OpenAI Clears US Government Review, Sets GPT-5.6 Sol, Terra, and Luna for Public Launch Thursday
OpenAI confirmed GPT-5.6's three-model family launches publicly July 9 after a government-requested early review. Sol hits 88.8% on Terminal-Bench 2.1, with Terra matching GPT-5.5 at half the price.
OpenAI confirmed on X that GPT-5.6 Sol, Terra, and Luna launch publicly this Thursday, July 9, after the US government requested and completed an early review of the model family before wider release — OpenAI says it’s now working with the Administration on a cyber executive order framework for future launches.
Three models, one family
Sol is the flagship — OpenAI’s “strongest model” to date, with enhanced coding, biology, and cybersecurity capabilities, including the ability to identify and patch security vulnerabilities within defined safety boundaries. Terra targets everyday use, matching GPT-5.5-level performance at half the cost. Luna is built for speed and low cost, aimed at high-volume, latency-sensitive workloads.
Pricing lands at $5/$30 per million input/output tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. Terra at $2.50/$15 undercuts GPT-5.5’s pricing while matching its benchmark scores — a direct move to push mid-tier workloads off the older model.
The numbers
On Terminal-Bench 2.1, Sol scores 88.8%, and Sol Ultra — a new reasoning mode that recruits subagents rather than running as a single agent — reaches 91.9%. That’s up from GPT-5.5’s 88.0%. Terra scores 82.5% and Luna, despite being the cheapest tier, edges out Terra at 84.3%.
Training data now runs through roughly May 2026, about two months more recent than GPT-5.5’s April cutoff. Developer reports point to a context window around 1.5 million tokens, though OpenAI hasn’t officially confirmed that figure in launch documentation.
Rollout timeline
GPT-5.6 wasn’t sprung on the market overnight. OpenAI opened a limited preview via the API and Codex on June 26 for trusted partners, then spent roughly two weeks in government-requested review before Thursday’s global preview expansion. That review cycle — unusual for a model launch — reflects Sol’s cybersecurity capabilities specifically; a model that can autonomously find and fix vulnerabilities draws different scrutiny than a general-purpose chat model.
Why it matters
The three-tier pricing structure is the real story for anyone building on the API: Terra’s launch effectively obsoletes GPT-5.5 for cost-sensitive production workloads, while Sol Ultra’s subagent architecture pushes OpenAI further into the same multi-agent reasoning territory Anthropic and Google have been racing toward. Combined with the government review precedent, expect future frontier launches from every US lab to carry a similar pre-release checkpoint whenever cybersecurity or offensive-capability benchmarks are involved.