ChatGPT vs Claude vs Gemini: Which AI Subscription Is Worth Paying For in September 2026?

Disclosure: this comparison contains no affiliate links. OpenAI, Anthropic and Google do not run affiliate programs for their consumer subscriptions, so there is nothing here for us to earn a commission on. Our affiliate disclosure explains how the rest of the site is funded. Editorial placement is decided on merit and is not for sale.

By Lennit Herrera, founder and editor, AI Tool Stack Reviews  ·  Last updated September 18, 2026

The verdict: if you write, code or work with long documents, Claude Pro is the best $20 you can spend — it is the only one of the three whose paid models carry a 1M-token context window, and the only one that does not use your chats to train its models unless you switch that on. If you want the widest feature surface — custom GPTs, projects, scheduled tasks, a coding agent and desktop work tools in one app — ChatGPT Plus at $19.99 is the better all-rounder. And if you already live in Gmail and Drive, Google AI Pro at $19.99 is the most stuff per dollar on this page: it bundles 5 TB of storage and YouTube Premium Lite with the Gemini app. The under-covered option is Google AI Plus at $4.99 a month, the cheapest real AI subscription any of the three sells.

Prices and plan limits verified against each company’s own pricing page, model documentation and help centre on September 18, 2026. Benchmark positions were read the same day from two independent leaderboards, not from vendor marketing.

On this page

• The short list: which one, in one line each
• Comparison table: price, model, context, privacy default
• How we compared them
• ChatGPT: the widest feature surface
• Claude: the long-document and writing pick
• Gemini: the most bundled value
• Every paid tier, cheapest to most expensive
• What the independent leaderboards actually say
• Who trains on your chats by default
• Context windows: the number nobody advertises
• Who each one is right for
• When not to pay for any of them
• Frequently asked questions

The short list

• Claude Pro — for writing, coding and anything where you paste in a long document.
• ChatGPT Plus — for the person who wants one app that does everything, including images, voice and scheduled tasks.
• Google AI Pro — for anyone whose work already lives in Gmail, Docs and Drive, or who wants the storage anyway.
• Google AI Plus ($4.99) — for the casual user who has outgrown a free tier but will never use $20 of anything.
• All three free tiers — genuinely enough for occasional questions. Most people should start here for a week.

Comparison table

Sources: OpenAI’s ChatGPT pricing page, Anthropic’s plan page and Google’s Gemini subscriptions page, all read September 18, 2026.

How we compared them

We compared three assistants across thirteen consumer and small-team plans against the companies’ own pricing pages, model documentation and help centres on September 18, 2026, plus two independent leaderboards read the same day. We do not test these products ourselves and we do not publish scores of our own. Where a performance claim appears below, it comes from named third-party ranking data and is linked to the source; where a price or a limit appears, it comes from the company’s own page.

Two things are deliberately excluded. We do not quote benchmark numbers that a vendor published about its own model without saying so, because every lab picks the tests it wins. And we do not rate these assistants on “writing quality” as though it were a measurement — that judgement depends on your prompt, your field and your taste, and anyone presenting it as a number is selling something. What we can do is put the verifiable facts side by side: what each plan costs, which model it unlocks, how much text it will hold, and what the company does with what you type.

Prices are United States list prices. ChatGPT’s subscription prices below come from Apple’s App Store listing for OpenAI’s ChatGPT app, which itemises every in-app purchase; OpenAI’s own pricing page describes the plans and their limits but blocks automated readers from the price figures themselves.

ChatGPT: the widest feature surface

• Paid tier price: $19.99/month (Plus). Go is $8; Pro is $100 for 5x usage or $200 for 20x.
• Free tier: Yes — unlimited everyday text chats with GPT-5.6 Luna, with limited uploads, image generation, voice and deep research.
• Top model on the paid tier: GPT-6 Astra, plus GPT-5.6 Sol, Terra and Luna.
• Context window, as published: 54K on the instant model, 256K on the reasoning models. Pro raises those to 128K and 400K.
• Trains on your chats by default: Yes. OpenAI’s own plan table lists “content is used to train our models” on Free, Go, Plus and Pro, with an opt-out.

Why ChatGPT? Because it is the only one of the three where the subscription is a bundle of products rather than a chat window. Plus includes projects, scheduled tasks, custom GPTs, image generation, voice, the Codex coding agent, and ChatGPT Work across desktop, web and mobile. If your problem is that you keep paying for four small tools, this is the one that most plausibly replaces several of them. It also has the most generous free tier on the thing most people actually do: OpenAI describes everyday text chats as unlimited on every tier including Free, subject to abuse guardrails.

