According to OpenAI, GPT-5.6 Sol is its most capable model so far. The company highlighted improvements in long-horizon agentic tasks across coding, biology, and cybersecurity, while introducing two new operating modes. A new max reasoning effort setting gives Sol additional time to work through complex problems, while an ultra mode uses multiple subagents to accelerate more demanding workflows.
OpenAI also released benchmark results intended to demonstrate the models’ capabilities. The company said GPT-5.6 Sol achieved a new state-of-the-art score on Terminal-Bench 2.1 for command-line coding tasks, showed stronger performance than GPT-5.5 on GeneBench v1 while using fewer output tokens, and improved cybersecurity performance on ExploitBench and ExploitGym as reasoning effort increased.
Much of the announcement focused on safety rather than raw performance. OpenAI said GPT-5.6 Sol launches with “our most robust safety stack to date,” describing stronger protections for higher-risk activity, sensitive cyber requests, and repeated misuse. The company said it spent several weeks identifying weaknesses, testing the models against real-world attack scenarios, and strengthening safeguards before the preview.
OpenAI said the models were developed with multiple layers of protection rather than relying on a single defense. Those safeguards include behavior trained directly into the models, real-time systems that evaluate outputs as they are generated, account-level monitoring for repeated misuse, differentiated access controls, and ongoing testing throughout the preview period.
The company acknowledged that those protections may occasionally interfere with legitimate work. During the preview, some requests may be blocked or delayed while additional review takes place, particularly in areas where defensive and offensive cybersecurity activities can appear similar. OpenAI said one objective of the limited rollout is to refine those safeguards and reduce unnecessary interruptions before wider availability.
For cybersecurity specifically, OpenAI said GPT-5.6 Sol is intended to be more effective at helping users identify and remediate vulnerabilities than at carrying out complete attacks. The company also said the model does not exceed the Cyber Critical threshold under its Preparedness Framework, noting that while it identified exploitation components during testing, it did not autonomously generate a complete exploit chain under the evaluated conditions.
To strengthen those protections, OpenAI said it dedicated more than 700,000 A100-equivalent GPU hours to automated red teaming designed to uncover broad jailbreak techniques. The company also worked with third-party security experts and said human red-teaming will continue throughout the preview period as additional weaknesses are identified and addressed.
The rollout begins through the API and Codex for a select group of trusted partners and organizations. OpenAI said it plans to expand availability to ChatGPT, Codex, and API users in the coming weeks.
OpenAI said the limited rollout follows discussions with the U.S. government before launch. At the government’s request, the company is beginning with a restricted preview before expanding access more broadly. “We don’t believe this kind of government access process should become the long-term default,” OpenAI said. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”
The company added that it views the current approach as a temporary measure while working with the Administration on “the cyber Executive Order framework and a repeatable process for future model releases.”
OpenAI also introduced a new naming system with GPT-5.6. The numerical version identifies the model generation, while the Sol, Terra, and Luna names represent long-term capability tiers that can evolve independently over time. Pricing is set per one million tokens, with Sol priced at $5 for input and $30 for output, Terra at $2.50 for input and $15 for output, and Luna at $1 for input and $6 for output. The company also announced updated prompt caching features and said GPT-5.6 Sol will launch on Cerebras in July, with access initially limited to select customers as capacity expands.
This analysis is based on reporting from OpenAI.
Image courtesy of OpenAI.
This article was generated with AI assistance and reviewed for accuracy and quality.
