GPT‑5.6 Is Here: OpenAI's "Sol, Terra, Luna" Trio Lands After an Unprecedented Government Review

The wait is over — and this launch was unlike any in AI history.

Today, OpenAI officially released the GPT‑5.6 family to the public: Sol, the new flagship; Terra, a balanced everyday workhorse; and Luna, the budget speedster. But the headline isn't just the models. It's how they got here: GPT‑5.6 is the first frontier AI release to clear a U.S. government security review before reaching the public.

The launch Washington paused

When OpenAI first rolled out GPT‑5.6 in late June, almost nobody could touch it. Access was limited to a small group of trusted partners at the request of the Trump administration, following a June executive order asking AI companies to voluntarily submit their most powerful models for federal review before public release.

The Department of Commerce's Center for AI Standards and Innovation put the models through additional testing, with OpenAI dispatching technical experts to Washington to field questions in real time. The concern? GPT‑5.6's cybersecurity capabilities are genuinely startling — more on that below. Approval came through this week, and Sam Altman marked the moment with two words on X: "Happy building."

The timing carries extra drama: the launch arrives just after rival Anthropic restored access to its Claude Fable 5 and Mythos 5 models following its own weeks-long standoff with the government over an export control directive. Frontier AI, it seems, now ships through Washington.

Three tiers, one philosophy: more intelligence per token

OpenAI's pitch for GPT‑5.6 is refreshingly economic: not just smarter, but cheaper per unit of useful work.

• Sol ($5/$30 per 1M tokens in/out) — the flagship, claiming state-of-the-art results across coding, knowledge work, cybersecurity, and science
• Terra ($2.50/$15) — matches or beats GPT‑5.5 at half the price
• Luna ($1/$6) — the fastest, cheapest tier, reportedly outperforming Claude Opus 4.8 on coding at a fraction of the cost

The naming scheme is deliberate: the number marks the generation, while Sol, Terra, and Luna are durable capability tiers that can now evolve on their own schedules.

The numbers that matter

On Agents' Last Exam — a grueling benchmark of long-running professional workflows across 55 fields — Sol posted a new high of 53.6, beating Anthropic's Claude Fable 5 by over 13 points. On the Artificial Analysis Coding Agent Index, Sol set a new record of 80, edging out Fable 5 while using less than half the output tokens and roughly a third less money.

Sol also claimed new state-of-the-art marks on BrowseComp (92.2% with the new multi-agent mode) and OSWorld 2.0 (62.6%), the leading tests of agentic browsing and computer use.

The honest caveat: it's not a clean sweep. Anthropic's Fable 5 still leads on SWE-Bench Pro by a wide margin, edges Sol on the Artificial Analysis Intelligence Index, and holds the crown on FrontierMath's hardest tier. Investor Matt Shumer's early take captured the split verdict: an amazing model, but Fable still won most of his personal tests. The realistic picture is a two-horse frontier race where your best model depends on your task — and your budget.

Meet "ultra": the swarm setting

The most intriguing new feature is ultra — a maximum-capability mode that coordinates four AI agents working in parallel by default, trading more tokens for better answers, faster. On benchmarks like BrowseComp and Terminal-Bench, adding parallel agents pushed scores up while cutting time-to-result. Developers can build their own multi-agent swarms through a new beta in the Responses API, and some configurations scale to 16 agents.

Alongside it comes Programmatic Tool Calling: instead of ping-ponging every tool response back through the model, GPT‑5.6 writes and runs lightweight programs that coordinate tools, filter intermediate data, and adapt on the fly. Early adopters report token savings of 24–63% on complex workflows.

Design taste, finally?

One of the more surprising claims: GPT‑5.6 has developed genuine design judgment. It doesn't just generate frontend code — it uses computer-use skills to inspect the rendered result, catch visual glitches, and polish before delivery. Canva reported it beat competitors on slide creation while being 1.6x more token-efficient; Triple Whale scored it 4.4/5 on frontend QA versus 4.0 for GPT‑5.5. It can also infer a slide deck's entire design system — typography, spacing, master-slide rules — and apply it faithfully to new content.

The cybersecurity elephant in the room

Here's why the government wanted a look first. On ExploitBench, which measures progress toward working code exploits, Sol scored 73.5% versus GPT‑5.5's 47.9%. On ExploitGym, which turns real-world vulnerabilities into functional exploits, it more than doubled its predecessor's pass rate.

OpenAI's response is a layered "Trusted Access" system: the most sensitive cyber capabilities are reserved for identity-verified individuals and vetted organizations doing legitimate defensive work — vulnerability triage, malware analysis, patch validation. Everyone else hits safeguards that block roughly ten times more potentially harmful activity than previous models. Before launch, OpenAI burned approximately 700,000 GPU-hours on automated red teaming alone.

Notably, OpenAI argues that overblocking is itself a security risk — locking defenders out while attackers turn to open-source alternatives. It's a philosophical stake in the ground for how dual-use AI should be governed.

The recursive twist

Buried in the announcement is perhaps the most consequential detail: GPT‑5.6 is accelerating OpenAI itself. Internal researchers' daily token usage more than doubled versus GPT‑5.5's peak, the share of research compute devoted to internal coding inference grew 100-fold in six months, and OpenAI now publishes an internal "RSI Index" — a bundle of evaluations measuring progress toward recursive self-improvement — where Sol jumped 16.2 points over GPT‑5.5.

The models are being used to build the next models. That flywheel, more than any single benchmark, may be the real story of 2026.

The bottom line

GPT‑5.6 arrives as the most efficient frontier model family yet, wrapped in the heaviest safety apparatus OpenAI has ever built, blessed by a government process that didn't exist a year ago. Rollout is underway across ChatGPT, Codex, and the API over the next 24 hours, with a new Grok release reportedly waiting in the wings.

The speculation phase is over. Now the builders get to vote.