OpenAI Releases GPT-5.6 Sol in Limited Preview With Stronger Cyber Safeguards

OpenAI has started a limited preview of GPT-5.6, led by its new flagship model, GPT-5.6 Sol. The new model family also includes Terra for balanced everyday work and Luna for faster, lower-cost tasks.

The rollout is limited for now. OpenAI says GPT-5.6 Sol, Terra, and Luna are first going to a small group of trusted partners whose participation has been shared with the U.S. government.

The company says broader access is planned in the coming weeks across ChatGPT, Codex, and the API. For users and developers, this means GPT-5.6 is available, but not yet in a normal public release.

GPT-5.6 Sol leads the new model family

GPT-5.6 introduces a new naming system. The number identifies the model generation, while Sol, Terra, and Luna identify capability tiers that can improve on separate schedules.

Sol is the flagship model. Terra is built for everyday professional work and is priced lower than GPT-5.5 for comparable performance. Luna is the fastest and cheapest option in the family.

Model	Role	Price per 1M tokens
GPT-5.6 Sol	Flagship model for advanced reasoning, coding, science, and cybersecurity	$5 input, $30 output
GPT-5.6 Terra	Balanced model for everyday work	$2.50 input, $15 output
GPT-5.6 Luna	Fast and affordable model for high-volume tasks	$1 input, $6 output

GPT-5.6 Sol will also launch on Cerebras in July at up to 750 tokens per second for select customers. OpenAI says access to that faster deployment will start with limited enterprise availability as capacity expands.

Why access is limited at launch

The limited rollout follows a U.S. government request tied to national security concerns. Reuters reported that OpenAI delayed a full public launch and shared details of the vetted partner group with authorities.

Axios reported that the White House Office of the National Cyber Director and the Office of Science and Technology Policy asked OpenAI to limit the initial release while a security testing framework is developed.

The request fits a broader federal push around frontier AI. The White House AI security order calls for cooperation between the federal government and AI developers on advanced systems with national security implications.

OpenAI says the access model should not become permanent

OpenAI says it previewed GPT-5.6 capabilities to the government before launch and agreed to a phased release as a temporary step. The company also says it does not want government customer approval to become the normal model for future AI releases.

The concern is practical. Restricting access too tightly could slow developers, enterprises, security teams, and global partners that need advanced models for defensive and productive work.

Still, OpenAI says it believes the limited preview gives it the strongest path to broader availability while the administration works on a repeatable process for future releases.

GPT-5.6 Sol brings stronger cyber capabilities

OpenAI describes Sol as its most capable model yet for cybersecurity. The company says it improves long-horizon security work, including vulnerability research, exploit analysis, patch development, code review, debugging, security education, and defensive testing.

The GPT-5.6 Preview System Card says OpenAI is treating Sol, Terra, and Luna as High capability in cybersecurity under its Preparedness Framework. However, OpenAI says none of the models cross the Cyber Critical threshold.

OpenAI says Sol performed strongly in evaluations involving Chromium and Firefox. It found bugs and exploitation primitives, but it did not autonomously produce a functional full-chain exploit under the tested conditions.

What the cyber benchmarks show

OpenAI says GPT-5.6 Sol is competitive with Mythos Preview on ExploitBench while using about one-third of the output tokens. It also reports improvements across the GPT-5.6 family on ExploitGym as reasoning effort increases.

ExploitGym is a benchmark focused on whether AI agents can turn known vulnerabilities into working exploits that create real security impact. The public paper describes exploitation as a hard, dual-use task that can help defenders but also lower barriers for attackers.

That dual-use nature explains why GPT-5.6 is launching with stronger restrictions and more safeguards. OpenAI says Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks.

Area	GPT-5.6 Sol update	Security meaning
Cybersecurity	Stronger long-horizon vulnerability research and exploitation analysis	More useful for defenders, but more sensitive to misuse
Coding	New state of the art on Terminal-Bench 2.1, according to OpenAI	Better handling of command-line workflows and multi-step development tasks
Biology	Improved GeneBench v1 results while using fewer tokens	Higher capability requires closer safety evaluation
Reasoning modes	New max reasoning effort and ultra mode with subagents	More capacity for complex, long-running tasks

OpenAI adds layered cyberattack protections

OpenAI says no single safeguard can stop determined misuse. For GPT-5.6, the company is using model-level refusals, real-time cyber and biology misuse classifiers, account-level review, differentiated access, monitoring, enforcement, and continued testing.

