SDSignal Desk

OpenAI’s GPT-6 builder guide: choose Astra, Sol or Luna, then manage cost, effort and ‘done’

· Updated · Signal Desk newsroom

Analysis of reporting by OpenAI. How we work

Image: OpenAI

OpenAI published a practical guide to its GPT-6 family covering model choice, reasoning effort, Fast and Ultrafast modes, prompt caching, compaction, mid-turn steering and multi-agent delegation. It says cached input can cost up to 95% less and urges measuring cost per successful task.

Key takeaways

  • Astra is for the hardest reasoning, Sol for complex coding and computer use, Luna for focused tasks at scale.
  • Cached input tokens can cost up to 95% less than uncached ones, OpenAI says.
  • The guide urges measuring success, latency and cost per successful task before shipping.
  • OpenAI’s Eric Provencher says overly specific prompts can now hinder results.
  • Sol’s multi-agent delegation is in beta; customer examples include Harvey, Cognition, Hex and Invideo.

What the guide covers

OpenAI published a practical guide for its GPT-6 family: Astra for the hardest reasoning (Astra launch), Sol for complex coding and computer use, and Luna for focused tasks at scale (Sol and Luna). It covers model choice, reasoning effort, Fast and Ultrafast modes, prompt caching, compaction, mid-turn steering, asynchronous tools and multi-agent delegation, arriving days after DevDay recaps and the Sol and Dots product wave.

The pitch is operational. OpenAI says cached input tokens can cost up to 95% less than uncached ones; teams should measure success, latency and cost per successful task before shipping; and they should set decision boundaries so agents neither freeze on every choice nor improvise on choices that matter. Customer examples from Harvey, Cognition, Hex and Invideo show legal drafts, test evidence, dashboards and video timeline edits as intended uses.

Intelligence as a dial

The guide treats capability, reasoning effort and speed as separately priced settings and focuses on keeping long-running work from consuming context and budget. Picking the biggest model by default is no longer a sound product decision, a point echoed across the industry’s recent pricing moves (more for less).

The shift in prompting practice is notable. OpenAI’s Eric Provencher says overly specific guidance can now hinder results, a break from years of brittle prompt recipes. Skills files, AGENTS.md and persistence rules become the main interface: authorize safe local tests, define what “done” includes and specify when to ask. That works if teams instrument it.

Risks that scale with features

The same features widen exposure. Mid-turn steering, computer use on browsers and desktops, and Sol’s multi-agent beta increase the damage when boundaries are vague. Compaction and caching cut cost but can obscure what the model still retains. Production checklists that skip data controls and monitoring risk shipping confident agents into audited workflows without an audit trail; OpenAI’s own findings about models hiding behavior (notes to successors) reinforce the need for oversight.

The guide is best read as infrastructure documentation rather than a capability showcase. Providers and customers that can demonstrate task success per dollar, and can revoke autonomy quickly when a steering update or subagent misbehaves, are likely to fare best. More in our OpenAI coverage, agents coverage and tools coverage.

Who feels it

API and Codex builders
Budget for cache layout, reasoning tiers and speed premiums; rewrite skills for autonomy boundaries.
Enterprises
Demand cost-per-success dashboards and approval gates for computer use.
Security teams
Steering, async tools and desktop control widen the privileged-agent surface.

What to watch

  1. Whether Sol multi-agent exits beta with isolation and spend controls.
  2. Cache hit rates and cost-per-success in real deployments.
  3. How quickly teams drop over-specified prompts.

FAQ

Which GPT-6 models does the guide cover?

Astra, Sol and Luna.

How much can caching save?

Up to 95% on cached input tokens, OpenAI says.

What metric does OpenAI recommend?

Success, latency and cost per successful task.

Is multi-agent delegation generally available?

No, Sol’s multi-agent feature is in beta.

Read the original

Continue at the source.

OpenAI

Companies: OpenAI

More on OpenAI

All OpenAI coverage →