OpenAI unveiled GPT-5.6 on Friday in a limited preview, less than 24 hours after reports that the Trump administration had asked the company to stagger the release. The launch covers three models — Sol, the flagship; Terra, a mid-tier model aimed at high-volume work; and Luna, a fast everyday model — and OpenAI is pricing Sol at $5 input and $30 output per million tokens.
That price puts Sol at roughly half the cost of Anthropic's Claude Fable 5, which runs $10 input and $50 output per million tokens. Terra is priced at half of Sol, and Luna sits at less than half of Terra. The pricing structure is the most aggressive tiering OpenAI has shipped in a flagship release, and it lands while Anthropic is leaning the other way on per-token cost.
OpenAI says the model suite is especially capable at coding, cybersecurity, and biology, and that it holds focus better than prior generations on long-horizon agentic tasks. Sol also gets two additional modes: a 'max' mode for deeper reasoning and an 'ultra' mode that orchestrates sub-agents. The sub-agent mode echoes OpenClaw, the project associated with Peter Steinberger, who joined OpenAI earlier this year.
Key facts
- 01GPT-5.6 Sol is priced at $5 input and $30 output per million tokens — roughly half the cost of Anthropic's Claude Fable 5 at $10 / $50.
- 02The release ships in three tiers: Sol (flagship), Terra (half Sol's price), and Luna (less than half Terra's price).
- 03OpenAI dedicated about 700,000 A100e GPU hours to automated red-teaming ahead of launch.
- 04Third-party testers will continue probing the model for two more weeks during the preview period.
- 05The launch came less than 24 hours after news broke that the Trump administration had asked OpenAI to stagger the release.
Safety occupied the majority of the announcement post, an unusual ratio for an OpenAI model launch and a direct read of the regulatory mood in Washington. The company said GPT-5.6 was trained to refuse prohibited cyber assistance even when users try to disguise intent or jailbreak the model — a clear nod to the recent jailbreaking issues at Anthropic. OpenAI also said Sol is 'better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks,' and that it does not cross the cyber-critical threshold under the company's preparedness framework.
“GPT‐5.6 is trained to refuse prohibited cyber assistance, including when users attempt to disguise their intent or jailbreak the model.”— OpenAI, company statement
That framework itself was revised in April 2026, with some prior areas of study removed, which complicates the claim. OpenAI described Sol as carrying its 'most robust safety stack to date' with strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse. The company dedicated roughly 700,000 A100e GPU hours to automated red-teaming and engaged third-party testers, who will continue evaluating the model for the next two weeks.
OpenAI flagged a real tradeoff in the preview: safeguards may occasionally block legitimate work, particularly in dual-use areas where defensive and offensive cyber activity look similar at first glance. The company framed that friction as part of what the preview is designed to surface. Earlier reporting this week indicated the Trump administration will approve customers on a case-by-case basis during the preview window, which we covered when OpenAI's delay was first disclosed.
General availability is expected in the coming weeks. OpenAI said it cooperated with the US government ahead of launch but signaled it does not want the arrangement to become permanent. The company wrote that this kind of government access process keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
“We don't believe this kind of government access process should become the long-term default.”— OpenAI, company statement
OpenAI framed the current setup as a short-term step it believes is the strongest path to broader availability while it works with the administration to develop the cyber Executive Order framework and a repeatable process for future model releases. In practice, that means the White House is now in the release loop for frontier models from both OpenAI and Anthropic, a shift in how US AI policy actually operates day to day.
The aggressive pricing matters more than the model card. By undercutting Claude Fable 5 by half on both input and output, OpenAI is signaling that the cost of frontier intelligence is meant to fall, not rise, and that it is willing to absorb margin to keep developers from migrating. Anthropic has been winning share among paying consumers and coding-heavy workloads; the response here is a price cut at the top of the lineup, not a feature war.
The bigger structural story is the case-by-case approval regime. A federal gate on who can use a commercial AI model — even temporarily — sets a precedent that every future frontier release will have to navigate, and it raises the cost of being a frontier lab in a way that smaller players cannot absorb. OpenAI's pushback in the announcement is calibrated: cooperate now, lobby for a repeatable framework, but make clear this should not be the default. Whether that line holds depends less on OpenAI than on what the next jailbreak headline looks like.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
