On 26 June 2026, OpenAI launched GPT-5.6. Not one model, but three at once: Sol, Terra and Luna. And for the first time in the company's history, the launch of a flagship model comes with a phrase no one would have imagined reading a year ago: access is deliberately restricted, at the request of the US government.
If you follow AI news, this scenario should ring a bell. It's exactly what happened to Anthropic a few weeks earlier. We'll take a complete look at what actually came out, what these models are worth, and why this release is as much a political signal as a tech story.
What actually came out: three models, not one
GPT-5.6 abandons the single-model logic. OpenAI introduces a new naming convention where the number denotes the generation, while Sol, Terra and Luna denote durable capability tiers that will each evolve at their own pace. It's the same logic as the Opus, Sonnet and Haiku range at Anthropic.
Here's how the three break down. Sol is the flagship, the most powerful model, reserved for the hardest tasks: long-horizon coding, security research, deep reasoning. Terra is the balanced everyday model, pitched as performing as well as GPT-5.5 but at half the price. Luna is the fast, economical model, built for volume and high-frequency simple tasks.
Sol also carries two new features the others don't have. A max mode, a new reasoning setting that gives the model the most time to think, and an ultra mode that splits a task across several sub-agents working in parallel. It's this ultra mode that produces the best benchmark scores.
Pricing: Sol holds, Terra undercuts, Luna crushes
This is where the strategy becomes clear. Here are the official prices per million tokens.
| Model | Input | Output | Positioning |
|---|---|---|---|
| GPT-5.6 Sol | $5 | $30 | Flagship, same price as GPT-5.5 |
| GPT-5.6 Terra | $2.50 | $15 | Balanced, GPT-5.5 quality at half price |
| GPT-5.6 Luna | $1 | $6 | Fast and economical, high volume |
Two things stand out. Sol keeps exactly the same price as GPT-5.5, making it a free upgrade for anyone already running on the old flagship. And Terra offers a level close to GPT-5.5 for half the price, which looks like the new norm: last generation's flagship quality, at a mid-range price.
OpenAI has also revamped its prompt caching system, a technical detail that weighs heavily for agents. Cache reads keep a 90% discount, with explicit breakpoints and a minimum lifetime of 30 minutes. For an agent that resends a large stable context at every step, like a codebase or a long system prompt, these savings matter as much as the raw token price.
Benchmarks: powerful, but read with caution
On Terminal-Bench 2.1, the reference test for autonomous command-line coding, Sol in ultra mode hits 91.9%, and standard Sol 88.8%. For comparison, Claude Mythos 5 sits around 88.0%, Terra and Fable 5 are tied at 84.3%, and GPT-5.5 at 83.4%.
An amusing detail that shows a marketing tier isn't a per-task guarantee: on this specific benchmark, Luna, the cheapest model, beats Terra, which is positioned higher. The tiers describe an intelligence-speed-cost balance averaged over many tasks, not an absolute hierarchy on every test.
On the cyber front, OpenAI claims Sol rivals Anthropic's Mythos Preview on the ExploitBench benchmark while using about a third of the output tokens. But the company acknowledges that Sol did not produce a complete, functional exploit autonomously under the tested conditions, and that Mythos remains ahead.
The real story: the state takes hold of the tap
Here's what makes this launch historic. GPT-5.6 is not available to the general public. During the preview phase, it's only accessible via the API and Codex, to around twenty partner companies whose participation was validated by the US government.
OpenAI says it presented its plans and the models' capabilities to the administration in advance, and started with restricted access at its request. This is a direct continuation of what happened to Anthropic, which was forced to cut access to Fable 5 and Mythos 5. The difference is that OpenAI is no longer alone in the crosshairs: Washington now treats the most advanced models as products requiring government review before wide release.
OpenAI doesn't hide its discomfort. The company said it doesn't believe this kind of government access process should become the long-term norm, arguing it deprives users and developers of the best tools. But it accepts this short phase, presented as the fastest path to wide availability, while building a repeatable regulatory framework with the administration around cybersecurity.
The underlying reason is in the system card. OpenAI classifies Sol, Terra and Luna as High capability in both cybersecurity and biological and chemical risk. None reach the critical threshold for AI self-improvement, but this High level is enough to trigger reinforced guardrails and a staged release. The company says it spent over 700,000 GPU compute hours automatically searching for jailbreak-type flaws.
For the general public: patience, and a domino effect on prices
Let's be clear: if you use ChatGPT daily, you can't touch GPT-5.6 yet. Wide deployment on ChatGPT, Codex and the API is announced for the coming weeks, but without a firm date, and it also depends on the government's green light.
What concerns you right now is the effect on prices. When the flagship holds its price and the mid-tier model halves its own at equal quality, the whole market price grid comes under downward pressure. This dynamic adds to the one triggered by Chinese open-source with GLM-5.2. For the end user, competition is working at full tilt, and that's good news.
For the enterprise: routing becomes the key skill
For a company, the real novelty of GPT-5.6 isn't a smarter model, it's that model choice becomes an architecture decision in its own right. The obvious logic is routing: send heavy, ambiguous tasks to Sol, regular production work to Terra, and routine volume to Luna.
Many products won't call Sol by default. They'll call Luna or Terra first, then only escalate to Sol when the task is difficult, sensitive, or costly to get wrong. This escalation mechanic, combined with the new cache, also changes the game on prompt engineering: a poorly structured prompt is no longer just a quality problem, it's now a direct cost problem.
Then there's the angle that matters to any technical leadership: access predictability. The lesson of recent weeks, between Anthropic's cut and OpenAI's locked-down preview, is that a cutting-edge model can be restricted or withdrawn by political decision overnight. Building your entire infrastructure on a single closed vendor becomes a strategic risk, not just a budgetary one.
The real message of this launch
GPT-5.6 is technically impressive: three models, aggressive pricing, novel reasoning modes, scores that beat the competition on agentic coding. But that's not the story. The story is that a cutting-edge model launch has become an event that's half technological, half geopolitical.
Raw capability no longer defines a model. What defines it now is the capability you can actually call, when you need it, without a government pulling the plug. In mid-2026, the frontier is no longer just about benchmarks. It's about access.