GPT-6 Sol and Luna for Product Managers: The Complete Tier Guide

GPT-6 is OpenAI's first model family to ship with explicit commercial tiers from launch day. Astra (September 3) set the capability ceiling. Sol and Luna (September 22) fill the volume and cost positions below it. This is a deliberate portfolio play: every price point, from maximum reasoning to maximum throughput, now has a GPT-6 SKU.

The GPT-5.x generation (5.5, 5.6 Sol, 5.6 Terra Luna) used incrementally different model IDs that required individual evaluation to understand capability tradeoffs. GPT-6 makes the hierarchy explicit in the name. This matters for product architecture: you can now design routing rules around a named tier rather than a specific model ID.

Luna is OpenAI's answer to a question every high-scale AI product team eventually asks: "What is the cheapest capable model for tasks that have a clear right answer?" At $0.50 per million tokens, Luna is approximately 4x cheaper than Sol and represents OpenAI's direct competition to high-volume open-weight alternatives.

What Luna is explicitly designed for, in OpenAI's own framing: "high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions." That phrase is a product specification, not just marketing copy. Luna was tuned for reliability on well-defined tasks, not for open-ended reasoning. If your use case has a ground truth to measure against, Luna is the right starting point.

Sol sits between Astra and Luna in the capability-cost curve. At $2/$10 per million input/output tokens, it is priced similarly to where GPT-4 class models were in 2024, but it ships with GPT-6 generation quality. Sol is the tier that most production AI features should target in late 2026.

The practical heuristic: use Sol as your default production model for new AI features. Then route down to Luna for any task your eval confirms is well-defined, and route up to Astra for any task where Sol produces outputs that require substantial human correction.

The GPT-6 family creates a 20x cost range within a single provider. Luna at $0.50/mtok versus Astra is approximately a 20x difference on equivalent tasks. For a product processing one billion tokens per month, routing 70% of traffic to Luna versus Sol can mean the difference between a $500K and a $2M annual inference bill.

The catch: routing logic is not free. You need evals to confirm Luna is producing acceptable quality on each task type before you route to it in production. A Luna routing decision that produces 5% more user-corrected outputs eliminates the cost savings fast. Build your eval harness before you build your router.

A routing architecture assigns each request type in your product to the appropriate GPT-6 tier at call time. The goal is minimum cost at acceptable quality, not maximum quality everywhere. Here is a practical framework for building one.

Products currently running on GPT-5.5, GPT-5.6 Sol, or GPT-5.6 Terra Luna should evaluate migration to the GPT-6 tier. The GPT-6 generation ships with improved instruction following, better structured output reliability, and a longer context window. The migration is not automatic: GPT-6 models may produce different outputs for the same prompt, and your existing evals need to run clean before you promote.