OpenAI has announced a limited preview of its GPT-5.6 model series, introducing three tiers: Sol, Terra, and Luna, with Sol positioned as the flagship model featuring the company’s most advanced cybersecurity capabilities and its most robust safety architecture to date.
The release, disclosed on June 26, 2026, shows that Terra delivers performance competitive with GPT-5.5 at half the cost, while Luna offers strong capabilities at the lowest price point in the lineup. GPT-5.6 Sol represents a significant leap in AI-assisted security work.
On ExploitBench, the model performs competitively with Anthropic’s Mythos Preview while consuming only about one-third as many output tokens, a major efficiency gain for long-horizon security tasks, including vulnerability research and exploitation analysis.
On ExploitGym, a benchmark developed by UC Berkeley researchers in collaboration with OpenAI and other frontier AI labs, all three GPT-5.6 models demonstrate measurable improvements in cyber task performance as reasoning depth increases.
GPT-5.6 Sol With Stronger Cyber Capabilities
The model also achieves state-of-the-art results on TerminalBench 2.1, scoring 88.8% on command-line workflows requiring planning, iteration, and tool coordination, with Sol Ultra pushing that figure further to 91.9%.
OpenAI also introduced a new ultra mode that goes beyond single-agent capabilities by leveraging subagents to accelerate complex, long-horizon work.
GPT-5.6 Sol additionally shows strong improvements in biology workflows on GeneBench v1, achieving better results than GPT-5.5 while using fewer output tokens.
OpenAI clarified that GPT-5.6 Sol does not cross the “Cyber Critical” threshold under its Preparedness Framework. While testing against Chromium and Firefox revealed the model could identify bugs and exploitation primitives, the building blocks of an exploit, it did not autonomously produce a functional full-chain exploit under controlled conditions.
Recognizing that advanced cyber capability carries dual-use risk, OpenAI deployed a multi-layer safeguard stack specifically hardened for GPT-5.6.
At the model level, GPT-5.6 Sol is trained to refuse prohibited cyber assistance, including attempts to disguise intent or jailbreak the system.
Real-time misuse classifiers provide a second layer, evaluating output as it is generated; for higher-risk cases, generation is paused while a larger reasoning model reviews the full conversation context before deciding whether to release or withhold the output.
Account-level behavioral analysis adds a third dimension, enabling OpenAI’s systems to distinguish persistent malicious patterns from legitimate dual-use security research, recognizing that offensive and defensive work can initially appear similar in context.
Differentiated access controls further ensure that the most sensitive capabilities are not broadly available by default, preserving access for code review, vulnerability research, patch development, and defensive testing while restricting prohibited offensive activity.
To pressure-test these protections, OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming focused on universal jailbreak attacks that generalize across prompts and contexts, rather than on exploiting narrow, single-case vulnerabilities.
Human expert red-teaming continues in parallel throughout the preview period, complementing automated testing with creative adversarial scenarios that automated systems may not anticipate.
In an unusual step, OpenAI coordinated with the U.S. government ahead of the launch, sharing model capabilities and release plans in advance.
At the government’s request, GPT-5.6 is launching first to a limited group of trusted partners whose participation has been disclosed to federal authorities, as OpenAI works alongside the Administration to develop a cyber Executive Order framework and a repeatable process for future model releases.
OpenAI stated it does not intend for this process to become a long-term standard, noting it restricts access for defenders, developers, and global partners who urgently need these tools.
GPT-5.6 is currently available via API and Codex to a select group of trusted partners, with broader access across ChatGPT and the API expected within weeks.
Sol is priced at $5 input / $30 output per 1M tokens, Terra at $2.50 input / $15 output, and Luna at $1 input / $6 output. Cache reads receive a 90% discount, while cache writes are billed at 1.25x the uncached input rate with a 30-minute minimum cache life.
OpenAI also announced a partnership with Cerebras, targeting up to 750 tokens per second for Sol beginning in July 2026.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
