GPT-5.6 Sol is a new flagship artificial intelligence model launched in a limited preview alongside two other models named Terra and Luna.
The developer stated that these models offer different options for speed, cost, and intelligence. Meanwhile, the initial release is restricted to a small group of trusted partners.
This phased rollout follows coordination with the United States government to test capabilities before a wider release. Consequently, the company is using this preview period to gather feedback from users. They plan to make the models generally available in the coming weeks.
Capabilities of GPT-5.6 Sol
The new GPT-5.6 Sol model demonstrates improved capabilities in coding, biology, and cybersecurity tasks.
Specifically, it introduces a max reasoning effort mode to allow deeper processing. In addition, an ultra mode uses subagents to accelerate complex work.
On the Terminal-Bench 2.1 benchmark, the model achieved a score of 88.8% for command-line workflows. Furthermore, the ultra version scored 91.9% on the same test. These results show strong planning and tool coordination.
In biology, the model achieved stronger results on GeneBench v1 than previous versions while using fewer tokens. Meanwhile, in cybersecurity evaluations, it identified bugs in Chromium and Firefox. However, it did not autonomously produce a functional full-chain exploit.
Layered Safety and Safeguards
The company implemented a layered safety stack to prevent misuse of the artificial intelligence models. For instance, the system uses real-time classifiers to evaluate outputs during generation. If a potential violation is detected, the process pauses for review.
These safeguards are designed to support legitimate defensive work like vulnerability research and debugging. Nevertheless, users may experience occasional blocks or delays during the preview. The company aims to refine these controls based on user feedback.
Automated Red Teaming Methods
To improve safety, the developer dedicated over 700,000 A100-equivalent GPU hours to automated red teaming. This process uses models to find weaknesses and universal jailbreaks. Additionally, third-party experts are conducting human red teaming.
A rapid-response process is in place to remediate newly discovered vulnerabilities. Consequently, these findings are added to ongoing evaluations. This approach helps the system adapt to changing attack tactics.
Pricing and Availability Details
The models are priced per million tokens. Specifically, GPT-5.6 Sol costs $5 for input and $30 for output. Meanwhile, Terra costs $2.50 for input and $15 for output, and Luna costs $1 for input and $6 for output. These models are available via API and apps like ChatGPT.
The company also plans to launch the flagship model on Cerebras in July 2026. This deployment will deliver up to 750 tokens per second. Initially, access to this high-speed option will be limited to select customers.
