GPT-6 Sol vs GPT-6 Luna: Pricing, What Each Is For, and Which One You Actually Need

2026-09-23 · 5 min read

openaigptmodel-releaseai-tools

On September 22, 2026 — about three weeks after GPT-6 Astra landed as the flagship — OpenAI shipped the rest of the generation: GPT-6 Sol and GPT-6 Luna. Neither is a bigger model. Both are cheaper ones, and that is the whole point of the release.

The headline number is price. Each new model costs roughly half of the GPT-5.6 variant it replaces, while scoring better on OpenAI's own evaluations. That combination is rare enough to be worth a few minutes of your attention, because it changes which model is the correct default for most everyday work.

What actually changed

Model	Input / 1M tokens	Output / 1M tokens	Predecessor price
GPT-6 Astra	$10	$50	— (flagship)
GPT-6 Sol	$2	$10	GPT-5.6 Sol: $4 / $20
GPT-6 Luna	$0.10	$0.50	GPT-5.6 Luna: $0.20 / $1.20

Both new models carry a ~1.05M token context window with up to 128K output tokens, and both were trained with the methods OpenAI developed for Astra — which is why the gains land on the same axes Astra improved: factuality, coding, computer use, and professional work. OpenAI's specific claim is that GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol on its factuality evaluation, with lower error rates on coding too.

Two details worth knowing before you budget:

Requests over 272K input tokens are billed at 2x input and 1.5x output. The million-token window is real, but it is not priced linearly.
Knowledge cutoffs differ slightly: Sol is trained to April 20, 2026, Luna to May 18, 2026.

On availability, Sol went to ChatGPT Work, Codex, and the API; Luna also reaches Free and Go users in the desktop app. Rollout was gradual through launch day.

Sol or Luna: the actual dividing line

The tier names describe compute, not quality, and the honest split is about whether your task needs reasoning or just reading.

GPT-6 Luna is for volume work where the answer is already in the input. Summarizing a document, extracting fields from a contract, classifying support tickets, answering a question from a retrieved passage, tagging, reformatting. At $0.10 per million input tokens, you can push volumes through Luna that would be economically absurd on a flagship — and for this class of task the flagship would not give you a better answer anyway, because there is no hard reasoning step to win.

GPT-6 Sol is for work where the model has to figure something out. Writing and debugging code, multi-step analysis, drafting something that requires judgment about what to include, following a long instruction set without drifting. This is where the halved error rate matters: on this kind of task, mistakes are expensive to catch downstream, so a model that is wrong half as often is worth more than the 20x price gap over Luna suggests.

GPT-6 Astra stays the pick only for long agentic runs. Astra's advantage shows up on extended, multi-step, tool-using jobs — the sort that run for minutes and chain a dozen decisions. If your task finishes in one or two turns, you are paying 5x Sol's price for reasoning depth that never gets exercised. The same logic applied to picking among the three GPT-5.6 variants, and it applies more sharply now that Sol is half the price it used to be.

A practical default: start on Luna, move the tasks that fail up to Sol, and reserve Astra for the agentic workflows that measurably need it. Almost everyone's instinct runs the other way — pick the best model, then economize later — and that instinct is what makes AI bills surprising.

How to try them without a subscription

Straight from OpenAI, Luna is the one you can reach on a free ChatGPT account; Sol sits behind paid tiers or the API, where you pay per token from the first call.

There is also the multi-model route, which is the only way to run a real side-by-side. aiwith.chat gives new accounts free credits and no ChatGPT subscription requirement. Being straight with you about what's live there today: GPT-6 Astra is available now as the default of its GPT-6 robot, alongside GPT-5.6 Sol / Terra / Luna, GPT-5.5, Claude and Gemini. GPT-6 Sol and GPT-6 Luna are not wired up yet — they launched yesterday, and we add models after checking them, not on announcement day. If Sol or Luna specifically is what you came for, check back rather than sign up expecting it today.

What the platform is genuinely useful for right now is the comparison that decides your default: run your own task on GPT-6 Astra, on GPT-5.6 Sol, and on Claude, and see where the quality actually stops improving. That boundary is the one worth knowing, and it barely moves when a cheaper tier arrives — if GPT-5.6 Sol already clears your bar, GPT-6 Sol at half the price is straightforwardly better, and Astra was never necessary.

The pattern to take away

OpenAI is no longer competing mainly on the top of its own lineup. Astra held the flagship spot for three weeks; the news since is that the tiers most people actually run got twice as cheap and measurably more accurate. For anyone paying per token, the cost-effective move this quarter is not adopting the newest flagship — it is auditing which of your workloads are quietly running on a model five to fifty times more expensive than the task requires.

Try GPT-6 free on aiwith.chat — or see what one plan covers if you're currently paying for ChatGPT and Claude separately.