What we like

• The broadest feature list of the three by a wide margin, and most of it sits on the $19.99 tier rather than being held back for $100.
• Unlimited everyday text chats on every plan, including the free one.
• GPT-6 Astra sits at the top of Artificial Analysis’s Intelligence Index, tied with Anthropic’s Fable 5.1 at an index of 53.
• A free plan for verified United States K–12 educators runs through June 2027, and nonprofits can get up to 75% off the business plans.

Watch out for

• The $8 Go plan may show you ads. OpenAI’s pricing page says so in plain text: “This plan may include ads.” That is the cheapest tier, and the one most likely to be suggested to a casual user.
• The smallest published context windows of the three. 54K tokens on the instant model at the Plus tier is roughly 40 pages of your own text — an order of magnitude below what Claude’s paid models hold.
• Training is on by default. You get an opt-out, but you have to go and find it.
• Usage is described as “expanded” and “limited” rather than as numbers, so you cannot tell in advance where Plus runs out.

Best for: the generalist who wants images, voice, agents, documents and chat on one subscription.
Skip it if: your work is long documents, or your employer would object to chats being used for training.

See OpenAI’s current ChatGPT plans →

If you are weighing ChatGPT against the smaller, cheaper tools rather than against Claude and Gemini, our roundup of the best ChatGPT alternatives compares six of them on price and output, and our guide to using AI for content covers the prompt structure that cuts editing time on any of these models.

Claude: the long-document and writing pick

• Paid tier price: $20/month, or $17/month on the annual plan billed at $200 up front. Max starts at $100.
• Free tier: Yes — Sonnet and Haiku, with web search, file creation, code execution, memory and Artifacts.
• Top model on the paid tier: Opus 5. Fable 5 and Fable 5.1 are not part of Pro’s usage limits — on Pro they run on pay-as-you-go usage credits.
• Context window, as published: up to 1M tokens on the current paid models.
• Trains on your chats by default: No. Anthropic’s privacy centre states chats are used for model improvement only if you choose to allow it.

Why Claude? The context window is the reason, and it is not a marketing number. Anthropic’s model documentation lists Fable 5.1, Opus 5 and Sonnet 5 at a 1M-token context window with 128K tokens of output, which it puts at roughly 555,000 words of input on the current tokeniser. Anthropic’s help centre says the same window applies to the newest models on paid chat plans. In practice that is the difference between pasting a contract and pasting a filing cabinet, and it is the clearest capability gap between these three products at the same price.

What we like

• The only one of the three that does not train on your conversations unless you opt in — and Incognito chats are excluded even then.
• 1M-token context on the paid models, with long conversations automatically summarised rather than cut off.
• Anthropic models hold the top three places on the LMArena text leaderboard, a crowdsourced blind-preference ranking, and Fable 5.1 ties GPT-6 Astra at the top of Artificial Analysis’s index.
• The annual plan is the biggest genuine discount on this page: $17 a month against $20, charged once as $200.
• Claude Code is included in the $20 plan rather than sold separately.

Watch out for

• The five-hour session cap is the thing people hit. Usage resets on a rolling five-hour session, and there is also a weekly limit that applies across all models. A long working afternoon can exhaust Pro.
• Fable, Anthropic’s strongest model, is not included in Pro at all. Anthropic’s plan table lists Fable as “usage credits” on Pro and as 50% of weekly limits only on Max and premium Team seats, and its help centre confirms the old 50%-on-Pro promotion ended on July 19, 2026. At $20 you get Opus 5, Sonnet and Haiku; the model that ties GPT-6 Astra at the top of Artificial Analysis’s index is a $100 Max plan or a metered extra.
• The thinnest bundle of the three. No image generation, no video, no storage, no consumer ecosystem — you are buying a text and code assistant and nothing else.
• Anthropic stores conversation data attached to thumbs-up and thumbs-down feedback for up to five years, de-linked from your user ID.

Best for: writers, analysts and developers who paste long source material and care about the privacy default.
Skip it if: you want images, voice or video, or you work in long unbroken sessions that would hit the five-hour cap.

See Anthropic’s current Claude plans →

If writing is the job you are buying for, it is worth checking whether a dedicated tool beats a general assistant: our best AI writing tools roundup and our Grammarly review cover the editing-focused options, and the Writesonic review looks at the marketing-copy end of the market.

