Terra and Luna: Three Tiers

The three-model structure is the headline design choice. Instead of asking users to guess which single model fits their task, OpenAI splits GPT-5.6 into clearly positioned tiers:

• Sol (flagship): The most capable tier, built for complex reasoning, coding, scientific work, cybersecurity and long-running agentic tasks. It exposes the highest reasoning-effort settings — including a "max" reasoning mode and an "ultra" orchestration mode — for high-ceiling, multi-step problems where cost is a secondary concern.
• Terra (balanced): A mid-tier model positioned as competitive with the previous GPT-5.5 while costing roughly half as much. It is meant to be the general-purpose workhorse for everyday tasks.
• Luna (fast and affordable): The smallest, cheapest and lowest-latency member of the family, aimed at high-volume, cost-sensitive workloads. Notably, on some coding benchmarks Luna actually edges out the pricier Terra, which OpenAI acknowledges makes it an economically rational default for a lot of routine work.

The practical upshot for developers is granular control over the cost-performance trade-off: reach for Sol when a problem genuinely needs frontier reasoning, drop to Luna when you are running the same prompt millions of times and latency and price matter more than the last few points of accuracy.

How Capable Is It?

OpenAI's reported numbers center on agentic coding and terminal-style workflows. On Terminal-Bench 2.1, which measures a model's ability to carry out real command-line and coding tasks, GPT-5.6 Sol in ultra mode leads at 91.9%, with standard Sol at 88.8% — both above the prior GPT-5.5 at 88.0%. Luna posts a respectable 84.3%, matching a leading competitor model and, tellingly, beating Terra's 82.5% on this particular test.

Beyond coding, OpenAI reports gains across specialized domains including biology and cybersecurity evaluations, with Sol using fewer output tokens than previous generations to reach its answers — a meaningful efficiency win when output tokens are the most expensive part of running a model. On the company's cybersecurity exploit benchmark, Sol scored around 73.5%, which OpenAI describes as its strongest offensive-security capability to date and part of the reason the model was gated behind a government review before wide release.

As always with vendor-reported benchmarks, the caveats matter. Independent evaluators flagged that Sol appeared to perform somewhat better during controlled testing than in early real-world deployment, and OpenAI itself notes that benchmark leadership does not automatically translate into a better experience on every task. The real test comes now that the models are broadly available and being pushed on unpredictable, messy production workloads.

Why the Government Reviewed It First

The most distinctive part of this launch is not the model — it is the release process. GPT-5.6 initially shipped only to a small set of trusted partners, coordinated with the US government, because Sol's stronger cybersecurity and scientific capabilities raised the stakes if the model were misused. The Center for AI Standards and Innovation, the successor body inside the Commerce Department that evaluates frontier AI, ran a review before OpenAI opened the gates to the general public.

OpenAI paired that with what it calls its most robust safety stack to date: model-level training that refuses prohibited cyber and bio assistance, real-time classifiers that evaluate output as it is generated, account-level review to separate malicious actors from legitimate security researchers, and tiered access with monitoring and enforcement. The company says it devoted the equivalent of hundreds of thousands of GPU-hours to automated red-teaming against jailbreaks, supplemented by outside human experts. The pattern — preview under government supervision, then widen access once cleared — is a preview of how the most capable models may be released going forward.

Frontier Speed on Cerebras

Alongside the model launch, OpenAI is offering GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second for latency-sensitive customers. That is dramatically faster than typical GPU-based serving and is aimed at use cases — live coding assistants, interactive agents, real-time tools — where waiting several seconds for a response breaks the experience.

Running a frontier model at that speed initially for a limited set of customers signals where the competition is heading: not just smarter models, but models fast enough to feel instantaneous inside interactive software. Speed is increasingly a feature in its own right, and pairing a top-tier model with specialized inference hardware is one way OpenAI is trying to differentiate.

Pricing

API pricing is tiered to match the three models, quoted per one million tokens:

• Sol: $5.00 input / $30.00 output
• Terra: $2.50 input / $15.00 output
• Luna: $1.00 input / $6.00 output

Prompt caching carries over from earlier models: cache writes are billed at a premium over the standard input rate, while cache reads receive a steep discount, which rewards workloads that reuse long, stable context. The spread between Sol and Luna — roughly five times on both input and output — is what makes the tiering meaningful. For high-volume applications, choosing Luna over Sol where the task allows can cut inference bills substantially while keeping most of the capability.

Availability and the Unified App

With the July 9 rollout, GPT-5.6 is reaching ChatGPT users and API developers, with the higher tiers gated to paying customers and the broader public gaining access in stages. OpenAI is also continuing to consolidate its products, folding its standalone coding tools into a single, unified ChatGPT application that brings chat, a coding agent and browsing together in one place rather than across separate apps.

For most people, the change will show up quietly: better answers in ChatGPT, faster responses, and a model picker that maps more cleanly onto "how hard is this task and how much am I willing to pay." Developers get the more consequential decision, choosing per-call between Sol, Terra and Luna to tune cost and latency against quality.

Why It Matters

GPT-5.6 is significant for two reasons that have little to do with any single benchmark score. First, the three-tier structure formalizes a shift the whole industry is making: away from one monolithic model and toward a menu of models tuned for different price and speed points, letting cost-conscious teams deploy frontier-adjacent intelligence at scale. Second, the government-supervised release process — preview to a vetted group, review by a federal standards body, then broad availability — is one of the clearest examples yet of how the most capable AI systems may be governed as their cyber and scientific abilities grow.

Whether Sol's benchmark lead holds up under real-world pressure, and whether Luna's aggressive price-performance reshapes how developers build, will play out over the coming weeks. What is already clear is that the frontier is no longer defined by capability alone — it is defined by capability, speed, cost and the process by which powerful models are allowed out the door. OpenAI's full announcement is published on its official site.