GPT-6 Luna
pricing, context & API guide.

Cached input refers to tokens successfully read from a previously cached prefix. Cache writes are tokens written into the cache and have a different billable rate. The rates cannot be applied interchangeably; inspect actual usage fields from the API to see what was billed.

The table and calculator below use Standard direct API pricing, not a guaranteed invoice. Web search, computer use and other billable tools can add fees. OpenAI lists a 10% regional-processing premium where applicable; tax and operational costs are outside token prices. Check current provider pricing ↗.

OpenAI documents a strictly greater than 272,000 input-token boundary. At 272,000 input tokens or fewer, the shorter schedule applies. At 272,001 tokens and beyond, Standard pricing changes to 2× input, 2× cached input, 2× cache writes and 1.5× output across the whole request. This is a request-level tier switch, not an incremental overage fee.

The sharp price step means applications hovering near 272K input tokens need careful context budgeting. Consider retrieval, document chunking, compaction and prompt deduplication when doing so preserves the information required for correctness. A larger window can process large inputs, but it does not guarantee perfect recall, and a small reduction in prompt length does not necessarily reduce task failure.

One million tokens of context is capacity—not the cheap-rate allowance. Luna can accept long text or image-grounded input within its 1,050,000-token context, but the extended schedule must be included in estimates for large documents, repositories and agent histories.

Worked GPT-6 Luna cost examples

Worked examples hold measured token counts constant so the arithmetic can be compared. They do not predict model-quality differences or actual outputs. For low-volume budgets, remember that tool calls, retries and operations can outweigh the base Luna token rate.

Headline rate ratio: at short-context Standard list rates, Luna's uncached input and output are 20× cheaper than GPT-6.1 Sol and 100× cheaper than Astra. Those ratios apply to the stated token price categories, not to end-to-end task cost or quality. More retries, larger output, extra tools, and lower task success can erase savings.

When comparing “Luna vs Sol,” use the right release: GPT-6 Sol was the original middle tier, while GPT-6.1 Sol is OpenAI's newer recommended model. This page compares against 6.1 Sol rather than treating the names as equivalent. See also the GPT-6 family hub for the broader model-selection map.

from openai import OpenAI

client = OpenAI()  # Set OPENAI_API_KEY in your environment
response = client.responses.create(
    model="gpt-6-luna",
    reasoning={"effort": "medium"},
    input="Classify this support message and explain the label briefly.",
)
print(response.output_text)

Prompting advice: keep instructions explicit, provide a handful of representative edge cases when useful, separate data from instructions and require a verifiable output contract. For repeated system prompts, keep prefix content stable to improve the chance of cache reuse, then confirm hits in usage reports.

Parameter compatibility: with reasoning effort other than none, OpenAI advises removing temperature, top_p and certain token-logprob settings that are not supported. Consult the GPT-6 migration guide before copying parameters from older integrations.

SXF retains its separately maintained model-evaluation observations below, including evaluator names, benchmark versions, reasoning settings and evidence links. Use those cards to inspect individual measurements rather than treating one score as a universal ranking against Sol or Astra.

For purchasing or architecture decisions, compare real workflows under identical conditions: same prompt data, context length, reasoning effort, allowed tools, success criteria and privacy policy. Record cost per accepted task and latency per accepted task. A cheaper per-token model is valuable only if it performs the required work reliably.

Long-context benchmarks are especially easy to overgeneralize. A model can accept large input while still missing relevant details or making unsupported inferences. Check source citation fidelity and exact retrieval, not just whether it can ingest the document.

Method: price facts and specifications are sourced to OpenAI. Illustrative request costs are calculated from published Standard price fields. Workload recommendations are editorial interpretations, not a new vendor benchmark. The independently tracked SXF change history, recent signals and evaluation evidence remain attached to this model page.