A $20 AI subscription looks like the responsible choice. It is a tenth of the flagship price, it is the number every review quotes, and all three major vendors sell something at or near it. The problem is what $20 has become. It is not “the same model with a smaller allowance”. On two of the three vendors it is the first tier that includes the frontier model at all, and on all three the tools that make a subscription worth having — deep research, long documents, a coding agent — sit on separate meters the price does not describe.
So “which $20 plan is best” is the wrong question. The useful one is: where is the ceiling, what happens when I touch it, and how often does my work touch it? A cheap tier is cheap only if you never hit the ceiling. If you hit it every week, $20 is the most expensive way to buy the work.
Everything below comes from vendor documentation fetched on 26 October 2026 — plan tables, limit articles, model-availability notes. No hands-on testing is claimed.
The $20 class, as it exists today
| Plan | Price | Models unlocked | What is metered | Buying past the cap |
|---|---|---|---|---|
| ChatGPT Go | $8/month | GPT-5.6 Luna; no Sol | Messages with tools, uploads, images, voice | Not sold; upgrade only. Ads may appear |
| ChatGPT Plus | $20/month | GPT-6 Sol in chat; Astra in Work and Codex | Expanded messages, uploads, research, Codex, Work | Credits, eligible features only |
| Claude Pro | $20/month, $17 on annual ($200 up front) | Sonnet, Haiku, Opus, Claude Code | Five-hour session cap plus a weekly cap | Usage credits |
| Google AI Plus | $4.99/month | Flash-Lite and Flash | Gemini compute budget, 2× free | AI credits |
| Google AI Pro | $19.99/month | Adds Pro and Gemini 4 Argon | Gemini compute budget, 4× free | AI credits |
| ChatGPT Pro | $100 / $200 / $500 | Adds GPT-6 Pro; Ultrafast only at $500 | Included allowance per tier | Credits, or a higher tier |
| Claude Max | from $100/month | 5× or 20× Pro usage | The same two clocks, enlarged | Usage credits |
| Google AI Ultra | $99.99 or $199.99 | Highest model access, Tasks, Genie | 5× or 20× the AI Pro budget | AI credits |
Three vendors, four price points, three incompatible units of measurement. OpenAI meters models and tools. Anthropic meters time. Google meters compute. That alone should end the idea that these plans are interchangeable at $20, and Google’s own table makes the point twice: AI Plus at $4.99 is the only sub-$20 plan that unlocks a better model than free, and the only one that does not unlock the frontier model. Cheap and useful are different axes.
The advice that is wrong, and the mechanism
Search for guidance on this decision and you will find the same line in a dozen forms: the $20 plan is the same model with a message cap, so start cheap and upgrade when you hit the wall. It sounds prudent. It is wrong in a way that costs money, because the wall is not where the advice says it is.
Model gating. A $20 plan is not a throttled frontier model; on two of three vendors it is the entry ticket to one. OpenAI’s model documentation states that Free and Go users “do not have access to GPT-5.6 Sol” and default to GPT-5.6 Luna. GPT-6 Pro, which runs on Astra, is sold on Pro, Business and Enterprise, while Plus gets Astra inside Work and Codex instead. Anthropic’s comparison puts Sonnet and Haiku on Free while Opus starts at Pro. Google is bluntest: Free gets “access to our Flash-Lite model”, AI Plus “access to our Flash model”, AI Pro “access to our Pro model”. If your work needs the strongest model — a long contract, a hard refactor, a synthesis across twenty sources — the cheap tiers are not slower versions of the same thing. They are a different product.
The meter is per tool. OpenAI lists Free as “limited” on uploads, image creation, voice, deep research, Codex and ChatGPT Work, and Plus as “expanded” on each. That is five counters, not one, and the counter you use decides whether $20 is enough. A writer uploading two PDFs a day never approaches the ceiling. A developer running an agent for an hour is somewhere else entirely, on the same $20.
