I pay for ChatGPT out of my own pocket and use it most days. I have also rewritten this review, because it contained two claims about a competitor that were wrong, and it recommended paying for something OpenAI switches off on August 30, 2026.
Both of those are worth stating plainly at the top rather than quietly fixing.
The thing with a deadline
The old version of this review said the $20 was justified by "memory, Canvas, higher rate limits, and DALL-E access."
DALL-E is going away. DALL-E 3 was deprecated on May 12, 2026, and OpenAI set August 30, 2026 as the retirement date for the DALL-E GPT inside ChatGPT. Image generation in ChatGPT now runs on ChatGPT Images 2.0 and the gpt-image family.
Your images do not stop working. The name you were sold does. But if you were choosing a subscription this week based on a review recommending DALL-E, you were being pointed at a product with a shutdown date, and a lot of ChatGPT reviews still do exactly that.
What this review got wrong about Claude
Two claims, both stated as fact, both false.
It said Claude Pro was "weaker for images, no voice mode, and no live web browsing by default." Claude has voice mode, on every plan including the free one, with continuous listening and connected tools it can use while you talk.
It also said that if Anthropic shipped cross-chat memory in Claude "that will be a real challenge to Plus. As of mid-2026 they still have not." Anthropic lists memory as a free tier feature on its own pricing page.
I checked those against Anthropic's pages rather than against my memory of them, which is what I should have done the first time. The Claude review now carries the corrected version, and if you want the two products side by side, Claude vs ChatGPT is the direct comparison.
$20 is now the middle of six tiers
This is the change that matters most for anyone deciding what to pay, and it is the reason I retitled this review.
When I first wrote it, ChatGPT had one sensible paid tier for individuals. Now there are six plans: Free, Go, Plus, Pro, Business, and Enterprise. Go is $8 a month in the US. Plus is still $20. Pro is no longer a single $200 wall, it comes in tiers at $100 and $200. The plan formerly called Team is now Business.
So "is $20 worth it" has stopped being the right question. The question is whether $20 is the right one of six, and for a decent number of people it is not. If you use ChatGPT a few times a week and keep bumping into free tier limits, Go at $8 is the honest answer, and no review was telling you that a year ago because it did not exist.
There is also metering to understand. Some models are charged per message in credits rather than simply being included, with the heavier reasoning modes costing more per message than the standard ones. Anthropic has moved the same way. Both companies now sell you an allowance rather than a flat product, and both are harder to forecast a month ahead than they used to be.
The reasoning models you read about are gone
If you have an older article open in another tab telling you to use o1 for hard problems, close it.
GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini were retired from ChatGPT on February 13, 2026. GPT-4.5 went on June 26, 2026. o3 is scheduled for August 26, 2026. The separate o-series is finished as a product line.
What replaced it is better, and simpler to explain. Reasoning is now a setting inside the GPT-5 family rather than a different model you switch to. You choose how hard you want it to think, it spends longer before answering, and you get the step by step working. The behaviour I used to praise o1 for is still there. It just is not called o1 any more.
I am deliberately not going to tell you which exact version number sits on your plan. OpenAI's own pages currently say different things in different places, listing GPT-5.2 Instant, GPT-5.4, and GPT-5.6 Sol depending on which page the search engine hands you. Open your model picker and read it. That is more reliable than any review, including this one. For how the families compare with Anthropic's, I went through it in GPT-5.6 vs Claude 5.
How to choose a thinking level
Since reasoning became a setting rather than a separate model, the practical skill changed. You are no longer picking between products, you are deciding how much compute to spend on a question, and the wrong choice costs you either time or accuracy.
My own rule after a few months of this: use the standard mode by default, and raise the thinking level when a wrong answer would cost me more than the wait. Drafting an email, summarising a document, explaining a concept, standard mode is faster and just as good. Debugging something across several files, checking a piece of arithmetic that feeds into a decision, working through a contract clause, that is worth the extra thirty seconds.
The failure mode I see people fall into is leaving it on the heaviest setting permanently, on the theory that more thinking is always better. It is not. You burn your allowance faster, because heavier reasoning costs more per message, and on a lot of ordinary tasks the answer is identical. The other failure mode is never touching it and then concluding the model is bad at maths.
One thing genuinely worth the upgrade: you can read the working. When the answer is wrong, seeing where it went wrong tells you whether to re-prompt or start again. A confident wrong answer with no trace is far more expensive than a slow right one.
The free tier and Go, honestly
Most reviews of Plus skip past the free tier in a sentence, which made sense in 2024 and does not now.
Free includes image generation and voice. It has memory. It searches the web. For someone who opens ChatGPT twice a week to draft something or check an idea, that is a working product, not a demo. The honest reason to leave it is the ceiling: you hit a cap, get moved to a lighter model, and notice the drop mid-task.
Go at $8 sits in the gap that used to be empty. More messages and more tool use than Free, longer memory, voice included on the same terms as the free tier. What it does not give you is the top of the model range or the full set of the newer features.
Here is the test I would actually apply. Spend a fortnight on Free and pay attention to two things: how often you hit a limit, and whether you hit it while doing something that mattered. If you never hit one, stay. If you hit them occasionally and it was mildly annoying, Go is $8. If you hit them in the middle of real work and it cost you momentum, that is the $20 signal, and you will know it without a comparison chart.
The thing nobody tells you is that moving down is fine too. If you paid for Plus during a busy quarter and your usage has dropped, Go now exists as a landing spot that did not before. Nothing punishes you for stepping down.
What Plus actually buys you
Stripped of the marketing, here is the honest list.
Much more usage, and the better models rather than being bumped to a lighter one mid-task. This is the thing you feel, and it is most of the value.
Image generation in the chat, without switching tools. Still capped, still generous for normal work.
