A model guide for the GPT-6 family

OpenAI published a practical guide for the GPT-6 family — Astra for hardest reasoning, Sol for complex coding and computer use, Luna for focused tasks at scale — covering model choice, reasoning effort, Fast/Ultrafast modes, prompt caching, compaction, mid-turn steering, async tools, and multi-agent delegation.

The pitch is operational: cached input tokens can cost up to 95% less than uncached ones; measure success, latency, and cost per successful task before shipping; set decision boundaries so agents don’t freeze on every choice or freestyle on ones that matter. Customer vignettes from Harvey, Cognition, Hex, and Invideo show legal drafts, test evidence, dashboards, and timeline edits as the intended use cases.

We’re past the era where “just use the biggest model” was an acceptable product decision. This guide reads like OpenAI admitting that GPT-6 only monetizes if builders treat intelligence as a dial — capability, reasoning effort, and speed priced separately — and keep long-running work from burning context and cash.

I’m watching the cultural shift in prompting as much as the feature list. Eric Provencher’s line that overly specific guidance can now hinder results is a real break from years of brittle prompt recipes. Skills, AGENTS.md, and persistence rules become the product surface: authorize safe local tests, define what “done” includes, and say when to ask. That’s useful AI if teams actually instrument it.

The downside scales with the same features. Mid-turn steering, computer use on browsers and desktops, and Sol’s multi-agent beta expand blast radius when boundaries are fuzzy. Compaction and caching cut cost; they also hide what the model still “knows.” Production checklists that skip data controls and monitoring will ship confident agents into audited workflows without an audit trail.

Our read: treat this as infrastructure documentation, not a capability flex. The labs that win the next year will be the ones that can prove task success per dollar — and can yank autonomy when a steering update or subagent goes sideways.

The guide lands days after DevDay 2026 recaps and the Sol/Dots product wave, positioning GPT-6 as a suite you operate, not a single chat endpoint.