OpenAI ships GPT-5.6 series (Sol, Terra, Luna) plus GPT-Live voice, after federal clearance

GPT-5.6 arrived as a tiered family rather than a single flagship: Sol is the frontier reasoning model, Terra a high-capability mid-tier, and Luna a hyper-fast, cheap option. OpenAI's headline claim is efficiency — Luna beats GPT-5.5 at its highest reasoning setting while costing 25x less, and the company positioned GPT-5.6 as 'a major step forward for health intelligence' with stronger performance at lower cost across the lineup.

Alongside the text models, OpenAI launched GPT-Live-1 and GPT-Live-1 mini, voice models designed to feel more human and interruptible, now the default for ChatGPT Voice across all tiers. President Greg Brockman framed voice as 'a new era of a natural interface,' predicting users will spend less time staring at screens.

Early third-party benchmarks temper the hype. LlamaIndex's day-zero ParseBench run found 'no change between GPT-5.6 Sol and GPT-5.5' on document understanding — strong on tables and text, still weak on charts and layout — and noted Luna is ~6x cheaper than Sol with only minor degradation, suggesting more reasoning tokens don't buy better visual understanding. Migration anecdotes were rosier: one HN post reported moving a production agent to GPT-5.6 for '2.2x faster, 27% cheaper.'

The federal-clearance framing is notable given the regulatory backdrop — the same government pressure that forced Anthropic to disable Fable 5. Safety scrutiny followed immediately: UK AISI reportedly found universal jailbreaks within hours despite OpenAI's 'robust safety stack,' reviving the debate over whether benchmark wins matter more than guardrail durability.