OpenAI launches GPT-5.6 Sol family with government-gated access; leads TerminalBench at 91.9%
OpenAI announced GPT-5.6, a three-model family — Sol (flagship), Terra (balanced), and Luna (affordable) — on June 26 in limited preview under a novel government-gated rollout. Sol and Terra are available only to ~20 U.S. government-vetted partners via API, with broader ChatGPT/API access pending formal federal cyber assessment by August 2026. This marks the first time an American frontier lab has released a model under government-managed access lists, driven by June 2 Trump cyber executive order.
Benchmark-wise, Sol achieves state-of-the-art on TerminalBench 2.1 (command-line automation): 91.9% in Ultra reasoning mode, 88.8% in standard Max mode—ahead of Anthropic's Claude Mythos 5 (88.0%) and previous GPT-5.5 (83.4%). On ExploitBench (cybersecurity), Sol matches Mythos Preview capability while using ~1/3 the output tokens. GeneBench v1 (biology) shows improvement over GPT-5.5 with lower token consumption. Terra is positioned as GPT-5.5-level capability at 2x lower cost; Luna as the volume tier.
Pricing reflects the tiering strategy: Sol at $5/$30 (same as GPT-5.5); Terra at $2.50/$15; Luna at $1/$6. OpenAI also introduced explicit prompt caching with 30-minute minimum lifetime and 90% reads discount, shifting costs for agentic systems away from pure token count toward architectural routing decisions.
For builders: this is OpenAI reframing frontier access as a *routing problem*, not a binary cost/capability trade. The government gating is temporary (per OpenAI's framing), but the model family structure—independent tiers that can advance separately—normalizes fine-grained capability/cost segmentation. Watch for competing multi-tier launches from Anthropic, DeepSeek, and others.
Sources
- Primary source
- openai.com
“GPT-5.6 Sol launches with our most robust safety stack to date...GPT-5.6 Sol is our most capable model yet for cybersecurity”
- venturebeat.com
“Sol scored 88.76% on TerminalBench 2.1 — ahead of both GPT-5.5's 83.4% and Claude Mythos 5's 88%”
- mlq.ai
“the company is restricting initial access to roughly 20 organizations individually approved by the US government. The launch marks the first time an American AI company has released a frontier model under a government-managed access list”
- kingy.ai
“Terra is competitive with GPT-5.5 while being 2x cheaper...Luna is the throughput play...explicit cache breakpoints, a 30-minute minimum cache lifetime, cache writes billed at 1.25x”