OpenAI's GPT-5.6 Sol Launches Under Unprecedented U.S. Government Restrictions
OpenAI's GPT-5.6 family launches in limited preview at US government request, with Sol setting new agentic coding and cybersecurity benchmarks at a new naming tier structure
Takeaways
- OpenAI previews GPT-5.6 Sol, Terra, and Luna -- a new flagship, balanced, and cost-efficient model family.
- Sol sets a new state of the art on Terminal-Bench 2.1 and matches Anthropic's restricted Mythos model on ExploitBench using 3x fewer tokens.
- Pricing: Sol at $5/$30, Terra at $2.50/$15, and Luna at $1/$6 per million input/output tokens; Terra delivers GPT-5.5 performance at half the cost.
- Two new modes: max reasoning effort for deeper thinking, and ultra mode that spawns parallel subagents for complex tasks.
- The U.S. government asked OpenAI to restrict the launch to vetted partners first -- the first time Washington has preemptively limited a frontier model release.
- General availability across ChatGPT, Codex, and the API is expected within a few weeks pending government-managed partner approvals.
OpenAI just announced the GPT-5.6 family: three models named Sol, Terra, and Luna. Sol is the new flagship, Terra is the balanced mid-tier, and Luna is the cost-efficient workhorse. But the launch comes with an unusual asterisk: at the request of the U.S. government, the models are starting in a limited preview for a small group of vetted partners rather than rolling out broadly. General availability is expected in the coming weeks.
A new naming system, three distinct tiers
GPT-5.6 introduces a new model naming convention. The number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence. Think of it like a product line: Sol is always the flagship, Terra is always the balanced option, and Luna is always the fast and cheap one, regardless of which generation they're in.
- Sol -- the new flagship, described as a step-function improvement over GPT-5.5
- Terra -- competitive performance to GPT-5.5 at 2x lower cost
- Luna -- OpenAI's most cost-efficient model, delivering strong capability at their lowest cost
What Sol can actually do
Sol's headline benchmark is Terminal-Bench 2.1, an upgraded version of the agentic coding eval that GPT-5.5 topped on its 2.0 version. Terminal-Bench evaluates models in live, containerized environments, testing strong reasoning under uncertainty, failure recovery, and tool-use discipline -- proving that the model doesn't just write code, but explores the environment, reads configurations, and tests its own solutions iteratively. GPT-5.5 scored 82.7% on Terminal-Bench 2.0; Sol sets a new state of the art on the harder 2.1 version.
Beyond coding, Sol also shows gains in biology. On GeneBench v1, which evaluates long-horizon genomics and quantitative-biology analyses, it achieves stronger results than GPT-5.5 while using fewer tokens. The token efficiency improvement matters for long-running agentic workflows where costs compound quickly.
Cybersecurity is where Sol's capabilities are most notable -- and most carefully managed. On ExploitBench, GPT-5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens. ExploitBench measures a model's ability to find and chain together software vulnerabilities -- the kind of task that has historically required senior security researchers. On ExploitGym, a benchmark created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, GPT-5.6 Sol, Terra, and Luna models all demonstrate strong improvements in cyber capabilities as reasoning increases.
Two new capability modes
GPT-5.6 introduces two new ways to push Sol harder on complex tasks:
- max reasoning effort -- a new reasoning effort level that gives Sol the most time to reason deeply , extending the existing none/low/medium/high/xhigh ladder
- ultra mode -- goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work . Instead of one model working sequentially, ultra mode can spawn parallel agents tackling different parts of a problem simultaneously
Ultra mode is particularly relevant for large-scale engineering tasks: imagine spawning one subagent to audit security, another to check test coverage, and a third to review maintainability, all in parallel, then merging their findings.
Pricing and availability
GPT-5.6 is priced per 1M tokens: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. For context, Sol matches GPT-5.5's price point while delivering meaningfully better performance, and Terra delivers GPT-5.5-level performance at half the cost.
There's also a new prompt caching model worth noting. GPT-5.6 introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the model's uncached input rate, while cache reads continue to receive the 90% cached-input discount. For agentic pipelines that repeatedly send large system prompts, this is a meaningful cost lever.
