OpenAI has introduced GPT-5.6, its latest family of artificial intelligence models, featuring three variants—Sol, Terra, and Luna. The company said the release will begin with a limited preview for a small group of trusted partners before expanding access to more users over the coming weeks.

Among the new models, GPT-5.6 Sol is positioned as OpenAI’s most advanced offering. Terra is designed as a more affordable alternative that delivers performance comparable to GPT-5.5 at roughly half the cost, while Luna is the company’s fastest and lowest-priced model for developers seeking cost-efficient AI capabilities.

Initially, the GPT-5.6 lineup will be available through OpenAI’s API and Codex platform, with ChatGPT integration planned at a later stage.

For the preview phase, OpenAI has set pricing at $5 per million input tokens and $30 per million output tokens for Sol. Terra is priced at $2.50 for input tokens and $15 for output tokens, while Luna costs $1 and $6 per million input and output tokens, respectively. The company also confirmed that GPT-5.6 Sol will become available on Cerebras in July, offering processing speeds of up to 750 tokens per second for select customers.

Limited Launch Following US Government Consultation

OpenAI said the staggered rollout comes after discussions with the US government regarding the model’s capabilities. According to the company, government officials were briefed before the public announcement, leading to a decision to initially provide access only to a limited group of trusted partners whose participation had been shared with the administration.

Despite agreeing to the temporary arrangement, OpenAI emphasized that it does not want government review before public releases to become a permanent practice. The company argued that delaying access prevents developers, businesses, cybersecurity teams, and other users from benefiting from the latest AI technology.

OpenAI added that the limited preview is intended as a short-term measure while it works with the US administration on establishing a consistent framework for future AI model launches under the country’s cybersecurity executive order.

Anthropic Restores Partial Access to Mythos 5

Shortly after OpenAI’s announcement, Anthropic revealed that it had started restoring access to its Claude Mythos 5 cybersecurity model for a limited number of US organizations following discussions with the US government.

In a post on X, the company said the government had approved redeployment of Mythos 5 for organizations responsible for operating and defending critical infrastructure. Anthropic also stated that it continues to work with officials to broaden access to Mythos 5 and eventually restore general availability of its Fable 5 model.

Earlier this month, access to both models had been restricted. However, Anthropic has not disclosed which organizations will regain access or when wider availability is expected.

New Reasoning Features and Performance Improvements

OpenAI said GPT-5.6 introduces a new “max reasoning effort” mode along with an “ultra mode” that relies on multiple subagents to handle more demanding tasks.

According to the company, GPT-5.6 Sol achieved state-of-the-art results on TerminalBench 2.1, a benchmark focused on coding workflows. The enhanced Sol Ultra version recorded a score of 91.9%, surpassing Anthropic’s Mythos 5 and Fable 5 models. OpenAI also reported improved performance over GPT-5.5 in biology and cybersecurity evaluations.

The company further claimed that GPT-5.6 Sol matched the performance of Mythos Preview on ExploitBench while generating only about one-third of the output tokens, making it significantly more efficient.

Stronger AI Safety Measures

OpenAI said GPT-5.6 includes its strongest safety protections so far. The company strengthened its cybersecurity safeguards through multiple layers of defense, including automated red-teaming, continuous account monitoring, real-time misuse detection, and additional protections against high-risk cyber activities.

According to OpenAI, the model is more effective at helping users identify and fix software vulnerabilities than carrying out complete cyberattacks. Although GPT-5.6 remains below the Cyber Critical risk level defined in the company’s Preparedness Framework, OpenAI introduced additional safeguards because of the model’s increased capabilities.

To test the system before release, OpenAI said it invested more than 700,000 A100-equivalent GPU hours in automated red-teaming to uncover jailbreak techniques and other potential security weaknesses.

Also Read: OpenAI Poaches Transformer Pioneer Noam Shazeer From Google