OpenAI announced the GPT-5.6 model family on June 26, 2026 — Sol, Terra, and Luna. All three variants are initially available to approximately 20 selected organizations under a limited preview, after OpenAI coordinated its release with the Trump administration. The company openly criticizes this arrangement, arguing that government gatekeeping should not become the new norm.
Key takeaways
- GPT-5.6 Sol (flagship): $5/$30 per million tokens — for complex reasoning, coding, cybersecurity, and agentic workflows
- GPT-5.6 Terra: $2.50/$15 — for high-volume business production environments
- GPT-5.6 Luna (entry tier): $1/$6 — everyday automation, fastest in the family
- Sol (ultra mode) scored 91.91% on TerminalBench 2.1 — a new record, above Claude Mythos 5 (88%)
- General release planned for "the coming weeks," pending government review under Trump's June 2 Executive Order
Three models, three capability tiers
GPT-5.6 Sol targets the hardest problems: extended coding sessions, autonomous agentic workflows, vulnerability research, and security-focused applications. Terra is designed for high-volume production environments — customer support, document analysis, internal tooling. Luna is optimized for speed and everyday use: summarization, drafting, routine automation.
The naming shift matters. OpenAI is moving away from "nano" and "mini" suffixes. The new Sol/Terra/Luna scheme is meant to represent durable "capability tiers" that can advance independently, not size variants.
Benchmarks: Sol sets records, Terra close behind
The most telling numbers come from TerminalBench 2.1, which measures terminal task automation. Sol in ultra mode (with subagents) achieved 91.91% — a new record, beating Claude Mythos 5 (88%) and GPT-5.5 (83.4%). Sol in standard max mode reached 88.76%.
On Agent's Last Exam — a multi-step autonomous workflow benchmark — Sol is the only model to cross the 50% mark (50.9% in "code mode"). Luna, despite being the lightest tier, also edged out the previous generation's flagship on this test.
On ExploitBench (cybersecurity), Sol reached performance comparable to Mythos Preview while using roughly one third of the output tokens. That's an efficiency gain worth noting: equivalent capability at significantly lower operational cost.
New technical mechanisms
GPT-5.6 introduces two new reasoning modes. max reasoning provides extended deliberation for hard problems. ultra mode engages subagents — complex projects are split across parallel agents rather than handled in a single flow.
Prompt caching has also been restructured. Developers can now set explicit cache breakpoints with a guaranteed 30-minute minimum lifetime. Initial cache writes cost 1.25× the standard input rate, but subsequent reads within the cache window receive a 90% discount. For agentic systems with large context windows, this predictability is a meaningful financial guardrail.
Starting in July, Sol will be available on Cerebras hardware, claiming throughput of up to 750 tokens per second.
Safety and geopolitics
All three GPT-5.6 models are rated "High" risk by OpenAI for both cyber and biological/chemical capabilities — which carries compliance obligations for regulated sectors. OpenAI dedicated approximately 700,000 A100e GPU hours solely to automated red-teaming, targeting "universal jailbreaks."
The release context is directly linked to the Anthropic situation. Two weeks ago, the US government ordered Anthropic to pull Claude Fable 5 and Mythos 5 from the market after security researchers identified bypasses. OpenAI chose to coordinate its GPT-5.6 release with the White House and restrict access until a 30-day review mandated by Trump's June Executive Order is complete.
We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
— OpenAI, official blog post, June 26, 2026
Why it matters
GPT-5.6 is not just another generation of models — it is a test of the new power dynamic between AI labs and the US government. For two years, companies like OpenAI, Anthropic, and Google operated in a rapidly evolving but relatively hands-off regulatory environment. Now a precedent exists: the government can conditionally block or delay AI model releases on security grounds.
For enterprises using OpenAI models, this translates into concrete implications: new compliance risk categories for Sol and Terra in cybersecurity and life sciences, a restructured cost model for long-context workloads (new prompt caching), and in July — a potential step-change in inference speed for Sol via Cerebras.
For the industry, the key question is whether the temporary GPT-5.6 gatekeeping becomes a precedent for future models. If so, every AI company with global reach will need to negotiate its own access protocols with different governments.
What's next?
- OpenAI plans a broad GPT-5.6 release "in the coming weeks" — contingent on the 30-day government review (deadline: July 2, 2026)
- Sol on Cerebras hardware (750 tok/s) launches in July 2026 per OpenAI's announcement
- OpenAI is negotiating long-term enterprise safety compliance controls including customer-operated safety overrides
Sources
VentureBeat — OpenAI unveils GPT-5.6 Sol, Terra and Luna models
OpenAI Blog — Previewing GPT-5.6 Sol
TechCrunch — OpenAI limits GPT-5.6 rollout after government request
The Verge — OpenAI unveils GPT-5.6 amid US AI regulatory drama
