# ChatGPT Free vs Plus: Free runs GPT-6 Luna and paid plans run Sol. In my test, asking Luna again moved its cited sites as much as switching models

> ChatGPT Free runs GPT-6 Luna and Plus runs Sol. Asked the same questions, Luna's own repeat answers differed as much as Luna's and Sol's did.

Canonical: https://brandonlazovic.dev/articles/chatgpt-free-users-get-luna/  
Author: Brandon Lazovic  
Published: 2026-10-29

## The short version

- OpenAI says ChatGPT's Free and Go plans run GPT-6 Luna, while Plus, Pro, Business and Enterprise run GPT-6 Sol. OpenAI does not say how many of its weekly users are on each plan.
- Asked the same 10 questions three times each through OpenAI's API with web search, a Luna answer and a Sol answer shared a median 60% of their cited sites, and two Sol answers shared 65%.
- Two Luna answers to the same question shared a median 55% of their cited sites, so asking Luna again changed its sources about as much as switching to Sol did.
- Where the models did differ, the gap was large. On running shoes, Sol cited shoe makers' own sites that Luna never cited, while Luna cited review publishers that Sol never cited.

ChatGPT's free and paid plans run different models. OpenAI says the Free and Go plans run GPT-6 Luna, while Plus, Pro, Business and Enterprise run GPT-6 Sol. [1] I asked both models the same 10 everyday questions, three times each, through OpenAI's API with web search switched on, and compared which websites each answer cited. The two models cited somewhat different sites. But asking Luna the same question a second time changed its cited sites about as much as switching to Sol did. So a check of which sites ChatGPT cites, run once from one account, can measure chance as much as it measures the model.

> Two Luna answers to the same question shared a median 55% of their cited sites, slightly less than a Luna answer and a Sol answer shared, at 60%.

The test ran through OpenAI's API, the paid route developers use to call models by the same names, and not through the ChatGPT app. Every answer in this test was made to search. [2] The app decides for itself whether to search, and it may add instructions this test cannot see.

## Which model does ChatGPT Free use?

ChatGPT Free uses GPT-6 Luna, the same model as the Go plan, according to OpenAI's GPT-6 announcement of October 7, 2026. The paid Plus, Pro, Business and Enterprise plans use GPT-6 Sol. OpenAI's sentence is "GPT-6 in ChatGPT is powered by GPT-6 Sol for Plus, Pro, Business, and Enterprise tiers, and GPT-6 Luna for Free and Go tiers. Both are tuned for everyday conversation." [1]

The paid plans got GPT-6 first. OpenAI said the rollout started with Plus, Pro, Business and Enterprise and that "Starting tomorrow, the rollout expands to Free and Go tiers." [1]

OpenAI describes GPT-6 as "built for more than 1.2 billion people who use ChatGPT each week," but the announcement does not say how many of them are on the free plans. [1] It also says nothing about how either model chooses which sources to cite.

## Do ChatGPT Free and paid models cite the same websites?

ChatGPT's free and paid models, GPT-6 Luna and GPT-6 Sol, cited partly different websites for the same questions in my test. Overlap here means the share of sites two answers have in common out of all the sites either one cited. A Luna answer and a Sol answer to the same question shared a median 60% of their cited sites. [2] Two Sol answers to the same question shared 65%. [2]

So the models differed, but only by a little more than Sol differed from itself. Before running the test I set the bar the claim had to clear. If the overlap between the two models had been no lower than Sol's overlap with itself, the models would have cited alike within the noise. It came in lower, by five points. [2]

The table below gives each question's overlap, where 100% means two answers cited exactly the same sites.

| Question | Luna vs Sol | Sol vs Sol | Luna vs Luna |
|---|---|---|---|
| best running shoes for flat feet | 18% | 51% | 18% |
| best budget robot vacuum | 89% | 78% | 100% |
| how to remove a red wine stain from carpet | 46% | 42% | 33% |
| is a heat pump worth it in a cold climate | 53% | 44% | 47% |
| best travel rewards credit card | 67% | 100% | 100% |
| how long does a us passport renewal take | 100% | 100% | 100% |
| best project management software for small teams | 64% | 63% | 63% |
| how to speed up a slow windows laptop | 100% | 100% | 100% |
| best noise cancelling headphones under 300 dollars | 32% | 47% | 17% |
| what are the side effects of creatine | 56% | 67% | 33% |
| Median | 60% | 65% | 55% |

Question by question, the result splits. The two models overlapped less than Sol did with itself on four questions, more on four, and the same on two. [2]

## Does asking ChatGPT the same question twice change the sources?

Asking GPT-6 Luna, the model behind ChatGPT Free, the same question twice changed its cited sources about as much as switching to the paid model did. Two Luna answers to the same question shared a median <span class="hand-ring">55%</span> of their cited sites in my test, slightly below the 60% a Luna answer shared with a Sol answer. [2]

On the creatine question, two Luna answers shared a third of their sites, while a Luna answer and a Sol answer shared more than half. On the headphones question, Luna's own answers shared 17%. [2] So a single free-plan answer is one draw from a range of possible source lists, and the next person to ask may see a different one.

