OpenAI launched GPT-5.6 today — Sol, Terra, and Luna — and most developers cannot touch any of them. At the White House’s request, only around 20 government-approved companies have access while federal agencies run a security evaluation under a June 2 executive order. The models are real, the benchmarks are impressive, and a broad release is weeks away. But for now, the most capable AI model OpenAI has ever shipped sits behind a government access list that most developers are not on.
What Happened and Why #
On June 25, the Office of the National Cyber Director and the Office of Science and Technology Policy formally asked OpenAI to limit GPT-5.6’s initial rollout. The request tied to a Trump executive order signed June 2 that established a framework for government cybersecurity teams to evaluate powerful AI models before broad release — a window of up to 30 days. The government’s concern centers on Sol’s cybersecurity capabilities: the model reportedly matches Claude Mythos 5 in offensive security tasks, the kind of capability Washington wants to assess before it becomes freely available to anyone with an API key.
OpenAI complied but was unambiguous about its view. “We don’t believe this kind of government access process should become the long-term default,” the company stated, noting that restrictions keep advanced tools away from developers, enterprises, cyber defenders, and global partners who need them. That’s a pointed position — wrapped in compliance.
International developers are caught in the middle. EU, UK, India, and APAC developers cannot access GPT-5.6 through normal ChatGPT or API tiers until the review completes around July 2. A model positioned as global infrastructure is, for now, a US government procurement decision.
The Three Models: What You’re Waiting For #
GPT-5.6 introduces a new naming scheme. Sol, Terra, and Luna are permanent capability tiers, not one-off names. The version number (5.6) identifies the generation; the tier name identifies the capability level. Future Sol models will advance on their own cadence — a meaningful structural change in how OpenAI positions its model lineup.
Sol is the flagship. At $5 input / $30 output per million tokens, it gets a 1.5M token context window (up 43% from GPT-5.5’s 1.05M), two new reasoning modes — max
for deeper single-pass reasoning and `ultra`
for multi-subagent coordination — and the strongest agentic performance OpenAI has shipped.
Terra is the one most developers should be planning for. It delivers GPT-5.5-level performance at half the price: $2.50 input / $15 output per million tokens. If you are currently using GPT-5.5, Terra is your drop-in replacement when it goes GA.
Luna targets speed and cost at $1 input / $6 output per million tokens, described by OpenAI as “strong capability at lowest cost.”
| Model | Input / 1M tokens | Output / 1M tokens | Context Window |
|---|---|---|---|
| Sol | $5.00 | $30.00 | 1.5M tokens |
| Terra | $2.50 | $15.00 | 1.5M tokens |
| Luna | $1.00 | $6.00 | 1.5M tokens |
Sol’s Benchmark Performance #
On Terminal-Bench 2.1 — which tests command-line workflows requiring planning, iteration, and tool coordination — Sol Ultra scores 91.91% and Sol scores 88.8%, compared to Claude Mythos 5 at 88% and GPT-5.5 at 83.4%. Sol is also the only model above 50% on Agent’s Last Exam in code mode (50.9%), which measures complex multi-step task completion. On GeneBench v1, Sol achieves stronger biology results using fewer tokens.
| Model | Terminal-Bench 2.1 |
|---|---|
| Sol Ultra | 91.91% |
| Sol | 88.8% |
| Claude Mythos 5 | 88.0% |
| Fable 5 | 84.3% |
| GPT-5.5 | 83.4% |
The ultra
mode is the differentiator worth watching. Rather than extending a single reasoning chain, it spins up subagents to tackle complex tasks in parallel. That’s a meaningful architectural step beyond what GPT-5.5 offers with its reasoning-effort settings.
Prompt Caching Gets a Real API #
GPT-5.6 redesigns prompt caching with explicit cache breakpoints — developers can specify exactly what gets cached rather than relying on OpenAI’s inference. Cache lifetimes now have a guaranteed 30-minute minimum, up from best-effort. Cache writes cost 1.25x the uncached input rate; cache reads get a 90% discount. If you have cost optimization logic built around current caching behavior, plan a review before you migrate to GPT-5.6.
What to Do While You Wait #
Most developers: GPT-5.5 is still your production model. Nothing in your stack has changed today. But the GA release is coming, and preparation pays off:
Plan Terra as your GPT-5.5 replacement. Same performance, half the cost — the migration case is straightforward.Audit your chunking logic. The 1.5M context window may eliminate retrieval workarounds you’ve built around GPT-5.5’s limit.Review your caching implementation. Explicit breakpoints work differently from the current inference-based system.Watch for the GA announcement. OpenAI says “coming weeks.” The government’s 30-day evaluation window closes around July 2.
OpenAI has made clear it does not want to repeat this pattern. The government restriction was framed as a short-term concession, not a new release model. Whether Washington agrees will shape how the next major OpenAI model ships. In the meantime, the official GPT-5.6 announcement has full technical details, and the OpenAI pricing page will update when public API access opens. For broader context on the government restriction story, VentureBeat’s coverage has additional detail on the partner selection process.