GPT-6 model routing: Astra vs Sol vs Luna
Three tiers, a 100x price spread, one routing decision. GPT-6 Astra, Sol, and Luna went live September 22, 2026 — and the frontier has split into specialized tiers instead of one model for everything. Stop standardizing on a single model. Route by stage.
Three tiers, confirmed pricing
| Model | Input / 1M | Output / 1M | Positioning |
|---|---|---|---|
| GPT-6 Astra | $10 | $50 | Flagship — hardest reasoning |
| GPT-6 Sol | $2 | $10 | Efficiency — reasoning & coding |
| GPT-6 Luna | $0.10 | $0.50 | Ultra-efficiency — high volume |
All three are live in the OpenAI API as of September 22, 2026. The spread is the story: Luna's input price is two orders of magnitude below Astra's. That kind of gap doesn't reward picking a favorite — it rewards routing: sending each stage of a task to the cheapest tier that clears its bar.
Astra requires the Responses API — Chat Completions won't serve it. Sol and Luna work with both, but OpenAI's direction is clearly toward Responses (the Agents API builds on it), so migrating now avoids a forced move later. Test the migration on non-production traffic first.
Route by stage, not by habit
🌙 Luna — triage & volume
$0.10/$0.50. Classification, triage, first-pass drafting, monitoring, extraction — tasks where the work is simple, failures are cheap, and volume dominates. The bar: measure completion rate, because a cheap model that needs three retries isn't cheap.
☀️ Sol — reasoning & coding
$2/$10. The default for agent work: coding, multi-step reasoning, tool use. 5x cheaper than Astra on both input and output. If you're standardizing on one model today, make it Sol — then promote selectively.
✦ Astra — architecture decisions
$10/$50. The flagship, reserved for the hardest reasoning: architecture choices, ambiguous requirements, high-stakes review. Leads on Terminal-Bench-Science and AutomationBench. Use sparingly, where the capability premium demonstrably pays.
This is the same posture our Astra vs. Opus comparison recommends, and it's how governed missions are designed: the contract bounds cost per stage, and model choice follows the evidence. A mission that routes triage to Luna, execution to Sol, and review to Astra will beat a mission that runs everything on the flagship — on cost, with no capability loss where it counts.
Price per token is not price per task
The routing table only works if you measure the right thing. A $0.10/M model that completes 60% of tasks first-try, needs two retries, and demands heavy verification can easily cost more per completed task than a $2/M model at 95% first-try. The efficiency tiers make the economics of agents interesting — but only when you benchmark completion rate on your tasks, with your verification costs included.
Practical sequence: instrument cost-per-task measurement now, route a slice of volume traffic to Luna, compare completed-task cost against Sol, and promote or demote per stage on evidence. DevDay may add tiers or move prices — the routing discipline survives any price list.
Frequently asked questions
How much do GPT-6 Astra, Sol, and Luna cost?
Per million tokens input/output: Astra $10/$50, Sol $2/$10, Luna $0.10/$0.50. Live in the OpenAI API since September 22, 2026.
Which GPT-6 model should I use for coding agents?
Sol is the default for reasoning and coding at $2/$10. Reserve Astra for architecture decisions and the hardest reasoning. Benchmark both on your tasks before standardizing.
Do I need to migrate APIs for GPT-6?
Astra requires the Responses API. Sol and Luna work with both Responses and Chat Completions, but the platform direction is toward Responses — migrate on your schedule, not OpenAI's.
Is the cheapest model the cheapest completed task?
Not necessarily. Measure completion rate per task including retries and verification — a cheap model with a low first-try rate can cost more per completed task than a pricier one.
Route with evidence
💰 Cost per task
The full math: break-even, retries, verification cost, long sessions.
📊 Astra vs Opus 5.5
The flagship head-to-head: price and performance, September 2026.
How this guide is verified
Pricing figures on this page reflect publicly reported September 2026 OpenAI pricing for the GPT-6 lineup (Sol, Luna, Astra) — Sol at roughly $2 input / $10 output per million tokens, Luna at roughly $0.10 / $0.50 — reported after the models’ launch and consistent across multiple secondary sources. Capability notes (Astra’s Responses-API-only tool calling, key-governance controls) reflect the current public documentation. Treat prices as directional and confirm on OpenAI’s pricing page before committing spend. This page will be refreshed after the September 29, 2026 DevDay keynote with anything OpenAI confirms.