Gemini: the most bundled value

• Paid tier price: $19.99/month (Google AI Pro). AI Plus is $4.99; AI Ultra is $99.99 or $199.99.
• Free tier: Yes — access to 3.6 Flash, varying access to 3.1 Pro, image generation, Deep Research, Gemini Live and 15 GB of storage.
• Top model on the paid tier: Gemini 3.1 Pro.
• Context window, as published: Google does not publish a per-plan figure for the app; the underlying Gemini 3.1 Pro model is documented at 1,048,576 tokens.
• Trains on your chats by default: Yes, while the Keep Activity setting is on.

Why Gemini? Because at $19.99 you are not only buying an assistant. Google AI Pro includes 5 TB of storage across Gmail, Drive and Photos, a YouTube Premium Lite plan, Gemini inside Gmail, Docs and Vids, 1,000 Google Flow credits for video generation, higher limits in Google’s Jules coding agent, and Gemini in Chrome. If you were already paying Google for storage, the marginal cost of the AI is close to nothing. That bundle, not the model, is the argument — and it is a strong one for anyone whose files already live in Drive.

What we like

• The best free tier of the three on features: image generation, Deep Research and Gemini Live are all included at $0.
• $4.99 for Google AI Plus — twice the free usage limits and 400 GB of storage, and the only sub-$8 paid AI plan any of these three sells.
• 5 TB of storage at the $19.99 tier, which on its own is priced like a storage plan.
• Gemini 3.1 Pro accepts audio and video input, which neither ChatGPT’s nor Claude’s flagship does.

Watch out for

• Activity is kept and used by default. Google auto-deletes Gemini Apps Activity after 18 months unless you change it, chats selected for human review can be kept for up to three years, and turning Keep Activity off still leaves 72 hours of retention.
• The plans are a moving target. Google’s own pages describe the Gemini app as running 3.1 Pro while the Search integration is described as Gemini 3 Pro, and 3.1 Pro is still labelled preview in Google’s developer documentation.
• It ranks lowest of the three on the capability index. Artificial Analysis scores Gemini 3.1 Pro Preview at 30 on its Intelligence Index, against 53 for GPT-6 Astra and Claude Fable 5.1 — a much bigger gap than its LMArena position suggests.
• Several headline features, including Chrome auto browse and Gemini Agent, are United States only.

Best for: people whose documents, email and photos already live in Google, and anyone who wants the cheapest real upgrade at $4.99.
Skip it if: your priority is the strongest reasoning model, or you want your chats kept out of training by default.

See Google’s current Gemini plans →

For the two-way version of this fight in more detail, see our Gemini vs ChatGPT head-to-head.

Every paid tier, cheapest to most expensive

The pattern is the same at all three companies: $20 is the real consumer price, and $100 is the multiplier tier for people who hit limits daily. What changed in 2026 is the bottom of the ladder. Google now sells a $4.99 plan and OpenAI an $8 one, so the honest answer for a lot of casual users is no longer “free or $20” — it is a middle rung that did not exist a year ago. OpenAI has also started selling usage credits as one-off purchases, at $4 for 100 credits up to $40 for 1,000.

What the independent leaderboards actually say

Two credible independent rankings exist, they measure different things, and in September 2026 they disagree sharply. Both were read on September 18, 2026.

LMArena ranks models by blind human preference — people compare two anonymous answers and pick one. On its text leaderboard, Anthropic holds the top three slots, led by Claude Fable 5 at an Elo of 1505.7 across roughly 30,000 votes. Google’s Gemini 3.8 Flash sits ninth at 1493.0 and Gemini 3.1 Pro fifteenth at 1486.8, with OpenAI’s GPT-5.6 Sol eighteenth at 1483.5 and GPT-6 Astra twenty-fourth at 1479.8.

Artificial Analysis builds a composite Intelligence Index from capability evaluations instead. There, GPT-6 Astra and Claude Fable 5.1 tie at the top with an index of 53, Claude Opus 5 follows at 51, and Gemini 3.1 Pro Preview scores 30.

The useful conclusion is not that one of them is wrong. It is that preference and capability are different measurements, and neither should decide a $20 purchase on its own. Two details make that concrete. First, the top twenty-five models on LMArena’s text board are separated by about 28 Elo points and their confidence intervals overlap heavily, so most of that ranking is a statistical tie. Second, the entries are tuned variants with names like “high”, “max” and “xhigh” that do not map cleanly onto what you get inside a $20 consumer app. A leaderboard position is evidence, not a purchase decision.

