GPT-6 Sol and GPT-6 Luna pricing tracker: the official table, four billing rules, four worked examples
The official model pages list per-million rates: GPT-6 Sol $2 input, $0.20 cached input, $2.5 cache writes, $10 output; GPT-6 Luna $0.1, $0.01, $0.125, $0.5. Only published unit rates and examples computed from that table appear here, re-checked 2026-09-29.
Updated 2026-09-29
The four things worth remembering
Sol $2/$10, Luna $0.1/$0.5
Official per-million rates, published both in the 2026-09-22 changelog entry and on the two model pages, where they match.
Above 272K the whole request moves to 2x and 1.5x
The official wording is for the full request: the multiplier applies to the entire call, not just the tokens past 272K. This is where budgets most often go wrong.
Batch and Flex sit at 50%
The same official sentence puts Fast at 2x the applicable rates. This page does not guess how Fast composes with the long-context tier — the official text stops short of that.
Four calls cost $0.40, $0.02, $1.50 and $0.20
The shape is 100,000 input and 20,000 output (the third uses 300,000 input, so the long-context tier applies; the fourth runs on Batch). All computed from the official table; the breakdown is below.
What separates the two, in money terms
The capability line is the same for both: reasoning models, input modalities text, image, identical context and output limits. The unit prices differ — Sol at $2 input and $10 output, Luna at $0.1 and $0.5, a factor of twenty when computed from the official table. The launch post also keeps the line GPT-6 Astra continues to be our best model across the board., so «cheaper» is not the same as «replaces the flagship».
Official wording: Sol and Luna are 50% below GPT-5.6 promotional pricing
OpenAI's Sol/Luna launch post says reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing. That sentence covers both models, but the only old rate this page could verify word for word is the gpt-5.6-sol row ($4.00 / $20.00, the 2026-08-21 changelog entry, which also labels it promotional and valid at least through 2026-11-21). So the halving is shown with old figures for Sol only; for Luna the official sentence is quoted and no old number is printed.
Public timeline
On 2026-08-21 the official changelog recorded GPT-5.6 Sol moving to $4 input / $20 output, the same entry labelling it a promotional price valid at least through 2026-11-21. It is the only pre-cut rate for the Sol line this page can quote verbatim.
On 2026-09-22 OpenAI released GPT-6 Sol and GPT-6 Luna, and the same changelog entry states the standard rates straight away — Sol $2 / $0.20 / $10, Luna $0.10 / $0.01 / $0.50 — for prompts with up to 272K input tokens.
On 2026-09-25 the changelog fixed an image-encoding bug that it says had degraded image understanding in GPT-6 Sol and GPT-6 Luna, and recommends rerunning evaluations that use image inputs. Prices did not move; what changed is how those requests behave.
The confirmed column and the not-verified column
Confirmed · verbatim on official pages
Confirmed by the official model pages, checked word for word on 2026-09-29: both carry a 1,050,000 context window, 922,000 maximum input and 128,000 maximum output, accept input modalities text, image, and support reasoning effort none / low / medium (default) / high / xhigh / max. Knowledge cutoffs: Sol Apr 20, 2026, Luna May 18, 2026. The price table has four rows — Input, Cached input, Cache writes, Output.
Not verified · not carried here
As of 2026-09-29 three things are left out. First, GPT-5.6 Luna's old rates: they exist only inside a data block of the pricing archive captured 2026-09-23, where four value sets (Standard / Batch / Flex / Fast) sit side by side, and the column assignment cannot be settled — so no figure is printed. Second, the rumour that Codex now defaults to gpt-6-sol: no official sentence says so, so it stays unverified. Third, the self-reported benchmark figures in the launch post: this page notes that they exist and does not transcribe them.
The two side by side
GPT-6 Sol · $2 input / $10 output
Officially aimed at complex coding and agentic work; 1,050,000 context, 128,000 maximum output, cached input $0.20, cache writes $2.5.
