— We don’t believe this kind of government access process should become the long-term default, writes OpenAI, and says — we [are working] with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.
In its presentation, OpenAI details three models in the GPT-5.6 series. Sol is the state-of-the-art Mythos-beating model, while Terra is «a balanced model for everyday work,» and delivers on par with GPT-5.5 at half the cost, and Luna is the cost-focused model «delivering stronger capability at our lowest cost.»
Strong on cyber
The models, particularly Sol, are very strong on cyber defense, and are able to identify security vulnerabilities, but not able to execute automated attacks due to strong safeguards, built using 700,000 A100-equivalent GPU hours of red-teaming.
Sol beats Mythos 5 on coding and cybersecurity and uses fewer tokens for better performance on biology tasks. It also comes with a new max mode, which gives more time for better reasoning, and an ultra mode which uses subagents for more complex work.
Prices for Sol are at $5 input/$30 output, for Terra is $2.50 input/$15 output, and Luna is at $1 input/$6 output, and general availability is up to higher powers.
Read more: OpenAI’s presentation, launch thread, Axios, Reuters, and TechCrunch. Discussion on r/Singularity.