GPT-6 Sol and GPT-6 Luna API: the GPT-6 lineup in three tiers

With GPT-6 Sol and GPT-6 Luna, the GPT-6 family on FastMetal now has three tiers. Both went live on September 23, 2026, as gpt-6-sol and gpt-6-luna, next to GPT-6 Astra, which was already in the catalog.

OpenAI ranks them Astra, Sol, Luna, from the top. Use Astra for the hardest work, Sol for everyday development, and Luna for high volumes where speed matters.

All three are billed per million input and output tokens, and the rates step down from Astra to Sol to Luna. Current rates are on the GPT-6 Astra, GPT-6 Sol and GPT-6 Luna pages. OpenAI has since released GPT-6.1 Sol; see GPT-6.1 Sol and Sol Pro for how it measured against GPT-6 Sol.

Luna costs less than the generation before it

GPT-6 Luna has lower input and output rates than GPT-5.6 Luna. A new generation that is priced lower is unusual. For classification, summarization and other jobs that run in large numbers, the difference shows up directly in the monthly bill.

What both models support

• A 1,050,000-token context and up to 128,000 output tokens.
• Image and file input.
• Tool calling and structured output.
• A settable reasoning effort through the reasoning parameter.

At launch we confirmed that both answered with training opted out. Their route has changed since, and its training policy is now listed as not declared. Each model page shows the current data policy.

The Pro variants: similar token rates, more tokens

gpt-6-sol-pro and gpt-6-luna-pro run the same models with OpenAI's reasoning set to pro. The per-token rates are close to the standard models', but each request costs more, because the upstream adds about 1,400 input tokens to every call. The shorter the question, the bigger that overhead is in relative terms.

We sent the same question ("What is the capital of Japan?") through our API to each model on September 23, 2026, and recorded the input tokens billed:

On gpt-6-sol-pro, most of the overhead was treated as cached from the second call on, so repeated calls cost less each. On gpt-6-luna-pro, three calls in a row showed no cache hits. Caching behavior is set by the upstream provider, and we do not guarantee it.

For large numbers of short questions, use the standard models. The Pro variants fit hard problems where better answers are worth the fixed overhead.

Calling it

It is OpenAI-compatible, so existing code only needs a new model value:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Swap in gpt-6-sol to check quality, or gpt-6-astra for hard tasks. The same API key and balance work for all of them.

Which one to start with

Start with Sol. If its output has quality to spare for your task, move down to Luna. Billing is prepaid: you load a balance by card, from $5, and you can never spend more than you loaded, so switching between models while you test cannot run up a surprise bill. See pricing, and everything else we serve in the model catalog.