OpenAI has launched GPT-5.6, a new family of AI models comprising Sol, Terra, and Luna, and said it will begin with a limited preview for a select group of trusted partners before a broader rollout in the coming weeks.
The company said GPT-5.6 Sol is its flagship model, Terra is a lower-cost model with performance comparable to GPT-5.5, and Luna is its fastest and most affordable offering.
“Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost,” the company said in a statement.
The models will initially be available via the API and Codex, with access to ChatGPT planned later.
During the preview, pricing starts at $5 per million input tokens and $30 per million output tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. The company also said GPT-5.6 Sol will launch on Cerebras in July with speeds of up to 750 tokens per second for select customers.
OpenAI said the limited rollout follows discussions with the US government. “As part of our ongoing engagement with the US government, we previewed our plans and the models’ capabilities ahead of today’s launch,” the company said. “At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly.”
The company added that it does not want such a process to become standard. “We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them,” it said.
“We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, while we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases,” OpenAI added.
Anthropic Gets Green Light
Following OpenAI’s announcement, Anthropic on June 27 said it has started restoring access to its Claude Mythos 5 cybersecurity model for a limited group of US organisations after working with the US government.
“The government notified us that Mythos 5, our strongest cybersecurity model, can be redeployed to a set of US organisations that operate and defend critical infrastructure,” the company wrote on X.
The company added that it is continuing to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again.
The announcement follows restrictions imposed earlier this month that limited access to the two models. Anthropic did not disclose which organisations would regain access or provide a timeline for when broader access to Mythos 5 or public availability of Fable 5 could resume.
What GPT-5.6 Can Do
OpenAI said GPT-5.6 introduces a new ‘max reasoning effort’ mode and an ‘ultra mode’ that uses subagents to tackle complex tasks.
According to the company, GPT-5.6 Sol achieves state-of-the-art performance on TerminalBench 2.1 for coding workflows, with GPT-5.6 Sol Ultra scoring 91.9% to outperform Mythos 5 and Fable 5. OpenAI also said GPT-5.6 improved results on biology and cybersecurity benchmarks compared with GPT-5.5.
Moreover, OpenAI said that GPT-5.6 Sol delivered performance comparable to that of Mythos Preview on ExploitBench, using roughly one-third of the output tokens.
The company also highlighted changes to its safety measures, saying GPT-5.6 ships with its ‘most robust safeguards to date.’ OpenAI said it strengthened protections against higher-risk cyber activity through layered safeguards, automated red-teaming, account-level monitoring and real-time misuse detection.
“GPT-5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks,” OpenAI said.
It added that the model remains below the Cyber Critical risk threshold defined in its Preparedness Framework, but introduced additional safeguards because of its improved capabilities.
To improve the system’s resilience, OpenAI said it devoted more than 700,000 A100-equivalent GPU hours to automated red-teaming to identify jailbreaks and other weaknesses before release.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.