← Back to feed

OpenAI published a practical model selection guide for the GPT-6 family on October 2, the first documentation to formally position all three models as a tiered production stack rather than a research preview.

The Pricing Cascade

The guide confirms the full pricing matrix:

Model	Input	Output	Cached Input
GPT-6 Astra	$10.00/M	$50.00/M	$1.00/M
GPT-6.1 Sol	$2.00/M	$10.00/M	$0.10/M
GPT-6 Luna	$0.10/M	$0.50/M	$0.01/M

The output ratio between models is exactly 100:20:1 (Astra:Sol:Luna). Cache discounts are 90% for Astra and Luna, 95% for Sol — making Sol disproportionately attractive for any workload that reuses large system prompts, tool definitions, or reference material across requests.

At $0.10 per million cached input tokens, Sol running agents over a 200K-token codebase context costs roughly $0.02 per cache hit. For teams processing thousands of agent turns per day against stable repositories, that gap compounds fast.

Model Selection Logic

The guide frames selection around three factors: capability, cost, and latency. The model hierarchy is:

• Astra: “Most intelligent for the best results.” Use for multi-step workflows requiring maximum reasoning across code repositories, databases, and external APIs.
• Sol: “Near-Astra intelligence at a fifth of the price.” Use for production workloads where Astra-level quality is not strictly required and throughput matters.
• Luna: “Fast and efficient at scale.” Use for high-volume, lower-complexity tasks where latency and per-task cost dominate.

The guide does not quantify the intelligence gap between Astra and Sol with specific benchmark figures, but external evaluations place them at roughly 82.2 versus 81.6 on LiveBench overall — a narrow 0.6-point spread at one-fifth the output cost.

Production Patterns

Four operational areas get the most attention:

Prompt caching: Stable content (instructions, tool definitions, reference material) should precede dynamic content so the cache layer can reuse it. OpenAI claims cached input tokens cost up to 95% less than uncached, depending on model. A caching diagnostics dashboard is available in the developer console.

Compaction: For long-running conversations or agent loops, compaction reduces context size while preserving task state. This is the guide’s mechanism for avoiding context window exhaustion without discarding critical state.

Reasoning effort: Developers can tune the reasoning budget per request, trading cost and latency against output quality. The guide treats this as a first-class dial, not an afterthought — implying GPT-6 reasoning is expensive enough to warrant per-task calibration at scale.

Long-running work with delegation: Multi-step workflows should use steering (mid-task course correction), async tool calling (non-blocking tool execution), and agent delegation (subagent spawning) to keep independent work parallel. The guide recommends setting explicit criteria for when models should pause and request human input rather than proceeding autonomously.

Why This Matters

Publishing a production guide is a different signal than launching a model. OpenAI has had GPT-6 Astra since September 3 and GPT-6 Sol and Luna since September 22. A formal operator playbook means the family is now considered stable enough to build production systems on, not just run demos.

The guide’s emphasis on caching economics is pointed. At the scale that matters to enterprise customers — millions of agent turns per day — the difference between paying $1/M (Astra cached) and $0.10/M (Sol cached) for cached input tokens is not marginal. For any team running GPT-6 agents over large, stable contexts, Sol’s cached input cost is one-tenth of Astra’s cached rate, with an even wider gap on uncached input ($2/M vs $10/M).

The Luna tier ($0.01/M cached) remains the wild card. Nothing in current evaluations places Luna in contention with Sol or Astra on complex reasoning, but for classification, routing, or structured extraction tasks where a smaller model suffices, $0.01 per million cached tokens is a different category of economics.

Benchmark Context

Available external evaluations on LiveBench (September 2026 run, max effort):

• GPT-6 Astra: 82.2 overall, 57.3 agentic coding
• GPT-6.1 Sol: 81.6 overall, 54.5 agentic coding

No comparable independent benchmark data for Luna. GPT-6.1 Sol’s DeepSWE v1.1 score of 75.2% at high effort surpasses GPT-6 Sol’s 68.8% at max effort, at roughly 76% lower cost — data published by OpenAI at DevDay on September 29.