How Much Does 1 Million GPT-6 Tokens Cost?

The first time I quoted a client for an API-backed feature, I told them "about a dollar a month" and missed by two orders of magnitude. The mistake everyone makes: treating "1 million tokens" as one number. It's three numbers — how much is input, how much is output, and which tier you're standing on.

So here's the direct answer, all short-context Standard rates per 1M tokens: GPT-6 Luna is $0.10 input / $0.50 output. GPT-6.1 Sol is $2.00 / $10.00. GPT-6 Astra is $10.00 / $50.00. A realistic mixed million — say 700K input, 300K output — costs $0.22 on Luna, $4.40 on 6.1 Sol, and $22.00 on Astra. Same million tokens, a 100× spread. That's the whole story in one sentence.

The per-million price card

Everything below is per 1M tokens, short-context (requests up to 272K input tokens), Standard processing:

Prices verified Oct 21, 2026 — always confirm on the official pricing page before budgeting. Note the pattern: output is priced at exactly 5× fresh input on every tier. That's not a coincidence — it's the single fact that matters most for output-heavy workloads, because output is the most expensive token you'll ever buy.

A real million: the 70/30 mix

Nobody buys a million tokens of one kind. A typical RAG chatbot does roughly 70% input, 30% output. Here's that million:

See that sliver for Luna? That's not a rendering bug. If your workload doesn't genuinely need Astra-class reasoning, choosing the tier wrong is the most expensive line item on your invoice — bigger than caching, bigger than batching, bigger than every optimization trick combined. My 10 ways to cut your bill piece goes through the full list, but tier choice is where 90% of the money is.

Three things that move the number

The sticker price is only the starting point. Three levers change what a real million costs:

• Prompt caching. Cached input bills at a tenth of fresh input on Luna and Astra ($0.01 and $1.00 per 1M), and one-twentieth on 6.1 Sol ($0.10 vs $2.00 — halved when 6.1 launched). Take that 6.1 Sol million with 90% of input cached: 900K × $0.10 + 100K × $2.00 = $0.29 on the input side instead of $2.00. That's 85.5% off, and you changed nothing but a stable system prompt.
• Batch and Flex at 0.5×. Work that tolerates slower turnaround — nightly backfills, evals, dataset labeling — bills at half price. The Astra 70/30 million drops from $22.00 to $11.00. If the answer isn't needed in seconds, full price is optional.
• Long-context pricing above 272K input tokens. This is the trap door. Past 272K input tokens in a single request, the input rate roughly doubles and output rises 50%: Astra's 1M-token single input bills at the long-context rate of $20.00/M, not $10.00. Spread your usage across many small requests and you never touch this tier.

What it looks like on a monthly bill

Tokens per million are abstract; here's a concrete workload: a support chatbot doing 3M input + 1M output tokens per month, uncached, Standard:

Notice the output column doing most of the damage: output is 5× input per token, so the 1M output side of the chatbot costs more than all 3M input tokens combined on every tier. Cap max_tokens, ask for tight answers, prefer structured JSON over chatty prose — the cheapest output token is the one never generated. (If you're budgeting this for real, our cost estimation guide walks through the formula and the four gotchas that eat naive estimates.)

The bottom line

1M GPT-6 tokens costs anywhere from $0.10 to $50.00 depending on tier and token type — a range so wide the per-million rate is almost meaningless without the workload attached. Anchor on the 70/30 mix: $0.22 (Luna) / $4.40 (6.1 Sol) / $22.00 (Astra), then apply your caching and processing-mode discounts on top. And the number I wish someone had given me on day one: a million tokens per month is about 33K per day — a small internal tool, not a production chatbot. Scale your mental model accordingly.

FAQ

Is 1M tokens in one request realistic? GPT-6 tiers have ~1.05M context windows, so technically yes — but cross the 272K input-token threshold and long-context pricing kicks in (roughly 2× input, 1.5× output). One giant request is usually the most expensive way to spend a million tokens.

Which tier should I pick for bulk processing? If the task is classification, extraction, summarization, or translation at volume, Luna at $0.22 per mixed million is the default — it's 20× cheaper than 6.1 Sol and 100× cheaper than Astra. Reserve the expensive tiers for requests that demonstrably need them.

Does prompt caching really change the per-million math? Yes — it's the single biggest discount after tier choice. On 6.1 Sol, a million input tokens with 90% cached costs $0.29 instead of $2.00, an 85.5% saving on the input side, with zero quality tradeoff.

Sources

• OpenAI API pricing — Luna / 6.1 Sol / Astra per-token rates, cache rates, Batch and Flex multipliers, 272K long-context threshold (official)
• GPT-6.1 Sol API docs — $2.00/$0.10/$10.00 rates and model details (official)
• GPT-6 Luna API docs — $0.10/$0.01/$0.50 rates and model details (official)

Updated Oct 21, 2026. This is an independent analysis, not affiliated with OpenAI.