OpenAI launched the GPT-5.6 model family on June 26 — three tiers called Sol, Terra, and Luna — setting new benchmark records for coding, biology, and cybersecurity. Sol is the new flagship; Terra matches GPT-5.5 at half the price; Luna brings frontier capability at the lowest cost OpenAI has offered. The rollout comes with a catch: at the Trump administration’s request, access starts with a handpicked group of trusted partners before it opens to developers and the public.
OpenAI published its GPT-5.6 preview announcement on Thursday, framing the limited release as a short-term coordination step with the U.S. government before broader availability arrives “in the coming weeks.” The move signals that OpenAI’s most capable models are now subject to national security review before public launch — a precedent the company says it doesn’t want to become permanent.
GPT-5.6 Sol, Terra, Luna — What Each Model Does and Costs
The three-tier system replaces the old naming convention with something simpler: a generation number plus a permanent capability tier. Sol is always the flagship. Terra is always the balanced middle. Luna is always the fast, affordable layer. Future versions can advance each tier independently.
Pricing per one million tokens: Sol runs $5 input and $30 output — roughly in line with GPT-5.5. Terra is $2.50 input and $15 output, competitive with GPT-5.5 at 2× lower cost. Luna comes in at $1 input and $6 output, OpenAI’s lowest API price for a frontier-class model.
OpenAI is also launching GPT-5.6 Sol on Cerebras in July, targeting speeds up to 750 tokens per second — fast enough for real-time agentic loops that current infrastructure makes impractical. Prompt caching gets an upgrade too: explicit cache breakpoints, a 30-minute minimum cache life, and cache writes billed at 1.25× the uncached input rate (reads stay at the 90% discount).
The Benchmark Numbers That Actually Matter
Sol Ultra — which gives the model extended reasoning time plus multi-agent “ultra” mode — scores 91.9% on Terminal-Bench 2.1, the command-line workflow benchmark testing long-horizon planning and tool coordination. Plain Sol sits at 88.8%, narrowly edging Claude Mythos 5 at 88.0%. Luna still reaches 82.5%, well above where previous mid-tier OpenAI models landed.
Terra posts 84.3%, matching Claude Fable 5 and beating GPT-5.5’s 83.4% — at half the price. For teams running high-volume inference, that cost differential compounds fast.
On biology workflows, Sol shows meaningful gains on GenBench v1 (long-horizon genomics tasks) using fewer tokens than GPT-5.5. The cybersecurity picture is more complicated: Sol is competitive with Anthropic’s Mythos Preview on ExploitBench using roughly one-third the output tokens — making it cheaper for both offense and defense.
Why the Government Gets First Access
OpenAI’s decision to let the Trump administration coordinate the launch is unusual even by frontier-AI standards. The company frames it as a short-term measure while working with the Administration on a “cyber Executive Order framework” and a “repeatable process for future model releases.”
The framing matters: Sol’s cyber capabilities are strong enough that OpenAI spent over 700,000 A100-equivalent GPU hours on automated red-teaming before launch, using its own models to find jailbreaks at scale rather than relying solely on human testers. The company concludes that Sol does not cross the “Cyber Critical” threshold under its Preparedness Framework — but notes that the benchmark can’t capture every real-world use case.
The phased rollout, then, is less about safety failure and more about optics. OpenAI can argue it cooperated with national security reviewers before releasing a model that can find Chromium and Firefox vulnerabilities.
The Safeguard Stack Underneath
GPT-5.6 runs layered defenses: model-level training to refuse prohibited cyber assistance, real-time output classifiers that can pause generation for secondary review by a larger reasoning model, and account-level monitoring that looks across conversations to distinguish persistent malicious patterns from legitimate dual-use security work.
OpenAI is also testing “privacy-preserving detection” and “customer-operated safety controls” for enterprise customers — a nod to the fact that enterprises have legitimate reasons not to want OpenAI reading their security research queries.
Users in the preview may see occasional refusals or slower responses when the real-time classifiers kick in. OpenAI says that friction is intentional — it’s testing whether the safeguards block misuse without meaningfully disrupting legitimate work.
Frequently Asked Questions
What is GPT-5.6 Sol?
GPT-5.6 Sol is OpenAI’s current flagship model, launched in limited preview on June 26, 2026. It is the most capable model in the GPT-5.6 family, which also includes Terra (balanced) and Luna (fast and affordable). Sol sets new records on coding and biology benchmarks and introduces “ultra” mode, which uses multiple subagents to accelerate complex tasks.
What is the difference between GPT-5.6 Sol, Terra, and Luna?
Sol is the flagship — most capable, most expensive at $5/$30 per million tokens. Terra is the balanced option: performance close to GPT-5.5 at half the price ($2.50/$15). Luna is the fast, affordable tier at $1/$6 per million tokens — best for high-volume production workloads where cost matters more than top-end reasoning.
When will GPT-5.6 be available to everyone?
OpenAI said GPT-5.6 Sol, Terra, and Luna will be “generally available in the coming weeks” following the limited preview that began June 26, 2026. Initial access is through the API and Codex for a select group of trusted partners. Broader availability via ChatGPT and the public API is expected within weeks.
How does GPT-5.6 compare to Claude Opus 4.8?
On Terminal-Bench 2.1, GPT-5.6 Sol scores 88.8% versus Claude Opus 4.8’s 78.9% — a meaningful gap for agentic coding tasks. Claude Opus 4.8 holds the top spot on the Artificial Analysis Intelligence Index with a score of 61.4. The two models target different strengths: Sol excels at long-horizon command-line workflows; Claude Opus 4.8 leads on general intelligence benchmarks.
What is GPT-5.6 “ultra” mode?
Ultra mode is a new reasoning configuration in GPT-5.6 that goes beyond a single agent by coordinating multiple subagents in parallel to tackle complex, multi-step tasks. Combined with the new “max” reasoning effort (which gives the model extended time to think), ultra mode is designed for the most demanding agentic workloads — deep research, long-horizon code tasks, and multi-tool automation chains.
OpenAI’s GPT-5.6 represents the clearest evidence yet that AI capability and geopolitics are converging. Developers interested in what’s coming should watch the GPT-5.5 Cyber Daybreak rollout for context on how OpenAI handled the previous generation, and check our Claude Opus 4.8 vs GPT-5.5 head-to-head to see how the competitive landscape shifted, and read our guide to the best AI coding assistants to see how these new tiers map to real workflows.
Last Updated: June 2026
