GPT-5.6 Sol squares up to Fable 5: the same category, but much faster

OpenAI dropped GPT-5.6 overnight in three flavors —Sol, Luna and Terra— and the surprise is Sol. They let Anthropic create the category above Opus with the Mythos family (which Fable 5 comes from) and now they show up saying: that thing you see, I do it too, but much faster. Mitchell Hashimoto, with early access for almost a month, already runs it as his daily driver: «Sol is my default».

His take is sharp: Sol is that charismatic, efficient coworker you're jealous of; Fable is a reclusive genius, brilliant but with its fixations. For the day-to-day, Sol; when things get really hairy —debugging, security, performance— Fable remains unmatched. In Codex the default model is already Sol and the recommended effort is extra-high, and the speed has been a hit: what Fable did in 20 minutes, Sol solves in 5.

Big new features: native programmatic tool calling (the model writes scripts that batch several tools in one turn and drastically cut the back and forth), embedded multi-agents and explicit prompt caching with breakpoints, copied from Anthropic's cache management model. Mind the footnote in the table: cache writes are now billed too, at 1.25x the input price, just like Anthropic.

A prompting tip we loved: don't ask it to 'be concise' or for 'minimal text', because 5.6 is so sensitive to those instructions that it ends up thinking less. Replace it with prioritization: 'lead with the conclusion, include the evidence and the next action, omit secondary details'. Fewer magic spells, more natural prompts. Watch in the video.