On the speed front, OpenAI is launching GPT-5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed, with access initially limited to select customers as capacity expands.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Position |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | Flagship |
| GPT-5.6 Terra | $2.50 | $15.00 | Balanced |
| GPT-5.6 Luna | $1.00 | $6.00 | Cost-efficient |
The safety stack behind the most capable cyber model yet
OpenAI is unusually candid about the tension here: Sol is their most capable cybersecurity model, and that's precisely why it ships with their most robust safety system. The approach is layered rather than relying on any single guardrail:
- Model-level refusals trained to resist jailbreaks and disguised intent
- Real-time cyber and biology misuse classifiers that evaluate output as it is generated -- for higher risk cases, if they detect a potential violation, generation may be paused while a larger reasoning model reviews the conversation and its context
- Account-level review that looks across conversations to distinguish persistent malicious behavior from legitimate dual-use security work
- Differentiated access, so the most sensitive capabilities aren't broadly available by default
OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red teaming aimed at finding universal jailbreaks -- attacks that can work across many prompts or contexts, not just one narrow setting. That's a significant compute investment in safety testing alone, and it reflects how seriously OpenAI is treating the dual-use risk of a model this capable at offensive security tasks.
Importantly, OpenAI is explicit about Sol's current limits: in evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives -- the building blocks of an exploit -- but did not autonomously produce a functional full-chain exploit under the conditions tested. It does not cross the "Cyber Critical" threshold under OpenAI's Preparedness Framework.
The government intervention reshaping AI releases
The most consequential aspect of this launch may not be the benchmarks. This marks the first time the U.S. government has preemptively asked an American AI company to restrict the launch of a model before release. The request came from the White House's Office of the National Cyber Director and Office of Science and Technology Policy, and maps onto a broader executive order framework.
President Trump's executive order on "Promoting Advanced AI Innovation and Security" calls on AI companies to voluntarily share frontier models with the government for cybersecurity review for up to one month before public release, but participation remains voluntary. In practice, though, the OpenAI case shows just how voluntary that review really is -- Altman explained that the government would approve access "on a customer by customer basis" during the preview phase.
The backdrop is Anthropic's recent experience. On June 12, 2026, the Trump administration used an export control directive to force Anthropic to take its latest models, Fable 5 and Mythos 5, entirely offline to prevent access by foreign nationals, citing serious national security concerns. According to a source familiar with the situation, both OpenAI and the administration view GPT-5.6 as "on par" with Anthropic's Mythos in terms of advanced capabilities, particularly in cybersecurity.
OpenAI is cooperating but pushing back on the long-term precedent. The company states directly in its announcement that it doesn't believe this kind of government access process should become the default, arguing it keeps the best tools from developers, enterprises, cyber defenders, and global partners who need them. How this preview period unfolds for GPT-5.6 could set the template for how every major AI lab releases powerful models going forward.
What this means for developers right now
For most developers, the practical reality is a waiting period. During the preview, GPT-5.6 models will initially be available through the API and Codex to a select group of trusted partners and organizations, with plans to make them more broadly available to people using ChatGPT, Codex, and the API soon.
When access does open up, the use cases that will benefit most from Sol specifically are:
- Long-horizon agentic coding -- multi-step terminal workflows, debugging complex system failures, large-scale refactors
- Security research and vulnerability analysis -- finding and patching bugs faster, with the caveat that access for offensive use cases will be tightly controlled
- Genomics and quantitative biology -- long-running analysis pipelines where token efficiency matters
- High-volume API workloads -- Luna at $1/$6 per million tokens makes it competitive for anything where you need volume at low cost
Terra is probably the most immediately interesting model for production use: GPT-5.5-level intelligence at half the price is a straightforward upgrade path for teams already running on GPT-5.5. The new naming system also gives a clearer signal about what to expect from future generations -- when GPT-5.7 or GPT-6 arrives, Sol, Terra, and Luna will still mean the same things relative to each other.