GPT-6 Luna · $0.1 input / $0.5 output
Officially the efficient model for focused, high-volume work; the same limits as Sol (1,050,000 / 128,000); cached input $0.01, cache writes $0.125.
Four billing rules, each checkable
One: cached input is priced at 10% of the uncached input rate and cache writes at 1.25x. Two: for prompts above 272K input tokens the official wording is for the full request — 2x on input and cache, 1.5x on output for the whole call, not only for the excess. Three: Batch and Flex are 50% of Standard rates, Fast is 2x the applicable rates. Four: regional processing adds a 10% premium where available, and EU data residency runs on Standard processing only. All four read the same on both model pages; this page does not guess how two of them compose when both apply.
How this works on QCode
As of 2026-09-29 gpt-6-sol, gpt-6-luna and gpt-6-astra are all callable on QCode — one key, one compatible endpoint, the id goes in the model field. This site's 30-day record from the morning of 2026-09-29 counts 12,682 calls on gpt-6-sol, 2,433 on gpt-6-luna and 26,832 on gpt-6-astra. Billing follows OpenAI's official table and the four rules above; the examples here are arithmetic, not a quote, and never a service-level promise.
Common questions
How much does Sol cost versus Luna?
Twenty times on unit price: Sol is $2 input and $10 output, Luna $0.1 and $0.5, with identical limits (1,050,000 / 128,000). Computed from the official table, a 100,000-in / 20,000-out call costs $0.40 on Sol and $0.02 on Luna.
How does the long-context tier work?
Above 272K input tokens the multiplier lands on the whole call: 2x on input and cache, 1.5x on output, for the full request. Worked example: Sol with 300,000 input and 20,000 output ⇒ input at $4, output at $15, total $1.50 — computed from the official table, not a figure OpenAI publishes.
What about Batch and Fast?
One official sentence covers both: Batch and Flex are priced at 50% of Standard rates, Fast at 2x the applicable rates. Sol on Batch, same shape (100,000 in, 20,000 out), computes to $0.20 against $0.40 at standard rates.
Did OpenAI state the size of the cut?
Yes — reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing, which covers both models. The only old rate verified here word for word is gpt-5.6-sol at $4 / $20, so Sol's row shows the previous figure and Luna's keeps just the official sentence.
Did image capability change?
There is an official record: the 2026-09-25 changelog fixed an image-encoding bug it says had degraded image understanding in GPT-6 Sol and GPT-6 Luna, and recommends rerunning evaluations and retrying affected workflows. This is separate from pricing.
And where does Astra fit?
The launch post keeps one line in place: GPT-6 Astra continues to be our best model across the board. Astra's official rates are $10 input, $1 cached input, $12.5 cache writes and $50 output, and the same >272K rule appears on its model page. Quota and migration for Astra live on a separate page; this one stays on Sol and Luna pricing.
Sources
The three OpenAI model pages for GPT-6 Sol, GPT-6 Luna and GPT-6 Astra (price tables and billing rules), the OpenAI changelog entries for 2026-08-21, 2026-09-22 and 2026-09-25, the Sol/Luna launch post (the 50% sentence and the line keeping Astra as the best model), the OpenAI pricing archive captured 2026-09-23, and this site's 30-day usage record from 2026-09-29.
One key, swap the model and compare it yourself
gpt-6-sol and gpt-6-luna are both callable on QCode as of 2026-09-29; prices and discount tiers follow OpenAI's official table.
Related
GPT-6 overview guide
Capabilities, endpoints and choice across the three GPT-6 models — the half that is not pricing.
GPT-6 Sol status tracker
Release, quota and status history for Sol — the companion to this pricing page.
GPT-5.6 Luna pricing tracker
The previous generation's Luna table and its promotional wording — the baseline OpenAI's 50% sentence refers to.
QCode is independent of OpenAI. This page carries only what the official pages state verbatim; every per-call total is our arithmetic from that table, labelled as a computation rather than a quote or a commitment. Where an old rate could not be assigned to the right column, the cell is left empty rather than guessed. Internal usage numbers come from this site's 2026-09-29 record and describe availability, never a service level.