Who trains on your chats by default

This is the comparison that usually gets skipped, and it is the one with the clearest answer.

Anthropic does not use consumer Claude chats for training unless you opt in. Its privacy centre states that chats and coding sessions are used to improve models only if you choose to allow it, with narrow exceptions for safety-flagged conversations. Incognito chats are excluded regardless of the setting. Conversation data attached to thumbs-up or thumbs-down feedback is kept for up to five years, de-linked from your user ID.

OpenAI trains on consumer content by default with an opt-out. Its plan comparison table carries a “content is used to train our models” row marked “opt-out available” on Free, Go, Plus and Pro alike.

Google uses your chats while Keep Activity is on, which is the default. Per Google’s own support page, Gemini Apps Activity auto-deletes after 18 months unless you change it to 3 or 36 months or switch it off. Conversations selected for human review can be retained for up to three years and are not removed when you delete the chat. With Keep Activity off, or in a temporary chat, Google still holds the conversation for 72 hours.

None of this makes any of the three unusable. It does mean that if you paste client material, unreleased work or anything covered by a confidentiality obligation, the default settings matter more than the model does — and Claude is the only one where the safe setting is already the setting you have.

Context windows: the number nobody advertises

Every assistant will accept a long document. What differs is how much of it survives.

OpenAI is the only one of the three that publishes consumer context limits per plan, and it deserves credit for that even though the numbers are the least flattering: the instant model holds 27K tokens on Free, 54K on Go and Plus, and 128K on Pro, while the reasoning models hold 256K on Go and Plus and 400K on Pro. OpenAI notes that the share available for your own input is smaller than the total, because system instructions, memory and the model’s own reasoning consume part of the window. Its API flagship, GPT-6 Astra, is documented at a 1,050,000-token window — so the consumer app runs a fraction of what the model can do.

Anthropic publishes 1M-token windows on Fable 5.1, Opus 5 and Sonnet 5, and its help centre says the newest models on paid plans support up to 1M tokens, with 500K or 200K on others. Google documents Gemini 3.1 Pro at 1,048,576 tokens on Vertex AI but publishes no equivalent figure for the consumer Gemini app, so the honest answer for Gemini is that we cannot verify what the app allows.

If your work is a 200-page PDF, a year of meeting transcripts or a whole repository, this is the number that decides the outcome, and it is not the number on the pricing page. For transcripts specifically, a dedicated tool often beats all three — see our best AI transcription tools and best AI meeting note taker comparisons.

Who each one is right for

You write for a living. Claude Pro. The 1M-token window means you can give it the whole brief, the style guide and three past pieces at once, and the opt-in-only training default matters if any of that belongs to a client. The annual plan at $17 is the cheapest way to hold it for a year.

You code. Claude Pro or ChatGPT Plus, and the deciding factor is workflow rather than model quality. Claude Pro includes Claude Code in the $20 plan; ChatGPT Plus includes expanded Codex plus everything else. Anthropic’s models lead LMArena’s text board while GPT-6 Astra leads Artificial Analysis’s index, which is a reasonable sign that either will be adequate. If you want the editor rather than the chat window, our best AI coding assistants roundup, Cursor review and GitHub Copilot review go deeper.

You want one app that does everything. ChatGPT Plus. Images, voice, projects, scheduled tasks, custom GPTs and a coding agent on one $19.99 bill is not a claim the other two can make.

Your work lives in Google. Google AI Pro. Gemini inside Gmail and Docs plus 5 TB of storage and YouTube Premium Lite is the best-value bundle here, provided you would have paid for the storage anyway.

You use AI a few times a week. Google AI Plus at $4.99, or nothing at all. All three free tiers are genuinely usable, and OpenAI’s free plan offers unlimited everyday text chats.

You want images or video more than chat. None of these first. Start with our best free AI image generators and best AI video generators comparisons, and for narration our best AI voice generators guide and the ElevenLabs vs Murf head-to-head.

When not to pay for any of them

There are three cases where $20 a month is the wrong answer.

You have not exhausted a free tier. This is most people. ChatGPT Free gives unlimited everyday text chats, Claude Free gives Sonnet and Haiku with web search and Artifacts, and Gemini Free includes image generation and Deep Research. Run a week of your real work through the free tier of your first choice and see whether you hit a wall. If you do not, you have your answer.

Your bottleneck is a specific task, not general intelligence. Transcription, voice generation, image editing, resume formatting and SEO briefs all have purpose-built tools that beat a general assistant at a fraction of the price. Paying $20 for a chatbot to do a $9 job is the most common way people waste money on AI.

