GPT-6 Sol and Luna pricing: API rates, sample calculations and ChatGPT plans

The GPT-6 Sol and Luna prices via API start from $2 per million tokens in input and 10 in output for Sol, and from 0.10 and 0.50 dollars per Luna, with Standard processing and requests under 272,000 input tokens. The input read from the cache costs one tenth, Batch and Flex cost half, Fast mode twice as much. On the other hand, ChatGPT and Codex plans do not pay per token: the two models are included in the subscription, with usage limits different from model to model.

OpenAI presents the new tariffs as a 50% cut from the promotional prices of GPT-5.6 (announcement of 22 September 2026). Here you will find the item by item list, comparison with GPT-5.6 and with Astra, three sample accounts made with official rates and the items that a quote tends to forget. Prices verified on 23 September 2026 on OpenAI official price list: Before committing a budget, double-check them.

How much do Sol and Luna cost via API: the eight list items

Each model has four entries (input, cached input, cached write, output) and two bands, depending on whether the request remains below or exceeds 272,000 input tokens. The values are those of Standard processing, in dollars per million tokens.

The rules behind the numbers are written in the tabs of GPT-6 Sol and GPT-6 Luna. Cache input costs 10% of normal input, caching 1.25 times. When a request exceeds 272,000 input, input and cache tokens double and the output goes up one and a half times, for all the request and not just for the excess part.

Four ways to call the same model, four rates

The same request can be processed in different ways, and the price changes with the mode. For those who use APIs it is the easiest lever: it does not require changing models or rewriting prompts.

Two notes from price list. Fast mode is the new name of the old Priority processing, changed on July 30, 2026. And for Sol and Luna the Data residency in the EU is only available with Standard processing, with a 10% surcharge that applies to models released from March 5, 2026 onwards: applied to the price list means $2.20 in input and $11 in output for Sol, $0.11 and $0.55 for Luna.

Data residency does not activate with a click. According to the OpenAI Data Guide must be requested from the sales team, set by project and, for regions outside the United States, requires approval of controls on monitoring abuse and a modified retention agreement; requests then pass from the address eu.api.openai.com. For a company that processes personal data of customers or employees it is a choice to be made before the estimate, with the picture described in Business Data and AI: What Happens to What You Write.

Compared to GPT-5.6 and Astra: what half price means

The ad's "half price" is accurate on Sol: input, cache and output of GPT-6 Sol are exactly half of GPT-5.6 Sol. On Luna the input halves and the output drops from 1.20 to 0.50 dollars, a cut of 58% that in the launch table still appears as "50% cheaper" (announcement). The comparison term, however, is a promotional price: the GPT-5.6 Sol card indicates that it is valid until at least November 21, 2026, so the comparison will have to be redone if the conditions of GPT-5.6 change.

With Astra the ratio is net and equal on each line: Sol costs a fifth, Luna a cent. Who today uses Astra for everything can ask which steps would hold with Sol, taking into account that OpenAI continues to indicate Astra as the model for the most demanding jobs; Astra's list with its rules is in how much GPT-6 Astra costs. Those who stayed on GPT-5.6 Terra will notice that Sol costs equal in input and less in output, and the comparison of the old family remains in GPT-5.6 Sol vs Terra.

Three sample accounts with official prices

Volumes and tokens are declared hypotheses, chosen to show how the list items are combined: they are not measures made on a company. All prices are those of the official pricing in dollars.

Example 1: an internal assistant on Sol, with and without cache. Hypothesis: 20,000 requests per month to the assistant of the technical office. Each request has a fixed portion of 8,000 tokens (instructions, tool definitions, manual extracts) and 1,000 variable tokens, and produces 800 output tokens including reasoning. With the cache and an explicit break point after the fixed part, we assume 600 scriptures per month and 19,400 readings.

The cache more than halves the account, and after the cache the heaviest entry becomes the output: two thirds of the expense. The next lever is therefore the length of the answers and the level of reasoning, not the price of the input.

