OpenAI publishes a model guide for its GPT-6 family

OpenAI has spent the past month shipping GPT-6 models at a steady clip. Now it has published the instruction manual.

On October 2, 2026, the company released “A model guide for the GPT-6 family.” The document targets developers and startups who need to decide which model to use, how much thinking to ask of it, and how to move a workflow from prototype to production.

Four models, three tiers, one menu

The guide sorts the GPT-6 lineup by workload and reasoning capability. At the top sits Astra, which OpenAI positions as its flagship for complex reasoning, with state-of-the-art performance in fields such as coding and science.

Astra launched on September 3, 2026. It posted a 98% score on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.

Next come Sol and Luna, which launched on September 22, 2026. Their selling point is price. API rates for both come in 50% lower than their GPT-5.6 counterparts.

Sol costs $2 per million input tokens and $10 per million output tokens. Luna goes much lower, at $0.10 per million input tokens and $0.50 per million output tokens.

Then there is 6.1 Sol, released on September 29, 2026. OpenAI says it delivers near-Astra performance at one-fifth of Astra’s cost.

All GPT-6 models share a context window of 1.05M tokens, with a maximum output of 128K tokens.

The knobs developers can turn

Model selection is only part of the guide. OpenAI also covers adjustable reasoning effort, prompt improvement, skills, tool coordination and production preparation.

The guide also highlights asynchronous tool calling. With this feature, a model can keep working on other parts of a task while it waits for an external tool to finish running.

Enhanced caching rounds out the efficiency features. Along with adjustable reasoning, it is designed to support rapid iteration and smoother deployment across different industries.

From launch sprint to playbook

The GPT-6 rollout unfolded fast. Astra arrived first on September 3, Sol and Luna followed on September 22, and 6.1 Sol landed on September 29.

The reference point throughout is GPT-5.6, the generation that Sol and Luna undercut on price by half.

What this means for startups and the AI market

The clearest takeaway is price compression. A 50% cut on Sol and Luna, plus near-flagship performance at one-fifth of Astra’s cost from 6.1 Sol, lowers the cost of running AI features at scale.

There are trade-offs to watch. Benchmark results like Astra’s 98% on FrontierMath Tier 4 measure performance on specific test sets, and production workloads tend to be messier. Developers will want to confirm that near-Astra performance from 6.1 Sol holds up on their own data before they commit.