Some ceilings are measured in pages. OpenAI publishes input limits per plan: roughly 12 pages of text on Free, 40 on Go and Plus, 250 on Pro. The document you paste in is the limit, not the number of replies. Google’s meter rises with chat length; Anthropic’s clock shortens as attachments grow.
Together these explain the real failure mode. People do not outgrow $20 by sending too many messages. They outgrow it because the job changed shape — the document got longer, the research got deeper, the code needed an agent — and the plan was priced for a different shape.
The ceiling ledger
| Plan | Where the wall is | Reset | Buy past it? | What the vendor will not do |
|---|---|---|---|---|
| ChatGPT Plus | Per-model allowances; Work and Codex separate | Per model, reset time shown when available | Credits, eligible features | Support cannot reset a limit |
| ChatGPT Pro 100/200/500 | Included allowance per tier | Per model and weekly | Credits; Ultrafast only at $500 | Sell “unlimited” GPT-6 Pro |
| Claude Pro | Five-hour session plus weekly cap, all models | Session every five hours; weekly at a fixed time | Usage credits | Publish a fixed message count |
| Claude Max | The same clocks, 5× or 20× larger | Same | Usage credits | — |
| Gemini AI Pro/Ultra | Compute budget, drained faster by premium features | Every five hours until the weekly cap | AI credits | Sell a fixed number of prompts |
The ledger matters more than the price list, because price is the only number you can know in advance. Everything in the middle column is learned by using the product.
Two details deserve emphasis. First, OpenAI says support staff cannot reset ChatGPT or Codex limits, and that hitting one “does not mean your account was restricted” — the wall is a product feature, not an incident. Second, the wall moves. OpenAI’s Pro-tiers article says the $200 plan now ships with a lower allowance for new subscribers, with a grandfathering window from 22 September to 29 September 2026 and an old allowance ending on 29 October 2026. Anyone budgeting “Pro 200 gives me X” is budgeting against a number revised this month.
ChatGPT: the ceiling is a ladder
Go costs $8 and buys GPT-5.6 Luna with roughly ten times the free tier’s messages, uploads and image creation, per OpenAI’s launch post. It may carry ads: OpenAI says ads may appear on Free and Go, testing began in the US on 9 February 2026, and Plus and above are ad-free. Go has no legacy models, and its voice limits match free.
Plus costs $20 and buys GPT-6 Sol in chat, expanded research, scheduled tasks and custom GPTs, and roughly 40 pages of input. It does not buy GPT-6 Pro or GPT-5.6 Sol Pro — both are absent from OpenAI’s comparison table for Plus. It also has no annual option; OpenAI’s help centre says annual billing is unavailable for Go, Plus and Pro.
Pro runs from $100 to $500, and even there the word “unlimited” is conditional: “a Pro subscription does not include unlimited use of the GPT-6 Pro model”. Speed is priced as consumption — fast mode draws the allowance at 2.5×, Ultrafast at 8×. Buying speed is buying less of everything else.
For a writer, Plus at $20 is a reasonable buy. For anyone running an agent for an hour a day, the ladder is the constraint, and the next rung is not $30. It is $100.
Claude: the ceiling is a clock
Anthropic publishes no message count. There is a five-hour session limit and a weekly limit across all models, resetting at a fixed day and time assigned to your account rather than whenever you start. The Pro plan article says the number of messages varies with message length, attachment size, conversation length, model choice and features used.
That makes Claude easy to reason about in one direction and hard in another. Easy: context is not the upsell — all plans list up to 1M tokens, varying by model. Hard: you cannot compute cost per task before paying, because the unit is a clock whose speed depends on what you are doing.
Two documented details help you stay under it. Cached project content counts less when reused, and caches expire after inactivity, so a daily project is cheaper per question than a fortnightly one. And usage credits exist, so the answer to “I hit the wall at 3pm” is sometimes a small top-up rather than a jump to Max.
