OpenAI's GPT-5.6 Sol goes multi-agent

OpenAI just pulled back the curtain on its GPT-5.6 model family, but you probably can't get your hands on it today.

The new lineup consists of three tiers: the flagship Sol, the mid-range workhorse Terra, and the budget-conscious Luna. But instead of an immediate worldwide rollout, OpenAI is kicking things off with a tight preview for a handful of vetted partners. The catch? It happened at the explicit request of the U.S. government, as frontier cyber capabilities get too close for regulatory comfort.

Meet Sol, Terra, and Luna

OpenAI is shifting to a fresh naming scheme: numbers designate the architectural generation, while celestial names mark persistent capability tiers that can evolve on their own schedule.

Sol is the crown jewel. It features a new max reasoning mode to let the system think out problems longer, along with an Ultra mode that orchestrates multiple subagents to divide and conquer long-horizon assignments. On Terminal-Bench 2.1, which benchmarks multi-step command-line coding, Sol set a new state of the art. In biology workflows on GeneBench v1, it beat GPT-5.5 while burning fewer tokens.

Meanwhile, Terra delivers performance comparable to GPT-5.5 at half the price, and Luna lands as the quickest, cheapest option. In API pricing per million input and output tokens, Sol sits at $5 and $30, Terra at $2.50 and $15, and Luna at $1 and $6.

700k GPU hours and a real-time kill switch

Cybersecurity is where GPT-5.6 gets spicy. On the ExploitBench evaluation, Sol rivaled Mythos Preview while utilizing roughly one-third of the output tokens. In live browser audits across Chromium and Firefox, Sol managed to unearth zero-day bugs and exploitation primitives, though OpenAI noted it stopped short of independently chaining together a full-scale, weaponized exploit.

To lock this down, OpenAI burned through over 700,000 A100-equivalent GPU hours on automated red-teaming to patch universal jailbreaks. The safety stack is remarkably aggressive: real-time misuse classifiers watch output as it streams. If a query looks suspicious, generation pauses mid-sentence while a larger reasoning model inspects the full context. If it flags a violation, the text is killed before the user ever sees it.

OpenAI admitted that this inspection layer, combined with account-level tracking, will likely cause frustrating false positives and latency spikes for legitimate security researchers and defenders during the trial.

Why it matters

For enterprise builders and developers, Terra and Luna look like absolute steals. Cutting GPT-5.5-level reasoning costs by 50% gives immediate breathing room to margins, while Sol's native multi-agent orchestration removes the hassle of hand-crafting external routing pipelines for complex workflows.

For the broader industry, OpenAI's coordinated nod to the White House marks a turning point. Stating publicly that this restricted rollout happened at the government's request—while openly grumbling that government clearance should not become the industry default—shows just how close advanced cyber models are to being classified as sensitive dual-use munitions. Expect Anthropic and Google to face identical regulatory hurdles on upcoming frontier releases.

We think OpenAI's offensive-to-defensive framing makes sense on paper: giving defenders tools that spot vulnerabilities with a fraction of the compute is a real win. But in practice, pausing active generation to run heavy contextual moderation is a clumsy user experience. If frontier models become so restricted and jumpy that white-hat security researchers get constantly flagged as malicious actors, builders might seek out unconstrained open-weights models to do the dirty work.