OpenAI's GPT-5.6 Sol Launch Spurs Optimizations, Benchmark Gains and Usage Fixes

OpenAI released GPT-5.6 Sol publicly on July 9 2026 as the flagship of a three‑tier model family—Luna, Terra and Sol—after a U.S. government request limited the rollout to a small group of trusted partners.

The company reported that GPT‑5.6 Sol helped optimize its own inference pipeline, rewriting GPU kernels and improving speculative decoding, which cut end‑to‑end serving costs by about 20 % and boosted token‑generation efficiency by more than 15 %.

When researchers adjusted two OpenAI API settings—retaining reasoning and enabling compaction—GPT‑5.6 Sol’s score on the ARC‑AGI‑3 benchmark rose sharply, and token output fell by roughly six‑fold.

After power users complained that Sol’s new programmatic tool‑calling drained ChatGPT Work and Codex limits too quickly, OpenAI reset those limits and rolled out backend improvements that extend typical Sol sessions by roughly 18 %.

Separately, OpenAI altered ChatGPT’s policy so it now refuses direct requests to emulate the style of specific authors, offering only text that captures general stylistic traits. The change sparked backlash among writers on Reddit and other forums.