Quick answer
- Astra
- $10 in / $50 out per MTok
- GPT-6.1 Sol
- $2 in / $10 out per MTok
- Luna
- $0.10 in / $0.50 out per MTok
- Verified
- September 30, 2026
Verified specs side by side
| Astra | GPT-6.1 Sol | Luna | |
|---|---|---|---|
| Base API price | $10 in / $50 out per MTok | $2 in / $10 out per MTok | $0.10 in / $0.50 out per MTok |
| Cached input | $1 | $0.10 | $0.01 |
| Context / output | 1,050,000 tokens (128K max output) | 1,050,000 tokens (128K max output) | 1,050,000 tokens (128K max output) |
| Position | Highest capability | Complex coding and agentic work | Efficient / high volume |
| Reasoning effort | low to max | low to max | none to max |
| Knowledge cutoff | April 30, 2026 | April 30, 2026 | May 18, 2026 |
| Released | September 3, 2026 | September 29, 2026 | September 3, 2026 |
| API model | gpt-6-astra | gpt-6.1-sol | gpt-6-luna |
GPT-6.1 Sol is the newer Sol, and OpenAI's docs now point to it. GPT-6 Sol, released with Astra and Luna on September 3, 2026, stays selectable in Codex at $2/$10 with $0.20 cached input. It is the Sol to keep where a workload depends on reasoning effort none, which Chat Completions function calling requires and GPT-6.1 Sol does not accept.
Your AI workspace for shipping software.
Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.
Download AIDEN freeFree to start · macOS 12+ · No credit card required
A practical routing rule
Start with GPT-6.1 Sol when the task is ambiguous or requires substantial reasoning; OpenAI recommends it for complex coding and agentic work and describes it as near-Astra performance at one fifth of Astra's token price. Use Luna after the workflow is well specified and repeatable. Escalate to Astra when a failed attempt costs more than its premium or when you explicitly need OpenAI's highest-capability tier.
The premium is easy to underestimate. Astra costs 5x GPT-6.1 Sol per input and output token and 10x per cached input token, so the gap widens in long agent sessions that reuse a cached prompt prefix. Luna costs 20x less than GPT-6.1 Sol per token.
What the benchmarks say
We have not verified a vendor-published coding benchmark for GPT-6.1 Sol or Luna, so this page does not rank the three on one. The cross-vendor figures available for the family cover Astra: 57.9% on Terminal-Bench 4.0 and 53.3% on FrontierCode v1.1, as reported by OpenAI in Anthropic's Claude Opus 5.5 launch table. That match-up is in Claude Opus 5.5 vs GPT-6 Astra. Reasoning effort, per-task cost and Codex setup are in the best Codex model guide.
Official sources
GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and GPT-6 Sol model documentation, plus OpenAI's latest-model guide. Data verified September 30, 2026.