After Anthropic's Mythos and Fable have dominated the AI headlines in recent weeks, OpenAI has come back to the party.
On Friday, the company debuted a limited preview of GPT-5.6, the latest iteration of its flagship model. The company unveiled several versions of this model for different levels of tasks, and intends to make them generally available in the coming weeks.
The company said it previewed its plans and the models' capabilities ahead of launch as part of its "our ongoing engagement with the US government," which is why it has rolled out the model in limited preview to a small group of trusted partners. OpenAI said it will continue testing the model during this period before releasing it to the general public.
GPT-5.6 has three different classes: Sol, Terra, and Luna.
Sol is the "flagship model," representing its strongest iteration yet and proving strong agentic capabilities in coding, biology, and cybersecurity. Sol also includes a new "ultra" mode, which leverages subagents for complex work. As for cost, Sol is $5 per million input tokens and $30 per million output tokens.Terra is a "balanced model" for everyday work, featuring competitive performance to GPT-5.5 while being half the cost. Terra is $2.50 per million input tokens and $15 per million output tokens.Luna, meanwhile,is the cheapest and fastest model in the stack, coming in at just $1 per million input tokens and $6 per million output tokens.
According to TerminalBench 2.1, which tests workflows requiring planning, iteration, and tool coordination, both Sol and its Ultra mode surpass Claude Mythos 5. Terra surpasses Claude Fable 5 on the same benchmark, while Luna surpasses Opus 4.8.
Notably, OpenAI strongly emphasized the model family's cyber capabilities and security safeguards in its announcement, noting that all three models are fitted with its "most robust safeguards to date," and configurations that match each model's capabilities. The company said that, as the model becomes more capable, it's designed to hold up against adversarial pressures, while still handling cyber work such as code review, defensive testing, patch development and debugging.
"Our goal is to make prohibited offensive activity more difficult, uncertain, and detectable without unnecessarily limiting those beneficial uses," OpenAI said in its announcement.
Our Deeper View #
OpenAI's latest addition to the frontier model race represents the latest hop in these companies' trillion-dollar game of leapfrog. And with Anthropic's long-awaited release of Mythos and Fable now muzzled by the US government due to the commerce department's national security export controls, OpenAI is effectively in a league of its own for now. Though the company is playing it safe, only releasing these models in a limited preview to not stoke the government's ire, the release still gives it a leg up. The fact that GPT-5.6 is outperforming Anthropic's Mythos-class models in several benchmarks definitely doesn't hurt either. However, the bigger question remains: As these companies continue to outdo one another with the latest frontier capabilities, will these upgrades actually make a noticeable difference to users? And, while these models take turns inching past one another, at what point will they become commoditized?