Technical document describes the Sol version as a significant advancement over the previous generation and introduces new safety layers against misuse.
The system card for GPT-5.6, OpenAI's upcoming model, indicates that the main version, called Sol, represents a significant leap in capability compared to GPT-5.5. The developer itself classified the performance jump as a "step function" improvement. Despite this progress, analyses suggest the model still falls short of the internal Mythos project. The technical documentation emerges as the primary source of information on the model as the company prepares for a broader release.
The new generation is not limited to a single model, forming a family designed to offer different balances between intelligence, speed, and cost. The Sol version will maintain the previous generation's pricing, costing $5 and $30 for input and output, respectively. For developers seeking savings, the Terra version promises performance similar to GPT-5.5 at half the price ($2.50 and $15), while Luna is presented as the cheapest option in the portfolio ($1 and $6).
There is also a focus on processing speed. OpenAI indicated that GPT-5.6 could operate at 750 tokens per second on Cerebras infrastructure, a considerably high rate. However, the company warned that processing capacity in this format will be initially limited and did not detail the costs associated with this specific operation.
Regarding features, the model will introduce new levels of reasoning. In addition to an enhanced configuration called "Max," OpenAI detailed an "Ultra" mode. According to the documentation, this new level will allow GPT-5.6 to generate subagents to handle tasks autonomously, expanding its ability to execute complex commands.
To curb misuse in sensitive areas such as biology and cybersecurity, OpenAI will adopt a defense-in-depth strategy. The system will feature multiple layers of protection, including safeguards trained within the model itself, real-time checks during response generation, and account-level monitoring. The company emphasizes that no single barrier is sufficient against adaptive attacks, justifying the need for an integrated set of filters that will be tested during the model's preview phase.
The GPT-5.6 family includes three versions: Sol, which offers a significant capability leap maintaining previous pricing ($5 input/$30 output); Terra, providing GPT-5.5 performance at half the price ($2.50/$15); and Luna, the cheapest option ($1/$6).
The Ultra mode introduces new reasoning levels that allow GPT-5.6 to generate subagents. These subagents operate autonomously to handle tasks, significantly expanding the model's ability to execute complex commands.
OpenAI uses a defense-in-depth strategy against misuse in biology and cybersecurity. This includes safeguards trained within the model, real-time response checks, and account-level monitoring to ensure adaptive attacks are blocked by multiple layers of protection.