TL;DR

  • Sol lists at $2 and $10 per million input and output tokens. Luna lists at $0.10 and $0.50.
  • Both share a 1.05M token context window, 128K maximum output, vision, and the tool surface.
  • Luna is built for high-volume focused work such as extraction, classification, and simple coding. Sol is built for complex coding and agentic workflows.
  • On Sol, prompts over 272K tokens bill input at double and output at 1.5x, and Batch and Flex run at 50%.
  • A common split routes the bulk of requests to Luna and escalates the hard tail to Sol.

OpenAI shipped GPT-6 Sol and GPT-6 Luna on the same day, September 22, 2026, and the pair is designed to be used together rather than picked between. The announcement calls them two balances of capability and cost. The short version: Sol is the complex-work model at $2 and $10 per million input and output tokens, Luna is the high-volume model at $0.10 and $0.50, and the 20x input-price gap is the whole decision.

The two models at a glance

Dimension GPT-6 Sol GPT-6 Luna
Input price, per 1M tokens $2 $0.10
Output price, per 1M tokens $10 $0.50
Context window 1.05M tokens 1.05M tokens
Maximum output 128K tokens 128K tokens
Built for complex coding, reasoning, agentic workflows high-volume focused tasks: extraction, classification, simple coding
Extras vision, function calling, web search, file search, computer use same tool surface

Per OpenAI’s models overview, the hardware of the two models is the same. The difference is the capability OpenAI tuned into each tier, and what it costs to run your traffic there.

What Luna is for

Luna is the volume tier. OpenAI’s docs position it for high-volume, focused tasks: extraction, classification, and simple coding. The announcement also notes Free and Go users can try GPT-6 Luna in the desktop app, which fits its role as the everyday model of the family.

The price is what makes it interesting. A workload that sends 10K input tokens and generates 2K output tokens per request costs about $0.002 on Luna. On Sol the same request costs about $0.04. At consumer-product volume, 10K requests a day is roughly $20 a day on Luna versus $400 on Sol, so anything Luna can do well should not be running on Sol.

What Sol is for

Sol is the complexity tier. It is OpenAI’s model for complex coding and agentic workflows, built to carry much of Astra’s strength into work at scale, and it is the model you escalate to when a task needs multi-step reasoning, sustained tool use, or hard generation quality.

Sol also has the family’s pricing machinery to learn before you commit. Prompts over 272K input tokens bill input and cache rates at double and output at 1.5x, so big-context work costs more than the headline rate suggests. Batch and Flex processing run at 50% of standard, which fits overnight enrichment and other async jobs. Fast mode costs 2x for latency-sensitive calls.

The full Sol breakdown, including availability in ChatGPT Work and Codex, is in GPT-6 Sol explained.

A practical split

The economical architecture routes by task shape:

  1. Send extraction, classification, formatting, and simple generation to Luna.
  2. Send multi-step agent runs, complex coding, and anything with expensive failure modes to Sol.
  3. Move anything asynchronous to Batch or Flex on Sol, and keep prompts under 272K or model the step rate.

The routing rule can start as simple as a classifier or a prompt-shape check, and it improves with logs: track which tier handled each request and what happened after, then move the boundaries. The failure mode to watch is silent drift upward, where more and more traffic lands on Sol because nobody is measuring the savings Luna was supposed to deliver.

When to skip both and call Astra

One tier sits above this pair. GPT-6 Astra lists at $10 and $50 per million input and output tokens, and OpenAI positions it for the hardest end-to-end reasoning and professional work. Sol and Luna are described as carrying much of Astra’s strength at lower prices, which means most products should exhaust Sol before paying Astra rates. Astra earns its price on the rare tasks where Sol’s output still needs rework by a specialist.

And if the comparison in play is across vendors rather than within OpenAI, Claude Opus 5.5 vs GPT-6 Sol covers the same-day Anthropic release, with the Opus family’s own flagship-versus-daily-driver split in Opus 5.5 vs Sonnet 5.

Sources

FAQ

What is the difference between GPT-6 Sol and GPT-6 Luna?
Positioning and price. Both were released September 22, 2026 with the same 1.05M token context window and tool surface. Sol is built for complex coding and agentic workflows at $2 and $10 per million input and output tokens. Luna is built for high-volume focused tasks at $0.10 and $0.50.
Is GPT-6 Sol worth 20 times the price of GPT-6 Luna?
For complex, multi-step work, that is the bet OpenAI is making. For extraction, classification, and simple generation, Luna is the better deal. Teams that route by workload keep most of the quality at a fraction of the cost.
Can one product use both GPT-6 Sol and GPT-6 Luna?
Yes. Both are served through the same API surface, so a product can route routine requests to Luna and escalate complex or agentic requests to Sol with a model-id change.
When should a team use GPT-6 Astra instead?
OpenAI positions Astra, at $10 and $50 per million input and output tokens, for the hardest end-to-end reasoning and professional work. It is the escalation tier above Sol, not the default.