How ChatGPT’s Free Plan Changed in Two Years

The short answer
In 2024 the free plan was metered by messages on a good model: you got GPT-4o for a while, then it quietly dropped you to GPT-3.5 when the cap ran out. In 2026 the chat itself is unlimited on one model, and the meter moved to the tools - uploads, images, voice, deep research, Codex - plus ads that did not exist two years ago.
I have used the free plan for about eight months, and reading OpenAI's own release notes while writing this was the first time I understood why it feels both more generous and more limited than it did a year ago. Both are true, and they are not a contradiction: the company moved what it meters.
Free in November 2024 against free in November 2026
| Late 2024 | November 2026 | |
|---|---|---|
| Main chat | GPT-4o with a message cap, then fallback to GPT-3.5 | Unlimited text chats on GPT-6 Luna, subject to abuse guardrails |
| Model choice | One model at a time, decided by OpenAI | One model only (Luna). Sol and Terra stay paid |
| Reasoning | None on free | A Think button, no slider |
| Uploads | Files and photos allowed, cap unpublished | Allowed, but 3 uploads a day |
| Images | Not on free until later | Limited image creation |
| Voice | Preview on free | Limited voice chats |
| Ads | None | Ads may appear on Free and Go in some countries |
| Storage | Nothing permanent | Library with 500 MB |
Everything in the right-hand column comes from OpenAI's own pages, read in November 2026: the pricing page, the Free Tier FAQ, the File Uploads FAQ and the release notes.
The four moments that changed it
1. May 2024: the good model arrived, with a meter
GPT-4o reached free users on 13 May 2024. That announcement is the one that set the pattern for the next two years: free users got the same model as paying users, plus web answers, file uploads, photo questions, data analysis and memory, but with a message limit that depended on demand. OpenAI never published the number, and when you hit it the conversation dropped to GPT-3.5, a model that could not use any of those tools.
By mid-2025 the practical figure people reported was around 15 messages in a three-hour window. That is the version of "free" most students remember: a good assistant for fifteen minutes, then a dumber one.
2. February 2026: ads arrived on the free plan
Ad testing started in the United States on 9 February 2026, for logged-in adults on Free and Go. It expanded to Canada, Australia and New Zealand in March, then to the UK, Mexico, Brazil, Japan and South Korea in August. The format is one labelled ad unit below a response, and OpenAI says Plus, Pro, Business, Enterprise and Edu accounts stay ad-free, as do accounts belonging to under-18s.
This is the single biggest change in the two years, and the one nobody puts in a comparison table. The free plan stopped being a product and became an audience.
3. August 2026: unlimited chat, one model
On 6 August 2026 OpenAI announced that everyday text chat would be unlimited on Free and Go from the week of 10 August, with separate limits kept for images, uploads and the rest. At the same time free users were pinned to a single model, GPT-5.6 Luna, the smallest of that generation, with one reasoning mode behind a Think button and no slider.
Read that as a trade: you can talk as much as you like, but you cannot choose the brain, and the heavy tools keep their own meters.
4. October 2026: GPT-6 moved into the main chat
The 7 October release notes say the rollout started for Plus, Pro, Business and Enterprise in the Chat tab and expanded to Free and Go the next day. Paid plans use GPT-6 Sol; free and Go use GPT-6 Luna. The pricing page now reads "unlimited text chats with GPT-6" on the free card.
Two OpenAI pages still said GPT-5.6 Luna weeks after that, which tells you how carefully the model name on a free plan is maintained. If you are quoting a model name anywhere, quote the date you read it.
What actually got better
- The chat ceiling is gone. Two years ago a long study session ended in a fallback model. Now the ceiling is only on tools.
- A real reasoning button. Free users in 2024 had no way to ask for more thought; Think is now on the free plan, one tap, no slider.
- Storage that survives. A 500 MB Library is genuinely new, and it is the difference between a chat toy and a place you keep work.
What got quietly worse
- Three uploads a day. The file cap is the one that bites students hardest: three PDFs and you are done until tomorrow.
- You cannot pick the model. Sol and Terra exist and are paid-only. In 2024 you could at least wait for the good model to come back; now there is nothing to wait for.
- Ads. They are labelled and they are below the answer, but they are there, on a plan that used to be clean.
- The name on the tin is stale. OpenAI's own help pages lagged the model change by weeks. Do not trust a review that does not carry a date.
If you are a student, what this means
The free plan is better than it was for talking and worse for feeding it material. That is the exact trade a student notices in week two: you can ask questions all night, but you cannot drop four lecture PDFs in and expect them to be read.
My own rule, from running the same five prompts through ChatGPT Free and Claude Free: use free plans while the work is one conversation long, and pay the first time a tool limit, not a chat limit, stops you. If you are weighing the paid tiers instead, the Plus vs Pro comparison does that arithmetic.
Questions people ask
Is ChatGPT free still unlimited in 2026?
Everyday text chat is unlimited on Free, subject to abuse guardrails. Images, uploads, voice, deep research and Codex have their own limits, and file uploads are capped at 3 a day.
When did ChatGPT start showing ads?
Testing began in the United States on 9 February 2026 on Free and Go. It reached Canada, Australia and New Zealand in March 2026 and the UK, Mexico, Brazil, Japan and South Korea in August 2026.
Which model does the free plan use now?
GPT-6 Luna, since the rollout that reached Free and Go on 8 October 2026. Before that it was GPT-5.6 Luna, and before that a capped GPT-4o that fell back to GPT-3.5.