The narrator introduces GPT-5.6 Sol, Terra and Luna as preview models and separates reported vendor capability claims from his own KingBench 3 results. Release plans, pricing and inference-speed announcements are historical context rather than current purchasing advice.
In his reported test, Sol scores55 of70, Terra44 and Luna31. The task mix includes front-end work, graphics, a hard math problem and an agentic workflow that generates data, fine-tunes a model locally and builds a web interface. These are results from his small benchmark, not a broad independent capability guarantee.
Sol and Terra receive full marks on the math problem, and all three models receive full marks on the long-horizon fine-tuning task. The narrator contrasts that with lower scores on visual and front-end examples and acknowledges that his benchmark is weighted toward those areas.
He characterizes the release as a strong increment rather than a new paradigm, with task-dependent tradeoffs against Fable and Opus. The closing caveat is important: preview behavior can change before general availability, and the aggregate score does not capture every useful workload.
Watch on YouTube