Voice, with real back and forth conversation and interruption. I use it while cooking and driving and almost never at my desk, which I suspect is true of most people who claim it is essential.
Memory across conversations. I dismissed this at launch and I was wrong. Not re-explaining who I am and how I write at the start of every session saves more time than any single model upgrade did.
Canvas, which opens a document beside the chat instead of dumping two thousand words into the message window. This is where I do long drafts now.
Projects, which group chats with their own instructions and files, so a new conversation starts already knowing the context.
Custom GPTs, plus the largest third party ecosystem of any assistant. If you rely on community built tools, this is the strongest reason to be on ChatGPT rather than Claude.
Priority access when things are busy.
Where it still falls short
Hallucination has not been solved and will not be. ChatGPT states wrong things with complete confidence, and the better the writing gets the more convincing the wrong answers become. Web search reduces it and does not fix it. Verify anything that matters. That includes prices in this review, which is why I have dated them.
Quality drifts in very long conversations. Past roughly forty or fifty turns on a complex piece of work, I notice it losing the thread. Starting a fresh chat with a summary fixes it, and I have stopped treating that as a bug and started treating it as housekeeping.
Buying it is genuinely confusing now. Six tiers, credit metering, per message costs that differ by reasoning level, and model names that change every few months. I could not tell you what a heavy month costs without doing arithmetic, and that is a fair criticism of a consumer product.
For sustained code work I still reach for something else. Claude holds a detailed spec more tightly, and if you want the assistant inside your editor rather than a chat window, Cursor vs GitHub Copilot covers what those actually cost once agent usage is metered.
It rewards people who know what they want. Give it a vague brief and you get something that looks competent and says nothing. That is not a flaw in the model so much as a description of what it is for.
What I could not verify, and why that matters
I want to be straight about the sourcing on this one, because it is weaker than I would like.
OpenAI blocks automated access to its own site. Trying to read openai.com, help.openai.com, or chatgpt.com programmatically returns HTTP 403. So unlike the Claude review, where I checked every claim against Anthropic's live pages, the prices and dates here come from OpenAI's published pages as indexed by search, plus their public announcements.
Two things I found circulating that I am not going to repeat as fact: that ChatGPT Go carries advertising, and that the $200 Pro tier has a one million token context window. Both appear only in third party blogs quoting each other, which is the same pattern I refused to launder in the Claude comparison. They may well be true. I have not seen either from OpenAI.
The retirement dates and the tier structure I am more confident about, because those trace to OpenAI's own announcements and help pages.
Which tier is actually right for you
Free is enough if you use it occasionally and can live with the caps. It includes image generation and voice now, so it is a real product rather than a trial.
Go at $8 is the tier most people have not heard of and some should be on. Occasional but regular use, annoyed by free limits, not willing to spend $20.
Plus at $20 is right if ChatGPT is part of most working days. Writers, developers, researchers, anyone generating images as part of the job.
Pro at $100 is the new sensible step for people who exhaust Plus daily. Before this tier existed the only option was $200, which was a bad deal for a lot of the people who took it.
Pro at $200 is for running it constantly as core infrastructure.
Business matters for one reason more than any feature: it does not train on your content by default. If you handle client work, that is worth more than the admin controls.
If none of this earns its keep, the best free AI tools list covers what you can get for nothing. And if you would rather bill it to someone else, using ChatGPT for freelancing covers what clients actually pay for.
Pricing
| Plan | Price | Who it is for |
|---|---|---|
| Free | $0 | Occasional use, with caps and image generation included |
| Go | $8 monthly (US) | Regular but light use, priced below Plus |
| Plus | $20 monthly | Daily use for work |
| Pro | $100 monthly | Exhausting Plus most days |
| Pro | $200 monthly | Running ChatGPT as core infrastructure |
| Business | Per seat, from 2 users | Teams, and no training on your content by default |
| Enterprise | Custom | Security, compliance, admin at scale |
Prices checked on August 6, 2026 against OpenAI's published plan pages. This is a product that repriced three times in a year, so verify before you pay.
Alternatives at the same money
Claude Pro at $20 is the closest competitor and the gap is smaller than it was. Better at long form writing and holding a brief. Weaker for image generation and a smaller ecosystem.
Gemini at around the same price is worth a look if you live in Google Workspace, where the document and mail integration does real work. ChatGPT vs Gemini goes through where the difference actually shows.
Perplexity Pro at $20 is the better research tool and a worse general assistant. It cites sources for every claim, which for factual work is a different way of using AI rather than a better version of the same thing.
Plenty of people pay for two. ChatGPT plus Claude at $40 total is the common pair, and I run it myself. If you can only pick one, pick on your dominant task rather than on feature counts.
My verdict
ChatGPT Plus is still a good subscription and I am still paying for it. The models got better, memory and Canvas changed how I work, and the free tier improving does not bother me because it brings more people to the point where paying makes sense.
What changed is that Plus is now one option among six, and the shopping is harder than the product. I cannot predict a month's cost without arithmetic, the model names move constantly, and the image generator readers were told to pay for is being retired, on a date OpenAI set as August 30, 2026.
Rating: 8 out of 10, down from 8.5.
The product did not get worse. Buying it did. Hallucination is still there, competitors still beat it at specific things, and now there is a genuine risk of paying $20 when $8 would do or $200 when $100 would do. That is a real cost to a real reader, so it comes off the score.
Try ChatGPT. Free tier first, then pick your tier once you know how you actually use it.
Verified on August 6, 2026 against OpenAI's model retirement announcements, published plan pages, and the ChatGPT Images 2.0 announcement, and against Anthropic's pricing and support pages for the Claude comparisons. OpenAI blocks automated access to its own site, so the OpenAI figures here are second hand in a way the Anthropic ones are not. Treat the dated claims as dated.
