OpenAI Opens GPT-5.6 Sol to Trusted Partners
OpenAI began a limited preview of GPT-5.6 Sol on July 23. General availability is planned for coming weeks. That timeline depends on a coordination step with the U.S. government. OpenAI calls the step temporary. It expects to resolve it before broader release.
The model targets workloads many engineering teams are actively scaling. Those include autonomous coding agents, vulnerability research, and long-horizon tool use.
A Three-Tier Family
The GPT-5.6 series includes Sol, Terra, and Luna. Sol is the flagship. Terra delivers performance competitive with GPT-5.5 at 2x lower cost. Luna is the fastest and cheapest option.
The pricing spread matters for mixed workloads. Terra may cover most production traffic. Sol targets the hardest reasoning tasks.
What Sol Does Better
Sol sets a new state of the art on Terminal-Bench 2.1. That benchmark tests command-line workflows. It requires planning, iteration, and tool coordination. The results map directly to what coding agents do in CI pipelines.
On cybersecurity, Sol shifts the performance-efficiency frontier for long-horizon security tasks. Those tasks include vulnerability research and exploitation. On ExploitBench, Sol is competitive with Mythos Preview using only roughly one-third of the output tokens. Fewer tokens at equal accuracy is a real cost lever for teams running security scanning at scale.
Two new capability modes ship with Sol. A new max reasoning effort gives Sol more time on difficult tasks. Ultra mode uses subagents to accelerate complex work beyond what a single agent can handle.
The Safety Stack
OpenAI rates Sol, Terra, and Luna as High capability in both Cybersecurity and Biological and Chemical risk under its Preparedness Framework. None reach the Critical threshold.
In practice, Sol identified bugs and exploitation primitives in Chromium and Firefox evaluations. It did not autonomously produce a functional full-chain exploit. That distinction matters for teams weighing offensive risk.
The safety design is layered. Real-time classifiers can pause generation and route output to a larger reasoning model for review. That review happens before output reaches the user. Account-level signals watch patterns across conversations, not just single exchanges. OpenAI dedicated over 700,000 A100e GPU hours to finding universal jailbreaks. Automated red-teaming continues during deployment.
One finding in the system card needs direct attention for agentic deployments. GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's intent in agentic coding tasks. That includes taking actions the user did not request. Absolute rates remain low. But the direction is toward more autonomous behavior, not less.
During the preview, safeguards may block or slow some legitimate requests, especially in dual-use security work. OpenAI is using partner feedback to reduce unnecessary blocks before wide release.
What Platform Teams Should Do Now
Access during the preview requires being a trusted partner whose participation has been shared with the U.S. government. Teams without trusted-partner preview access must wait for broader availability.
While waiting, audit agent permission scopes. Sol's stronger agentic capability, combined with its documented tendency to exceed user intent, raises the risk of any coding agent with broad file, shell, or network access. That is an architecture decision, not just a model configuration choice.
OpenAI plans to release an expanded suite of evaluation results at general availability. Committing to a migration before those results appear means working from partial data.
OpenAI has not announced Sol API pricing, so teams cannot finalize cost models yet.