OpenAI Launches Limited Preview of GPT-5.6 Sol Model
OpenAI has launched a limited partner preview of its GPT-5.6 series, led by the flagship Sol model. General availability for all three series entries is planned for the coming weeks. The initial preview is restricted to a small cohort of trusted partners.
OpenAI’s official GPT-5.6 Sol preview announcement
OpenAI emphasized it does not want this restricted access process to become a long-term default for new model rollouts. The company notes the limited preview currently restricts access for developers, enterprise users, and cyber defenders who need the tools most.
OpenAI’s official GPT-5.6 Sol preview announcement
GPT-5.6 Sol Delivers Agentic Performance Gains Across Key Benchmarks
OpenAI shared initial evaluation data highlighting Sol’s improved agentic capabilities for long-horizon, tool-using workflows that require planning, iteration, and cross-step coordination. For coding tasks, the model achieves the highest publicly reported score to date on the Terminal-Bench 2.1 benchmark, which tests command-line workflows that require multi-step tool use.
OpenAI’s official GPT-5.6 Sol preview announcement
On GeneBench v1, a benchmark that evaluates genomics and quantitative biology analyses, Sol outperforms GPT-5.5 while using fewer tokens per task, per OpenAI’s testing.
OpenAI’s official GPT-5.6 Sol preview announcement
For cybersecurity workloads, Sol shifts the performance-efficiency frontier for long-horizon vulnerability research and exploitation work. On ExploitBench, it is competitive with Mythos Preview while using only roughly one-third of the output tokens required by the competing model.
OpenAI’s official GPT-5.6 Sol preview announcement
All three GPT-5.6 models demonstrate measurable gains in cyber task performance as reasoning effort scales on ExploitGym, a benchmark built by a UC Berkeley research group working alongside OpenAI and other leading frontier AI development teams.
OpenAI’s official GPT-5.6 Sol preview announcement
New Reasoning Modes for GPT-5.6 Sol Long-Horizon Workflows
The GPT-5.6 series introduces two new reasoning effort options. The max setting gives the model additional time to reason deeply for complex, high-stakes tasks. The ultra mode leverages subagents to accelerate multi-step work.
OpenAI’s official GPT-5.6 Sol preview announcement
Layered Safeguards Tailored to GPT-5.6 Sol Capability Levels
OpenAI built the GPT-5.6 series with layered, model-specific safeguards designed to enable legitimate defensive work such as vulnerability research, patch development, and security education. The safeguards also make prohibited offensive activity more difficult and detectable.
OpenAI’s official GPT-5.6 Sol preview announcement
OpenAI’s internal testing found the Sol model performs more consistently at identifying and remediating security flaws than it does at executing full, end-to-end cyber exploitation workflows. The model does not meet the Cyber Critical risk threshold defined in the company’s Preparedness Framework.
OpenAI’s official GPT-5.6 Sol preview announcement
In tests with Chromium and Firefox, Sol identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit under tested conditions, per OpenAI’s internal safety evaluations.
OpenAI’s official GPT-5.6 Sol preview announcement
The safety stack includes model-level refusals for prohibited cyber assistance, plus real-time cyber and biology misuse classifiers that evaluate output as it is generated. For high-risk cases, the system pauses for review by a larger reasoning model before disallowed output is delivered to the user.
OpenAI’s official GPT-5.6 Sol preview announcement
Exact safeguard configurations vary across the GPT-5.6 series models to match their differing capability levels. The phased preview approach is paired with ongoing safeguard pressure-testing to ensure the model’s increased capabilities are matched to appropriate access controls as it scales to broader users.
OpenAI’s official GPT-5.6 Sol preview announcement
Full GPT-5.6 Series Lineup and General Availability Timeline
Alongside the flagship Sol, the GPT-5.6 series includes two additional models. Terra is a balanced everyday model with performance competitive to GPT-5.5 at half the cost, per OpenAI’s announcement.
OpenAI’s official GPT-5.6 Sol preview announcement
All three models will move to general availability in the coming weeks following the limited preview period. During this preview window, OpenAI will continue coordinating with partners to refine safeguards and performance ahead of broader release.
OpenAI’s official GPT-5.6 Sol preview announcement