Think of asking two friends for a restaurant pick, and then asking the first friend again the next day. If the first friend's second answer differs from their first as much as it differs from the other friend's, the gap between the two friends tells you little.

## Where do the free and paid ChatGPT models differ most?

GPT-6 Luna and GPT-6 Sol differed most in my test on two shopping questions, "best running shoes for flat feet" and, next, the headphones question. On running shoes, a Luna answer and a Sol answer shared 18% of their cited sites, while Sol's own answers shared 51%. Sol cited shoe makers' own sites, newbalance.com, saucony.com and hoka.com, which Luna never cited. [2]

Luna went to review publishers on that question instead, citing runrepeat.com, runnersworld.com and doctorsofrunning.com, none of which Sol cited. [2]

The same lean showed in the sites pooled across all 10 questions, with exceptions on both sides, though I noticed it only after seeing the data, so it describes one small test and does not show how the models behave in general. The table below lists every site that only one of the two models cited, across 30 answers each.

| Cited only by Luna | Cited only by Sol |
|---|---|
| basecamp.com, doctorsofrunning.com, nerdwallet.com, nlr.gov, runnersworld.com, runrepeat.com, soundguys.com, wired.com | acsm.org, consumerreports.org, hoka.com, newbalance.com, notion.com, patientpop.com, roborock.com, saucony.com, sherwin-williams.com, soundcore.com |

Luna's list leans to review publishers, and Sol's to makers' own sites. [2] For a brand, that is the kind of difference worth checking for its own category, on both models, before drawing a conclusion from either.

## How did I compare GPT-6 Luna and Sol?

I compared GPT-6 Luna and GPT-6 Sol by sending each the same 10 questions three times through OpenAI's Responses API, with the web search tool required so every answer searched and cited, on October 26, 2026. Both models ran at the same, lowest reasoning setting. The whole test, including a two-call trial run before the 60 answers, cost $2.13 by my own tally from each call's reported usage, against a cap of $3. [2]

Each cited link was reduced to its website, so two pages on the same site count as one match. For every question I compared each Luna answer with each Sol answer, and each model's three answers with one another, and took the share of sites in common. Every answer searched at least once and cited at least one site. [2]

The test has limits a reader should weigh. The API is not the ChatGPT app, which decides for itself whether to search. It is 10 questions on one day, in US English, from one network, and the website is a coarse unit. I checked four questions' answer-by-answer site lists by hand against the table, and they matched.

## What does the free and paid model split mean for AI visibility checks?

The split between ChatGPT's free and paid models means an AI visibility check, a test of whether ChatGPT cites a given site, should say which plan and model it ran on. The bigger lesson from my test is repetition. One answer is a sample, so ask the same question several times on each model and report how often a site appears across the answers. [2]

A check run only from a paid account tests Sol, while every Free and Go user gets Luna. [1] If a brand's own site shows up on Sol and a review site takes its place on Luna, as on the running-shoe question, the paid-account check reports a result free-plan users do not see. [2]

How ChatGPT picks its sources in the first place is covered in [the article on ChatGPT's source selection policy](/articles/chatgpt-source-selection-policy/). Swapping one model for another in your own pipeline is a separate risk, covered in [the article on a half-price model swap](/articles/half-price-model-swap/). And counting the visits those citations send is its own problem, covered in [the article on ChatGPT clicks arriving as direct traffic](/articles/chatgpt-clicks-arrive-as-direct/).

## Sources

1. OpenAI: GPT-6 and Intelligent UI for everyone (2026-10-07) — https://openai.com/index/gpt-6-for-everyone/
2. First-hand: 10 questions asked three times each of gpt-6-luna and gpt-6-sol through OpenAI's Responses API with the web search tool required, cited domains compared, 60 answers analyzed on 2026-10-26, $2.13 spent across 62 calls including a two-call pilot