Nobody can honestly tell you which US$20 AI subscription is best, and any article that opens with a winner is selling you something. The three main options have converged on almost the same price, and the published usage limits are vague enough that you cannot compare them on a spec sheet.
What you can do is work out which of them you are actually deciding between, and whether you should be buying a subscription at all. For a large number of practitioners the honest answer in August 2026 is that the API costs a couple of dollars a month for the same work.
What are you actually deciding when three AI subscriptions all cost about US$20?
You are not choosing a price. You are choosing which usage ceiling, which model family, and which set of built-in features you want as your default. At roughly US$156 a month in Hong Kong dollar terms for a year of one seat, the wrong default is an expensive habit rather than a disaster.
The decision splits three ways. Buy one seat for interactive work. Skip the seat and pay per token through the API for repeatable production work. Or buy nothing yet because you have not hit a limit that costs you anything.
Most people never consider the second option, which is why they overpay. Most people also skip the third, which is why they own two subscriptions and use one.
What do the paid tiers cost in August 2026?
Published list prices as of August 2026, in US dollars per month for individual plans. All three providers now run a budget tier under US$10, a standard tier at about US$20, and a power tier at US$100 to US$200.
Subscription tiers, monthly list price
--- ChatGPT Go US$8, ChatGPT Plus US$20, ChatGPT Pro US$200
--- Claude Pro US$20, Claude Max US$100 for 5 times the usage, Claude Max US$200 for 20 times the usage
--- Google AI Plus US$7.99, Google AI Pro US$19.99, Google AI Ultra US$99.99 for 5 times and US$199.99 for 20 times
--- Google Workspace Business Standard US$14 per user per month, which folds Gemini into Gmail, Docs, Sheets, Slides and Calendar rather than selling it separately
API list price, per million tokens
--- Claude Opus 5, released 24 July 2026: US$5 input, US$25 output. Batch is half price at US$2.50 and US$12.50. A prompt cache read costs US$0.50 per million.
--- GPT-5.5: US$5 input, US$30 output. Batch is US$2.50 and US$15.
--- Gemini 3.7 Flash, released 13 August 2026: US$0.75 input, US$3.75 output. This is introductory pricing that expires on 31 December 2026. Context is 1,048,576 input tokens and 65,536 output tokens, with a stated knowledge cutoff of March 2026.
The gap between the cheapest and the most expensive output token here is a factor of eight. That is the single most useful number in this article.
When is the API cheaper than a US$20 subscription?
The API wins when your work is repeatable and your token count is predictable. It loses badly when you work by conversation, attach large documents, and iterate.
Take a concrete case. You produce 20 long drafts a month from a fixed brief: roughly 8,000 tokens of input and 2,700 tokens of output each, so about 160,000 input and 54,000 output tokens a month. At list prices that is about US$0.32 on Gemini 3.7 Flash, about US$2.15 on Claude Opus 5, and about US$2.42 on GPT-5.5. Against a US$20 seat, the subscription is the expensive option by a wide margin.
Now flip it. You spend three hours a day in conversation, attaching a 100,000-token document and taking 30 turns. That is roughly 3 million input tokens in one day, which on Claude Opus 5 list pricing is about US$15 for the day. A US$20 monthly seat is dramatically cheaper, and this is exactly what the subscription is for. If token maths is new to you, our explainer on what a context window is covers why attached documents dominate the bill.
The honest test is one question: does the same instruction run repeatedly with predictable inputs? If yes, price it as an API job. If no, buy a seat.
Which subscription fits which kind of practitioner?
The verdict changes with the work, not with the brand. These are the splits that hold up in practice.
--- Writer, marketer or content creator working mainly in a chat window with file uploads. Buy one US$20 seat and choose on output taste, not price, because the prices are within twenty cents of each other. Run the same real brief through all three free tiers for a week first.
--- Anyone producing volume from a fixed prompt. Skip the seat. Use the cheapest capable model through the API, which on current list prices means Gemini 3.7 Flash at roughly a fifth of the output cost of the others.
--- Long-document analyst, researcher or anyone who hits limits by lunchtime. The thing you are buying is the ceiling. Claude Pro at US$20 or Max at US$100 is priced explicitly as a usage multiple, which makes it the easiest ceiling to reason about.
--- Already inside Google Workspace. Workspace Business Standard at US$14 per user per month may already cover what you were about to buy separately.
--- Working on a tight budget or still experimenting. Start on the US$8 tiers. ChatGPT Go and Google AI Plus exist precisely so you do not have to commit US$20 to find out.
--- A team of three or more each buying a personal seat. Stop and consolidate. Three individual seats with no shared prompts, no shared skills and no shared review process is the most common way practitioners waste money on AI.
If your problem is that you cannot keep track of which tool does what, a maintained catalogue is more useful than another comparison article. UD keeps one in the UD AI Directory.
What the price tags do not tell you
This is where a comparison article should be honest, so here is what these numbers cannot do for you.
--- One of the prices is temporary. Gemini 3.7 Flash's US$0.75 and US$3.75 rates are introductory and expire on 31 December 2026. Any cost model you build on them has an end date.
--- Model availability moves faster than your workflow. OpenAI no longer publishes GPT-5.5 on its current API pricing page, having moved promotion to the newer GPT-5.6 family. A model you standardise on today can quietly become the one nobody is optimising.
--- The US$20 usage limits are not published as hard numbers. They are described in relative terms and change without notice, so "enough usage" is not a specification you can plan a business process around.
--- Third-party comparisons are not controlled benchmarks. Review-site verdicts on which model writes better are informal testing, not matched experiments, and this article does not claim otherwise.
--- Hong Kong buyers pay in US dollars. Card fees and FX spread sit on top of every figure above, and regional availability of specific features is not guaranteed to match the US.
--- Nobody can tell you which model you will prefer. Output taste is the one variable in this decision that no amount of pricing research resolves.
How to run a two-week test before you commit
Two weeks on free tiers will tell you more than any comparison table, provided you test with your real work instead of trick questions. Use the same brief in all three tools, on the same day, and score the output rather than the vibe.
Try this prompt in each tool, unchanged
"You are helping me evaluate you against two other AI assistants for my actual work. Here is a real task I do every week: [paste the task, the raw input, and one example of an output I was happy with].
Produce the output. Then, separately, answer four questions in under 40 words each: 1) which parts of my brief were ambiguous, 2) what you guessed, 3) what you would need from me to do this better next time, 4) which parts of this task you are likely to do worse at than a specialised tool. Do not flatter the brief and do not tell me the output is excellent."
Score each tool on three things only: how close the first draft was to usable, how many turns it took to fix, and how often you hit a limit. Then buy the one that won, and cancel the others before the second billing cycle rather than after it.
The correct next step
If you have not hit a paywall or a usage limit in the last 30 days, buy nothing this month. If you are producing volume from a fixed prompt, price it through the API before you renew anything. If you live in a chat window all day, one seat is clearly worth it, and which one is a taste question you can settle in two weeks.
The expensive mistake is not picking the wrong US$20 plan. It is paying for three of them, using one, and never noticing that the work you actually repeat would have cost two dollars.
Spending well on AI is mostly about knowing which of your tasks is a habit and which is a process. We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
Reviewed by the UD AI team. Prices are publisher list prices checked in August 2026 and change without notice.
Stop Paying for Seats You Do Not Use
If the honest answer is that you need a role rather than another chatbot seat, that is a different purchase. We'll walk you through every step, from mapping which tasks are actually repeatable to choosing the right tool and getting it running in your daily workflow.
