OpenAI Previews GPT-5.6 Family Under Government Restrictions Amid Security Benchmark Gains

OpenAI announced a limited preview of its GPT-5.6 model family, consisting of Sol, Terra, and Luna. [3][7] Sol is the flagship model designed for complex coding and security research. [3] Terra is intended for high-volume business tasks, offering competitive performance to GPT-5.5 at twice the cost reduction. [5][6] Luna is positioned as the fastest and most affordable model, performing near GPT-5.5 levels on several tests. [3][6]

The models introduce a tiered pricing structure per one million tokens. [6] Sol costs $5 for input and $30 for output, Terra costs $2.50 and $15, and Luna costs $1 and $6. [6] During the preview, the models are available through the OpenAI API and Codex, but they are not accessible in ChatGPT. [7] OpenAI plans to launch Sol on Cerebras hardware in July for select customers, enabling speeds up to 750 tokens per second. [6]

GPT-5.6 introduces new inference capabilities, including a "max" reasoning effort for deep inference and an "ultra" mode that uses subagents to parallelize complex workflows. [1][6] Sol sets a new state-of-the-art score on Terminal-Bench 2.1, an evaluation testing command-line workflows. [5] On the SecureBio evaluations, the model scored 68.3 percent on World-Class Bio, improving upon GPT-5.5 by roughly nine percentage points. [5]

In cybersecurity testing, Sol matched the performance of Anthropic's Claude Mythos Preview on ExploitBench, though other sources indicate it outperformed Mythos on benchmarks like TerminalBench. [1] According to OpenAI, the model achieved competitive results on cybersecurity benchmarks while using approximately one-third of the output tokens required by another leading frontier system. [5] Sol and Terra demonstrated the capability to identify vulnerabilities, but OpenAI stated the models were unable to execute autonomous, end-to-end attacks against hardened targets. [2]

OpenAI restricted the initial preview of GPT-5.6 to approximately 20 organizations after sharing its release plans with the U.S. government. [3][6] This staggered rollout follows a June 2, 2026, executive order issued by President Donald Trump, which directed federal agencies to benchmark and assess the capabilities of new AI models. [3] The order outlined a 30-day process scheduled to conclude on July 2. [3]

The U.S. government previously issued an export control order against OpenAI competitor Anthropic regarding jailbreaks discovered in its Claude Fable 5 model. [3] Following that order, Anthropic removed public and private access to both Claude Fable 5 and its cybersecurity-focused counterpart, Claude Mythos 5. [3] OpenAI stated it plans a broader release for GPT-5.6 in the coming weeks. [3][6]

OpenAI's three-tiered release signals a structural shift in enterprise AI toward explicit workload routing, allowing developers to balance latency and token costs natively without relying exclusively on third-party orchestrators. The government-mandated limited preview underscores an increasingly aggressive regulatory environment for frontier models, especially as architectures push the boundaries on autonomous cybersecurity exploits and biological synthesis. This cautious, federally vetted launch stands in direct contrast to Anthropic's recent forced rollback of Claude Fable 5 and Mythos 5 due to export controls. By offering Sol on Cerebras infrastructure at 750 tokens per second, OpenAI is challenging competitors on raw inference velocity while maintaining benchmark parity with Claude Mythos Preview. What the sources don't address: How will federal agencies manage the bottleneck of continuously assessing iterative point-releases of frontier models under the new 30-day benchmarking mandates?