On price, Anthropic is the one vendor here selling a discounted year: $17 per month when billed as $200 up front, against $20 monthly. OpenAI offers no annual billing on Go, Plus or Pro; Google’s page quotes monthly prices. The question before signing an annual deal is whether your usage is steady or spiky, because the weekly allowance resets on a schedule and cannot be banked.
Max starts at $100 for 5× Pro usage and $200 for 20×, bundling $100 or $200 a month in API credits. If you already buy API access, that credit is a real offset; if not, it is a discount on something you were not going to buy.
Gemini: the ceiling is a budget
Google publishes limits as multipliers rather than units. Free gets Flash-Lite; AI Plus at $4.99 doubles the allowance and adds Flash; AI Pro at $19.99 quadruples it and adds Pro; Ultra is $99.99 for five times the AI Pro allowance, or $199.99 for twenty times. Google describes the meter in its own footnote as compute-based: it factors the complexity of the request, the features used and the length of the conversation, refreshes every five hours until a weekly limit is reached, and can be topped up by buying AI credits.
So Google’s ceiling is the most elastic and least predictable of the three. A short question is nearly free; a long analysis or a media generation is not. Two AI Pro users can have wildly different headroom because they do different kinds of work, not different volumes of the same work.
The help centre lists which features use that budget fastest — media generation, Deep Research, the Pro model, extended thinking and Deep Think — and adds a detail most comparisons miss: the low, medium and high effort setting on a single model buys a more thorough answer at a higher cost against the same limit.
The routing is the other half of the design, and it is the part people miss. From October 2026, Auto is the default for accounts without a paid plan: most prompts go to Flash-Lite, prompts that need deeper reasoning may go to Flash or Pro, and if smart model selection is switched off the account uses Flash-Lite for everything. The free tier is therefore not one model. It is a router that decides per prompt, which is also why two free users can get different answers to the same question.
Google also bundles hard to ignore. AI Pro includes 5 TB of storage, YouTube Premium Lite and Google Home Premium Standard; Ultra adds Project Genie and a YouTube Premium individual plan. If you already pay for storage and YouTube, the effective price of the AI is lower than the sticker. That does not make the plan cheap; it makes the comparison unfair in Google’s favour, which is worth knowing before you add a second subscription.
Three $20 plans are not $60 of the same thing
The scenario in the title is specific: someone quietly paying three subscriptions for one job. It happens because the job has three phases and each phase has a different bottleneck.
A realistic week: a 60-page report to read, a small internal tool to fix, and a research question needing twenty sources. The report is a context problem, so Plus at 40 pages is wrong for the whole document. The tool is a coding problem, so Codex on Plus, Claude Code on Pro, or Jules and Antigravity on AI Pro are the candidates. The research question is a metered-tools problem, and all three vendors meter it separately from chat.
Buying all three at $20 each is often correct — $60 is less than one $100 tier, and you get three strengths. The mistake is buying all three and still hitting the wall on the phase that matters, because none was sized for the heavy version of that phase. If 80% of your value comes from the coding agent, $60 spent to be blocked in the one place you needed to work is worse than one $100 tier aimed at the bottleneck.
A rule that survives contact with the pricing pages: count metered actions, not subscriptions. Which single action, repeated, would you pay to never be blocked on? If it is a coding agent, price the agent tiers. If it is long-document analysis, price input size. If it is research, price the deep-research allowance. Then buy one plan for that action and let free tiers cover the rest. For a wider view of which tools absorb work rather than adding a dashboard, see our comparison of AI productivity tools that help humans work.
The break-even test that needs no invented numbers
You cannot compute cost per task in advance, because no vendor publishes task counts. So measure backwards with two weeks of your own log.
Each working day, note which task hit a wall, what you did instead, and how many minutes the workaround cost. “Instead” includes waiting for the reset, switching tools, splitting a document by hand, or writing the section yourself.
Then compare minutes against the price gap, which is conveniently similar across vendors. ChatGPT Plus to Pro 100 is $80. Claude Pro to Max 5x is $80. Google AI Pro to Ultra is $80. At $40 an hour, two lost hours a month pays for any of those upgrades. At $15 an hour, you need more than five.
