OpenAI Launches GPT-5.6 Models to Government Partners Amidst Stricter Safety and Regulatory Scrutiny

GPT-5.6 Sol cleared 96.7% of internal cyberattack test challenges, and access to the new models is being governed by a government-assisted process with no public waitlist, no self-service enrollment, and no confirmed date for broad availability.
Sol launches with OpenAI's most robust safety stack to date, including strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse; the company also notes that all three GPT-5.6 models are classified at the 'High' capability level for cybersecurity and vulnerability risk under its Preparedness Framework.
On Terminal-Bench 2.1, Sol scores 91.9%—higher than Anthropic’s Mythos at 88%—highlighting a lead in advanced software engineering tasks within OpenAI’s evaluation suite.
In biology-focused benchmarks, Sol outperforms GPT-5.5 on GeneBench v1, achieving stronger results while using fewer tokens, signaling improved efficiency on long-horizon genomics analyses.
Pricing and performance positioning show Terra delivering GPT-5.5-level results at roughly half the cost, while Luna targets high-volume, low-cost usage, with Luna priced around $1 per million input tokens.
OpenAI has unveiled GPT-5.6, a three-tier model family led by Sol, its most powerful AI yet. Rather than a public launch, the company handed access to roughly 20 government-approved partners — a historic first for the AI industry, according to Forbes and OpenAI.
Sol cleared 96.7% of internal cyberattack test challenges, according to MLQ.ai. That score — and the model's ability to find vulnerabilities in hardened software like Chromium — convinced the Trump administration to treat it like a dual-use weapon, not a consumer product.
On June 2, 2026, President Trump signed an Executive Order creating a new category: "covered frontier models." Any AI that crosses high thresholds in cybersecurity or biology must pass a federal vetting process before going public, according to MLQ.ai. GPT-5.6 Sol hit those thresholds.
The White House Office of the National Cyber Director asked OpenAI to delay its broad release. OpenAI agreed. The company said it believes in "broad access" but is "starting with a limited preview" at the government's request, adding that "this should not become the long-term default," according to TechTimes. There is no public waitlist and no confirmed date for wider access.
On Terminal-Bench 2.1, Sol scored 91.9% — beating Anthropic's rival Mythos model at 88%, according to The PC Enthusiast. That test measures advanced software engineering. Sol ran in "Ultra" mode, which uses internal sub-agents to break big tasks into smaller ones and solve them in parallel.
In biology, Sol scored 68.3% on "World-Class Bio" benchmarks — a 9% jump over GPT-5.5 — according to Pulse2. It also beat GPT-5.5 on GeneBench v1 while using fewer tokens, meaning it gets better results with less computing power on long genomics tasks.
OpenAI built GPT-5.6 as a family. Sol is the flagship at $5 per million input tokens. Terra sits in the middle at $2.50 — offering GPT-5.5-level performance at roughly half the cost. Luna targets high-volume use at just $1 per million input tokens, according to QazInform.
All three models are rated "High" risk under OpenAI's Preparedness Framework for cybersecurity. OpenAI says Sol carries its strongest safety protections yet. The company spent 700,000 A100-equivalent GPU hours on automated red-teaming alone, according to TechMyMoney.
Not everyone is cheering the new framework. Tech liberty advocates say government-gated AI gives Fortune 500 firms and defense contractors a head start on agentic workflows while smaller developers are stuck with older models. Quartz noted the vetting process "keeps the best tools from global partners who need them most."
The precedent also raises questions about multinational companies trying to access Sol outside the US. Broad availability is "hoped for" in the coming weeks but remains under federal review through August 2026, according to TechTimes. Independent real-world testing — not just OpenAI's own benchmarks — will be needed to confirm whether Sol's gains hold up in practice.
Publishers
15
Articles
5
Reach
20