GPT-6 Astra vs Sol vs Luna: which to use, and what it costs

GPT-6 comes in 3 tiers: Astra for the hardest work, GPT-6.1 Sol for most complex work at a fifth of Astra's price, and Luna for high-volume, focused tasks at a hundredth. On Epoch AI's index, Astra and 6.1 Sol score within each other's range of doubt, so for most teams the default is Sol, with Luna for the many small calls and Astra only where your own tests show Sol falling short.

All 3 share a 1,050,000-token context window, 128,000 output tokens, structured outputs, tools and prompt caching. What differs is price, a few API settings and how they hold up on long, hard tasks.

Side by side

Scores are from the Epoch Capabilities Index by Epoch AI (CC BY 4.0). Astra's 90% range runs from 163.4 to 171.2 and 6.1 Sol's from 163.2 to 170.8, so a general benchmark cannot tell them apart. Luna's 156.3 is lower but above GPT-5's 150. Each model's page has every price: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna.

What each is for

GPT-6.1 Sol is the everyday model for complex work: code, computer use, documents and analysis. OpenAI's own advice is to compare it with Astra on your tasks; its default reasoning effort is medium, so raising effort is the first step before paying for Astra.

GPT-6 Luna is for focused, high-volume calls: extracting fields, classifying, small edits, choosing which tool or model handles a request. OpenAI's ladder starts it at low effort for simple extraction and at extra-high effort for problems with clear constraints, which is still far cheaper than Sol.

GPT-6 Astra is for the most demanding work: long multi-step workflows across code, browsers and professional software. OpenAI reports that in several evaluations Astra reached stronger results with substantially fewer output tokens, so its cost per finished task can be closer to Sol's than its token price suggests. Measure that on your own work before assuming it.

GPT-6 Sol, the older Sol, costs the same as 6.1 Sol but charges twice as much for cached input and scores lower. New work should start on 6.1 Sol.

The 272K step

Every GPT-6 tier bills a prompt over 272,000 input tokens at 2 times the input and cache rates and 1.5 times the output rate, for the whole request. On Astra, a 270,000-token prompt with a 2,000-token answer costs $2.80; at 300,000 tokens it costs $6.15. 30,000 more tokens more than double the bill. If you pass whole documents or long histories, keep prompts under the step: retrieve the relevant parts, or compact the history.

Differences that break a model swap

• Reasoning effort. Astra and 6.1 Sol reject none and minimal; use low. GPT-6 Sol and Luna accept none.
• Tools in Chat Completions. Astra and 6.1 Sol support Chat Completions, but tool calling needs the Responses API. GPT-6 Sol and Luna call functions in Chat Completions only with reasoning_effort set to none.
• Sampling. With any reasoning effort other than none, remove temperature, top_p and top_logprobs.

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6.1-sol",
    reasoning={"effort": "low"},  # "none" returns an error on 6.1 Sol and Astra
    input="Summarise this support ticket in one sentence: ...",
)
print(response.output_text)

A worked month

A contract review assistant: 10,000 requests a day (300,000 a month), each with a 6,000-token instructions block served from the cache, 2,000 new input tokens and 1,500 output tokens, reasoning included (reasoning tokens are billed as output).

• Astra: 300,000 × (2,000 × $10 + 6,000 × $1 + 1,500 × $50) ÷ 1 million = $30,300
• 6.1 Sol: 300,000 × (2,000 × $2 + 6,000 × $0.10 + 1,500 × $10) ÷ 1 million = $5,880
• Luna: 300,000 × (2,000 × $0.10 + 6,000 × $0.01 + 1,500 × $0.50) ÷ 1 million = $303

If tests show Luna handles 70% of contracts (the short, standard ones) and Sol the rest, the month is about $1,976, a third of all-Sol. Cache writes, at 1.25 times the input price, are left out; they are paid once per cache lifetime, not per request.

Try itGPT-6 Astra, Sol or Luna: what a month costs

Requests a day

10,000

Input tokens a request

8,000

Input served from the cache

75%

Output tokens, reasoning included

1,500

Processing

StandardBatch or Flex (50% off)

GPT-6 Astra score 166.4$30,300

GPT-6.1 Sol score 166.3$5,880

GPT-6 Sol score 162.6$6,060

GPT-6 Luna score 156.3$303

Luna is 19 times cheaper than 6.1 SolUnder the 272K step

What to do

1. Start new work on GPT-6.1 Sol at medium effort; raise effort before you switch to Astra.
2. Move well-defined, high-volume steps to Luna once your tests pass. Our guide to model routing shows how to send each request to the cheapest tier that handles it.
3. Keep prompts under 272,000 tokens, and cache the fixed part: OpenAI's 12-hour prompt cache explains when the longer cache pays.
4. Compare with other makers: which model finds the cheapest model that scores as well as you need, and Claude Opus vs Sonnet vs Haiku does the same job for Anthropic's tiers.

Why thinking before answering helps on hard problems, and costs output tokens, is explained in Chain of ThoughtChain of ThoughtHaving a model write out the steps of its reasoning before it gives an answer. Once a prompting trick, now built into "thinking" models..