That is the whole ROI argument, and it has nothing to do with benchmark scores. The cheapest plan is the one whose ceiling your work does not touch. The most expensive is the $20 tier you keep working around.
One caveat: the upgrades are not equivalent. ChatGPT Pro 100 and Claude Max 5x buy more of the same product; Google AI Ultra buys model access and video features AI Pro lacks at any volume. The deciding question is which wall you hit, not which plan has more of everything.
If the bottleneck is specifically the coding agent, vendor differences matter more than price gaps, and we went through them in Claude for coding in 2026. For writing-heavy work, the constraint is how much text you push through per day — the same metering logic that governs AI content generators.
What the $20 tier is genuinely enough for
Most people do not need $100, and pretending otherwise would be bad advice.
The $20 tier is enough when your work is one job, in short documents, with a human reviewing every step. Drafting, rewriting, summarising a report you have already read, answering questions about a contract under 40 pages, generating a first pass at code you will review line by line — all of that fits inside Plus, Pro or AI Pro.
It is also enough when you can tolerate the wall. If hitting a limit once a fortnight means pausing for an hour, the wall is not a cost. The calculation changes only when the wall lands on work you are paid to finish today.
The cheapest defensible setup in October 2026 is not a $20 plan by default. It is one paid plan aimed at your bottleneck, free tiers for everything else, and a monthly check on whether the wall moved. All three vendors revised limits this year, and two revised them in the last eight weeks. A subscription decision made in January is a decision about a product that no longer exists.
If you are still mapping tools to tasks rather than prices to plans, the tools archive is the place to start.
Sources
- https://chatgpt.com/pricing ” line-up, comparison table, context and input limits, Go ads note.
- https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus ” Plus $20/month, no annual billing.
- https://help.openai.com/en/articles/11989085-what-is-chatgpt-go ” Go contents, no legacy models.
- https://openai.com/index/introducing-chatgpt-go/ ” Go at $8/month, 10× free-tier messages and uploads.
- https://help.openai.com/en/articles/9793128-what-is-chatgpt-pro — Pro 100/200/500 pricing, reduced allowance, grandfathering dates, no support resets, 2.5×/8× speed consumption.
- https://help.openai.com/en/articles/11909943 ” model availability by plan, Sol absent on Free/Go, Pro-model limits.
- https://help.openai.com/en/articles/20001047 ” ads on Free and Go from 9 February 2026.
- https://claude.com/pricing — plan line-up, $17 annual / $20 monthly, Max 5×/20× and API credits, model access, 1M context.
- https://support.claude.com/en/articles/8325606-what-is-the-pro-plan ” five-hour reset, weekly cap, usage credits, no API included.
- https://support.claude.com/en/articles/9797557-usage-limit-best-practices ” what consumes usage, caching behaviour, usage settings.
- https://gemini.google/subscriptions/?hl=en ” Free/Plus/Pro/Ultra pricing, 2×/4×/5×–20× limits, model access, bundles, and footnote 2 on the compute-based meter that refreshes every five hours until the weekly limit and can be topped up with AI credits.
- https://support.google.com/gemini/answer/17004136?hl=en ” October 2026 change: Auto becomes the default for accounts without a Google AI plan, most prompts routed to Flash-Lite with deeper-reasoning prompts going to Flash or Pro, Flash-Lite for everything when smart model selection is off; compute-based limits (complexity, features, chat length) refreshing every five hours to a weekly cap, 2×/4×/5×–20× by plan, and the features that consume faster — media generation, Deep Research, the Pro model, extended thinking, Deep Think and higher effort settings.
- https://ai.google.dev/gemini-api/docs/rate-limits” API rate limits vary by model and tier and are read in AI Studio rather than published as fixed quotas, which is why no developer-side number appears in this piece.
Independent technology writer focused on artificial intelligence, emerging technologies, and digital innovation. Covers AI applications in sports, productivity, and online business.
