OpenAI launches GPT-5.6 family (Sol, Terra, Luna) after government cybersecurity delay

OpenAI's GPT-5.6 launch, rolled out globally over 24 hours across ChatGPT, Codex and the API, is the centerpiece of a frantic 72-hour stretch that also saw Meta and xAI ship frontier models. The family splits into three tiers: Sol as the workhorse flagship, Terra as a mid-tier enterprise model priced at $2.50 input / $15 output (matching GPT-5.5 at roughly half the cost), and Luna as the budget option at about $1 input, which OpenAI claims beats GPT-5.5's top reasoning setting at 25x lower cost.

The headline benchmark is Sol's 91.9% on Terminal-Bench 2.1, which developers say edges Anthropic's Claude Fable 5 (83.4%). But the community consensus is more nuanced: on SWE-Bench Pro, Claude Fable 5 (80.3%) still leads, and Reddit's summary — 'Fable is better base by large margin, but GPT is stronger exponent' — captures the split. LlamaIndex's day-zero ParseBench run found no meaningful gain over GPT-5.5 on document, chart and layout understanding, and noted Luna is ~6x cheaper than Sol with only minor degradation.

Notably, the rollout was staggered at the White House's request over cybersecurity concerns — mirroring earlier Anthropic release restrictions and fueling debate about government gating of frontier models. Pricing, not raw capability, is the story: Altman explicitly framed 5.6 as 'a huge step forward for dollars-per-task,' aligning the launch with the week's cost-cutting mood. Watch whether the Terminal-Bench lead translates to real-world SWE work, where practitioners still treat Claude as the production standard.