OpenAI has launched limited partner previews of its GPT-5.6 Sol flagship model, the lead release in the new GPT-5.6 family, with general public availability planned for the coming weeks OpenAI.
OpenAI shared details of the GPT-5.6 line and its capabilities with U.S. government stakeholders ahead of the limited partner preview launch OpenAI.
The GPT-5.6 line includes three distinct models built for different performance and cost profiles OpenAI.
Sol is the high-capability flagship designed for complex, multi-step agentic work. Terra offers a balanced performance profile for general use cases.
Luna is a low-latency, low-cost option that OpenAI describes as fast and affordable for high-volume, simple tasks OpenAI.
GPT-5.6 Sol introduces new deep reasoning and subagent capabilities
The flagship GPT-5.6 Sol model introduces two new operating modes for complex agentic tasks OpenAI.
The max reasoning effort setting allocates extra compute cycles for deep, step-by-step problem solving on high-complexity requests. The new ultra mode leverages specialized subagents to complete multi-step end-to-end workflows without manual intervention for each step OpenAI.
Benchmark leads in coding, biology and cybersecurity tasks
OpenAI released initial evaluation results to preview GPT-5.6 Sol’s performance across high-impact agentic workflows OpenAI.
For coding and command-line tasks, Sol sets a new state-of-the-art result on Terminal-Bench 2.1, a benchmark that tests end-to-end command-line workflows requiring multi-step planning, iterative debugging and external tool coordination OpenAI.
In quantitative biology testing, Sol outperforms the prior GPT-5.5 model on GeneBench v1, a genomics and long-horizon biology analysis benchmark, while using fewer total tokens per completed task OpenAI.
For cybersecurity use cases, Sol matches the performance of Mythos Preview on the ExploitBench benchmark using roughly one-third of the output tokens required by the prior top model. It also delivers improved results on ExploitGym, a cybersecurity evaluation benchmark created by a UC Berkeley research team in partnership with OpenAI and other leading frontier AI labs, with performance gains increasing as reasoning effort is raised OpenAI.
Who can access the GPT-5.6 preview right now?
The initial GPT-5.6 preview is restricted to a small cohort of vetted trusted partners OpenAI.
OpenAI stated it does not intend for the government access review process used for this preview to become a long-term default for future model releases OpenAI.
What safeguards are in place for GPT-5.6 Sol?
All three GPT-5.6 models ship with safeguard configurations tailored to each model’s specific capability level and associated risk profile OpenAI.
For the high-capability GPT-5.6 Sol model, safeguards include model-level refusals of requests for prohibited cyber assistance, including attempts to disguise malicious intent or jailbreak the model, and account-level access controls OpenAI.
OpenAI noted Sol does not cross its Cyber Critical threshold under the company’s official Preparedness Framework. In controlled testing using Chromium and Firefox browsers, the model identified software bugs and exploitation primitives, but did not autonomously produce a functional full-chain exploit under tested conditions OpenAI.
The company’s stated goal for these safeguards is to make prohibited offensive activity more difficult, uncertain and detectable, while preserving full access for legitimate defensive work including vulnerability research, patch development and security education OpenAI.
Bottom line: Users seeking advanced agentic, coding or cybersecurity AI capabilities can plan to access the full GPT-5.6 series, including the GPT-5.6 Sol flagship with its new max and ultra reasoning modes, in the coming weeks. Sol’s verified benchmark results include a new state-of-the-art score on Terminal-Bench 2.1 for end-to-end command-line workflows, outperformance of GPT-5.5 on GeneBench v1 with lower per-task token usage, and parity with the prior top cybersecurity model Mythos Preview on ExploitBench using roughly one-third the output tokens, all paired with tiered safeguards that block prohibited offensive use while preserving access for legitimate defensive security and research work OpenAI.
