OpenAI has formally begun a restricted preview of the GPT‑5.6 mannequin collection Sol, Terra, and Luna, positioning its flagship Sol as the corporate’s most succesful and security-hardened AI mannequin so far, accessible initially solely to a small group of trusted companions on the formal request of the Trump administration.
The GPT‑5.6 household introduces three distinct functionality tiers below a brand new naming system. Sol is the flagship mannequin; Terra is a balanced mannequin for on a regular basis work, delivering aggressive efficiency to GPT‑5.5 at 2x decrease value; and Luna is a quick, inexpensive mannequin designed to deliver sturdy AI functionality on the lowest worth level within the lineup.
Within the new naming conference, the quantity identifies the mannequin’s technology, whereas Sol, Terra, and Luna establish sturdy functionality tiers that may every advance on their very own cadence.
Pricing for GPT‑5.6 is structured per million tokens: Sol at $5 enter / $30 output, Terra at $2.50 enter / $15 output, and Luna at $1 enter / $6 output. OpenAI can also be launching GPT‑5.6 Sol on Cerebras at as much as 750 tokens per second in July, bringing unprecedented inference pace to pick enterprise prospects.
The managed rollout of GPT‑5.6 comes after the Trump administration formally requested that OpenAI stagger the general public launch, citing the mannequin’s superior capabilities and nationwide safety implications.
In response to reporting, the request got here from the Workplace of the Nationwide Cyber Director and the Workplace of Science and Know-how Coverage, with Commerce Secretary Howard Lutnick additionally reportedly advising towards an unrestricted launch.
OpenAI CEO Sam Altman knowledgeable staff that the federal government could be “approving entry buyer by buyer” in the course of the preview interval. OpenAI has been specific, nevertheless, that this association shouldn’t be its most well-liked long-term mannequin, stating: “We don’t consider this type of authorities entry course of ought to turn into the long-term default. It retains one of the best instruments from customers, builders, enterprises, cyber defenders, and world companions who want them.”
Through the preview interval, GPT‑5.6 fashions will probably be accessible through the API and Codex to pick trusted companions, with broader availability throughout ChatGPT, Codex, and the API deliberate for the approaching weeks.
GPT-5.6 Sol’s Cybersecurity Capabilities
GPT‑5.6 Sol is OpenAI’s most succesful mannequin for long-horizon safety duties, together with vulnerability analysis and exploitation, considerably shifting the performance-efficiency frontier.
On ExploitBench, Sol is aggressive with Anthropic’s Mythos Preview whereas utilizing solely roughly one-third of the output tokens. On ExploitGym, a benchmark developed by UC Berkeley researchers in collaboration with OpenAI and different frontier labs, Sol, Terra, and Luna all show vital enhancements in cyber capabilities as reasoning effort will increase.
For coding workflows, GPT‑5.6 Sol units a brand new state-of-the-art on Terminal-Bench 2.1, scoring 88.8%, with the Sol Extremely configuration pushing the benchmark to 91.9%, outperforming Claude Mythos 5 (84.3%) and GPT‑5.5 (88.0%). The mannequin additionally introduces a brand new extremely mode that leverages subagents to speed up advanced multi-step work past the aptitude of any single agent.
Layered Cyberattack Protections
GPT‑5.6 Sol launches with OpenAI’s most sturdy security stack so far, particularly designed to face up to real-world adversarial stress. The protection structure is multi-layered:
- Mannequin-level safeguards skilled to refuse prohibited cyber help, together with jailbreak makes an attempt
- Actual-time misuse classifiers that consider output as it’s generated, with the power to pause technology and route flagged content material to a bigger reasoning mannequin for assessment
- Account-level assessment that analyzes indicators throughout conversations to tell apart persistent malicious habits from reputable dual-use safety analysis
- Differentiated entry controls that protect entry for reputable defensive work with out making essentially the most delicate capabilities broadly accessible
- Automated red-teaming utilizing over 700,000 A100-equivalent GPU hours devoted to discovering common jailbreaks and hardening the safeguard stack
- Human professional red-teaming by third-party testers, persevering with by way of the preview interval
Critically, GPT‑5.6 Sol doesn’t cross the Cyber Crucial threshold below OpenAI’s Preparedness Framework. In evaluations involving Chromium and Firefox, it recognized bugs and exploitation primitives however didn’t autonomously produce a practical full-chain exploit below examined situations.
OpenAI’s said purpose is to make “prohibited offensive exercise tougher, unsure, and detectable” with out limiting reputable use instances resembling code assessment, vulnerability analysis, patch improvement, and defensive safety testing.
GPT‑5.6 introduces a max reasoning effort mode, permitting Sol the utmost time to motive deeply earlier than producing output. The brand new extremely mode goes past a single-agent structure, leveraging subagents in parallel to speed up advanced, long-horizon duties.
On GeneBench v1, which evaluates long-horizon genomics and quantitative biology analyses, Sol achieves stronger outcomes than GPT‑5.5 whereas utilizing fewer tokens, demonstrating effectivity good points alongside uncooked functionality enhancements.
Previous to the June 26, 2026, launch, The Data first reported that OpenAI had agreed to stagger the GPT‑5.6 release after the Trump administration formally requested managed entry, citing nationwide safety considerations.
The transfer adopted a high-profile incident on June 12, 2026, when the administration issued export management directives compelling Anthropic to take its Fable 5 and Mythos 5 fashions completely offline to stop entry by international nationals, setting a crucial precedent for frontier AI governance.
Shortly earlier than the GPT‑5.6 collection launch, OpenAI officially released GPT‑5.5‑Cyber, a specialised mannequin engineered for automated vulnerability detection, patch technology, and remediation at machine pace, as a part of its broader Dawn initiative.
The mannequin delivered state-of-the-art outcomes throughout three main benchmarks: CyberGym (85.6%), ExploitGym (39.5%), and SEC-bench Professional (69.8%) all surpassing GPT‑5.5 baselines.
🔒 CISO / Safety Chief: Your Subsequent Breach Could Not Have a Face: Be a part of the “” LIVE webinar with ISC2