The system card also describes deployment simulation, expert testing, and evaluation work across several risk areas. OpenAI says GPT-5.6 Sol has about the same predicted rate of disallowed content violations as GPT-5.5 in its deployment simulation.

• Model-level safeguards refuse prohibited cyber assistance.
• Real-time classifiers can pause high-risk generations for additional review.
• Account-level review can identify repeated malicious behavior across conversations.
• Differentiated access keeps the most sensitive capability away from broad default use.
• Enterprise work may include privacy-preserving detection and customer-operated safety controls.

Automated and human red teaming shaped the release

OpenAI says it dedicated more than 700,000 A100-equivalent GPU hours to automated red teaming. The goal was to find universal jailbreaks that could work across many prompts or contexts, not only narrow test cases.

The company also used third-party human red teams and says that work will continue during the preview period. This matters because experts often find creative misuse patterns that automated testing can miss.

The preview period is also meant to test false blocks. OpenAI says safeguards may sometimes interrupt legitimate defensive work, especially where offensive and defensive security requests look similar at first.

GPT-5.6 follows OpenAI’s Daybreak cyber push

The release comes days after OpenAI expanded its Daybreak cyber defense initiative. That update included the full version of GPT-5.5-Cyber for trusted defenders, an updated Codex Security plugin, and partnerships aimed at faster vulnerability remediation.

OpenAI said GPT-5.5-Cyber reached 85.6% on CyberGym, 39.5% on ExploitGym, and 69.8% on SEC-bench Pro, beating GPT-5.5 baselines in those tests. The company positioned the model as a tool for authorized defenders who need more advanced cyber capability.

The newer GPT-5.6 release builds on the same trend. OpenAI wants powerful models to help defenders find and patch vulnerabilities, while making prohibited offensive use more difficult, uncertain, and detectable.

Anthropic’s Mythos case shaped the policy backdrop

GPT-5.6 also arrived shortly after a major Anthropic access restriction. In a June statement, Anthropic said the U.S. government issued an export control directive requiring suspension of Fable 5 and Mythos 5 access for foreign nationals.

Anthropic said it had to disable Fable 5 and Mythos 5 for all customers to ensure compliance. That incident raised the stakes for other frontier AI labs planning cyber-capable model releases.

Axios reported that officials viewed GPT-5.6 as having Mythos-like capability. Reuters also noted that U.S. officials want early access to frontier models to identify threats before wider deployment.

What developers and enterprises should know

During the preview, selected partners can use GPT-5.6 through the API and Codex. Broader access for ordinary ChatGPT, Codex, and API users is expected later, but OpenAI has not treated the first wave as a normal general release.

Pricing gives enterprises a clearer cost ladder. Sol targets the most demanding work, Terra aims for a balance between capability and cost, and Luna targets speed and affordability.

Developers should also expect more visible safety checks in sensitive domains. Some cyber requests may receive refusals, pauses, or extra review while OpenAI tunes safeguards during the preview.

Why GPT-5.6 matters for AI policy

GPT-5.6 shows how AI model launches are becoming part of national security policy. The AI security executive order sets a voluntary framework for early government access to covered frontier models, while companies continue to argue for broad access for legitimate users.

The challenge is finding the right balance. Models that can help defenders patch software faster can also support harmful cyber workflows if safeguards fail or access controls are weak.

The OpenAI preview makes that trade-off visible. GPT-5.6 Sol is more capable, more expensive than smaller tiers, and more tightly controlled at launch. It also shows why future model releases may involve closer coordination between AI labs, cyber agencies, enterprise customers, and independent testers.

For security teams, the practical takeaway is clear. GPT-5.6 Sol may become a major tool for defensive vulnerability research and patching, but access, monitoring, and governance will shape how quickly enterprises can use it at scale.

The broader OpenAI Daybreak effort and the Anthropic access suspension point to the same conclusion: frontier AI cyber capability is moving fast, and governments now want a larger role before the most powerful models reach the public.

Benchmarks such as ExploitGym will likely become more important as labs, regulators, and enterprises try to measure what these models can actually do in realistic cyber tasks.

FAQ