You need guaranteed accuracy. None of these three will reliably tell you when it is wrong. Every one of them can produce a confident, well-formatted, incorrect answer, and no subscription tier fixes that. Anything that goes to a client, a regulator or a court needs a human check, and the more fluent the output looks, the more that is true.

Frequently asked questions

Which AI assistant is best in September 2026?

There is no single answer, and the two credible independent rankings disagree. Anthropic’s models lead LMArena’s human-preference leaderboard, while GPT-6 Astra ties Claude Fable 5.1 at the top of Artificial Analysis’s capability index and Gemini 3.1 Pro sits well below both. For writing and long documents, Claude Pro is the strongest pick; for breadth of features, ChatGPT Plus; for value bundled with storage and Google apps, Google AI Pro.

Is the free version of any of them good enough?

For occasional use, yes. OpenAI lists unlimited everyday text chats on its free plan, Claude Free includes Sonnet and Haiku with web search and Artifacts, and Google’s free tier includes image generation, Deep Research and Gemini Live. The free tiers restrict uploads, image volume, voice and deep research rather than basic conversation.

What is the cheapest paid AI assistant?

Google AI Plus at $4.99 a month, which doubles the free usage limits and includes 400 GB of storage. ChatGPT Go is next at $8, but OpenAI’s pricing page states that plan may include ads.

Do these companies use my conversations to train their models?

It depends on the company. Anthropic uses consumer Claude chats only if you opt in. OpenAI’s plan table shows content is used for training on all consumer tiers, with an opt-out available. Google uses Gemini chats while the Keep Activity setting is on, which is the default, and keeps that activity for 18 months unless you change it.

Which one handles the longest documents?

Claude, on the published numbers. Anthropic documents a 1M-token context window on Fable 5.1, Opus 5 and Sonnet 5, and says paid plans get up to 1M tokens on the newest models. OpenAI publishes 256K for reasoning models on Plus and 400K on Pro. Google documents 1,048,576 tokens for the Gemini 3.1 Pro model but does not publish a figure for the consumer app.

Is Claude’s annual plan worth it?

If you are confident you will use it for a year, it is the largest discount of the three: $17 a month against $20 monthly, charged as $200 up front. No equivalent annual consumer discount is published for Google AI Pro, though Apple’s App Store does list a $200 annual ChatGPT Plus purchase alongside the $19.99 monthly one.

Can I switch later without losing anything?

Mostly yes. These are monthly subscriptions, and while conversation history is not portable between them in any meaningful way, nothing you create is locked to one vendor — documents, code and text all export normally. The practical cost of switching is the habits and the custom setups, not the data.

Do any of them pay affiliate commissions?

No, and that is worth saying plainly on a page like this. None of the three runs a consumer affiliate programme, so nobody ranking them can be paid by them for the recommendation. Treat any page that claims otherwise with suspicion.

The bottom line

In September 2026 these three subscriptions cost within pennies of each other and buy quite different things. Claude Pro buys the biggest context window and the only privacy default that assumes your work is yours, which makes it the right pick for writers, analysts and developers. ChatGPT Plus buys the widest feature surface, which makes it the safest choice for someone who cannot predict what they will need. Google AI Pro buys the most total value, provided you were already in the Google ecosystem. Below all of them, Google AI Plus at $4.99 is the first paid rung that makes sense for casual users, and the free tiers are better than the upgrade prompts suggest.

Start with the free tier of whichever matches the description above, use it for a week of real work, and pay for the one that runs out first. That is a more reliable signal than any leaderboard, including the two cited here.

Related reads

• Gemini vs ChatGPT: the two-way comparison
• Best ChatGPT alternatives
• Best AI writing tools
• Best AI coding assistants
• Perplexity vs Google for research
• All our AI tool comparisons

Verified September 18, 2026 against: OpenAI’s ChatGPT pricing page and Apple’s App Store listing for the ChatGPT app (subscription prices); OpenAI’s model documentation for GPT-6 Astra and chat-latest; Anthropic’s plan page, models overview and help-centre articles on usage limits and model training; Google’s Gemini subscriptions page, the Vertex AI documentation for Gemini 3.1 Pro, and Google’s Gemini Apps Activity support page; and two independent leaderboards, LMArena’s text board and Artificial Analysis’s Intelligence Index, both read the same day. Plans and prices change often — check the vendor’s page before you subscribe. — L.H.