OpenAI puts GPT-5.6 Sol behind a government-gated preview

OpenAI released GPT-5.6 Sol, Terra and Luna on June 26, placing the frontier model family into a limited preview after a U.S. government request. The controlled launch gives government visibility and selected customer access while OpenAI works on a repeatable release process. The models are priced at $5/$30 per million tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna.

OpenAI @OpenAI https://x.com/OpenAI introduced GPT-5.6 Sol, Terra and Luna on June 26, placing its newest frontier model family into a limited preview after what OpenAI described as a request from the U.S. government, according to a six-post thread on X https://x.com/openai/status/2070555272230384038?s=46 and an accompanying OpenAI blog post https://openai.com/index/previewing-gpt-5-6-sol/ . The release is a model launch and a policy signal in the same package. OpenAI says GPT-5.6 Sol is its new flagship model, Terra is a lower-cost model intended for everyday work, and Luna is the fastest and cheapest tier. But unlike a typical general rollout, OpenAI is first making the models available only to a small group of trusted partners through the API and Codex, with broader access planned in the coming weeks. OpenAI said it previewed the models' capabilities and release plan to the U.S. government before launch. At the government's request, OpenAI said it is beginning with partners whose participation has been shared with the government. OpenAI also said it does not want that kind of government access process to become the long-term default, arguing that it keeps tools away from users, developers, enterprises, cyber defenders and global partners that need them. That sentence is the center of the announcement. OpenAI is trying to move a more capable model family into the market while avoiding a broader release fight over cyber and bio risk. OpenAI said the preview is a short-term step while it works with the Administration on a cyber Executive Order framework and a repeatable process for future model releases. The effect is a controlled launch that gives government officials visibility, gives selected customers access, and gives OpenAI time to harden safety systems before the models reach a wider developer base. Three tiers, one naming reset OpenAI is also using GPT-5.6 to reset how it names models. In the new system, the number identifies the model generation, while Sol, Terra and Luna identify capability tiers that can move on their own cadence. Sol is the flagship tier. Terra is the balanced tier. Luna is the cost-efficient tier. OpenAI says Terra delivers performance competitive with GPT-5.5 at 2x lower cost, while Luna offers strong capability at OpenAI's lowest cost. Those are OpenAI's claims; OpenAI has not yet published the broader evaluation suite it says will come with general availability. The pricing is already public. GPT-5.6 Sol costs $5 per 1 million input tokens and $30 per 1 million output tokens. Terra costs $2.50 per 1 million input tokens and $15 per 1 million output tokens. Luna costs $1 per 1 million input tokens and $6 per 1 million output tokens. OpenAI is also changing prompt caching for GPT-5.6 and later models: cache writes are billed at 1.25x the uncached input rate, while cache reads continue to receive a 90% cached-input discount. That pricing puts the lineup into an increasingly familiar frontier-model shape: one expensive model for the hardest work, one middle model for default production workloads, and one cheaper model for high-volume traffic. For founders building on top of OpenAI, the immediate question is not just whether Sol is better. It is whether Terra or Luna can absorb enough everyday work to cut inference costs without forcing a product downgrade. OpenAI is selling capability, then explaining containment OpenAI says GPT-5.6 Sol is its strongest model yet, with improved agentic capabilities in coding, biology and cybersecurity. On Terminal-Bench 2.1, which tests command-line workflows requiring planning, iteration and tool coordination, OpenAI says Sol sets a new state of the art. OpenAI also says Sol improves on GeneBench v1, a genomics and quantitative-biology benchmark, while using fewer tokens than GPT-5.5. The cyber claims are more carefully framed. OpenAI says Sol is its most capable model yet for cybersecurity and improves long-horizon tasks such as vulnerability research and exploitation. On ExploitBench, OpenAI says Sol is competitive with Mythos Preview while using roughly one-third of the output tokens. OpenAI also says Sol, Terra and Luna show stronger cyber performance on ExploitGym as reasoning effort increases. OpenAI is not presenting those gains as unqualified upside. In the GPT-5.6 Preview System Card https://deploymentsafety.openai.com/gpt-5-6-preview , OpenAI classifies Sol, Terra and Luna as High capability in both Cybersecurity and Biological and Chemical risk under its Preparedness Framework. OpenAI says none of the models reaches the High threshold in AI Self-Improvement, and says the GPT-5.6 family does not reach the Cyber Critical threshold. The system card adds an important caveat to the launch copy: OpenAI says GPT-5.6 Sol and Terra can find vulnerabilities and components of exploits, but in OpenAI's cybersecurity testing they did not autonomously carry out end-to-end attacks against hardened targets. OpenAI also says separate evaluations found GPT-5.6 has a greater tendency than GPT-5.5 to go beyond the user's intent in agentic coding tasks, including taking or attempting actions the user had not requested, though OpenAI says absolute rates remain low. That is the tradeoff OpenAI is asking customers and regulators to accept. GPT-5.6 is more useful for defenders because it is better at finding and fixing weaknesses. The same capability pushes OpenAI closer to the line where a model can help with offensive activity. OpenAI's answer is not to delay the model indefinitely. It is to add more gates around who gets access, what outputs are allowed through, and how high-risk activity is monitored. The safety stack is now part of the product OpenAI says GPT-5.6 Sol launches with its most robust safety stack to date. The system includes model-level refusals for prohibited cyber assistance, real-time cyber and biology misuse classifiers that can pause generation, review by a larger reasoning model in higher-risk cases, account-level review across conversations, differentiated access and ongoing testing. OpenAI said it spent more than 700,000 A100-equivalent GPU hours on automated red teaming for universal jailbreaks, alongside human expert red teaming. OpenAI's stated goal is to find attacks that work across many prompts or contexts, rather than only narrow one-off failures. OpenAI says the preview period will continue that testing, including feedback on whether safeguards block legitimate defensive work. Those safeguards will not be invisible to users. OpenAI says preview users may see refusals or delays when generation is paused for review, and that safeguards may occasionally intervene on legitimate work in dual-use areas where defensive and offensive activity can look similar at first. For enterprise customers, OpenAI says it is working on longer-term approaches including privacy-preserving detection, customer-operated safety controls and access calibrated to the risk of a customer, user or workload. That is a practical concession to the market OpenAI wants to keep: large customers want stronger models, but many will resist safety systems that require broad inspection of sensitive internal work. A speed play through Cerebras OpenAI is also pairing Sol with a hardware distribution note. OpenAI says GPT-5.6 Sol will launch on Cerebras in July at up to 750 tokens per second, initially for select customers as capacity expands. The Cerebras detail matters because the cost and latency of frontier models increasingly decide which products can use them in production. A model that is strongest on paper but too slow for interactive coding, security triage or agentic workflows is limited to narrower use cases. OpenAI is presenting Sol not just as a larger step in reasoning capability, but as something it can make fast enough for customers who need frontier intelligence inside time-sensitive products. The immediate rollout remains narrow. OpenAI says the preview is limited to selected trusted partners in the API and Codex, with broader availability planned for ChatGPT, Codex and the API soon. Until OpenAI publishes the expanded evaluation suite and opens access beyond the preview group, the strongest claims around Sol, Terra and Luna remain OpenAI's own launch-time benchmarks. What is already clear is the shape of the release. GPT-5.6 is not just another model card and pricing table. It is OpenAI's attempt to define a release process for more capable dual-use models before government pressure defines one for it.