Example 2: a 300,000-token specification on Sol. Assumption: a tender document with 300,000 input tokens and 10,000 output tokens for the analysis. A single request crosses the threshold and pays the long-context rate: 0.3 million × $4 = $1.20 plus 0.01 million × $15 = $0.15, or $1.35. Split into two requests of 150,000 input tokens and 5,000 output tokens each, they remain in the short-context tier: $0.60 for input and $0.10 for output, $0.70 in total. For a hundred documents per month, that is $135 versus $70. However, splitting has a hidden cost: if clauses in the second half refer back to clauses in the first, the model does not see them together. It is often better to remove unnecessary attachments first and stay below the threshold with a single document.

Example 3: The nightly enrichment of the catalog on Luna. Hypothesis: 200,000 product cards per month to be completed with attributes and short texts, 1,200 input tokens and 300 output each, i.e. 240 million input tokens and 60 million output tokens.

Night work is the ideal candidate for Batch, who returns results within 24 hours. However, if data is to remain in Europe, with Luna Batch it is not available and you pay 120% more than Batch. On Luna the difference is worth a few tens of dollars; with larger volumes or with Sol, where each entry costs twenty times as much, the same choice moves figures that must be put to budget. Method and formulas are in cost of AI APIs: how to really calculate.

What do you pay if you use Sol and Luna in ChatGPT or Codex

In OpenAI apps there is no token counter on the bill: you pay the plan, and each model consumes a different share of the limits included. Codex price page It makes it clear that ChatGPT Work and Codex share prices, credits and usage limits. Prices are in dollars as the page shows them: at the time of purchase check the amount in euros.

The same page estimates how many local messages you send in a five-hour window: with Plus, from 15 to 150 with Sol, from 350 to 3,000 with Luna, from 5 to 45 with Astra, against 10-100 of the old GPT-5.6 Sol. They are estimates, not fixed limits, and may also be worth weekly limits. Beyond the included limits you pay in credits: per million tokens Sol consumes 50 credits in input, 5 in cache and 250 in output, Luna 2,5, 0.25 and 12.5, without a separate charge for writing in cache. Fast mode in Codex consumes 2.5 times the Standard credits (speed in Codex).

Many SMEs need both: subscriptions for people who use Work and Codex, APIs for processes that run on their own. If you wonder which model to put on which process, the reasoning is in GPT-6 Sol vs Luna; how you set the model inside Codex is explained in Codex with GPT-6 Sol.

The Voices That Escape Quotes

• Reasoning is paid for as an output. A high level on thousands of requests can weigh more than any input.
• Unnecessary cache writes. A block remains reusable for at least 30 minutes from the last use: if no one reads it back at that time, the writing cost an extra quarter for nothing (cache guide).
• The threshold of 272,000 tokens. An extra attachment can pass all the request at the long rate.
• Fast mode left on. In APIs it costs twice as much, in Codex it consumes 2.5 times the credits.
• EU residency excludes Batch and Flex. It must be decided before estimating, not after.
• The comparison with a promotional price. Savings on the old GPT-5.6 Sol are calculated on promotional rates, available until at least November 21, 2026.

Cost items that are not on the list (human review, failed attempts, maintenance) are collected in the hidden costs of AI projects.

How to Estimate Spending Before Signing

1. Spin fifty or a hundred real cases with the pattern and level of reasoning you would really use.
2. Read the actual tokens of input, cache and output that the API reports for each call, and checks in the cache panel how much of the input is reused.
3. Separate what can wait from what needs to be answered immediately: the first group goes on Batch or Flex.
4. Decide if you need EU residency before doing the math, because it changes the modes available.
5. Divide spending on completed tasks, including those made by hand. The method is in price per million tokens, explained.

With Sol and Luna the price per token drops, but the final account decides it caches, thresholds and processing modes more than the list. If you want an estimate made on your volumes before choosing between subscriptions and APIs, the Free AI Audit It starts with the processes and numbers you already have.