ChatGPT, Claude or Gemini: which one is actually worth $20 a month?

• What each plan costs, and which model you get
• Which one for what
• API prices for developers
• What happens to your data
• How to pick in a week
• When a chat plan isn’t enough

You pay per person per month, and the plan decides which models you get. Prices are before tax, from each maker’s page on 13 October 2026. For personal plans, ChatGPT and Google AI show us US prices.

ChatGPT Pro (with GPT-6 Astra), Claude Max (with Opus 5.5 and Fable 5.1) and Premium users on the Business and Team plans start at $100 a month.

Sources, checked 13 October 2026: the price pages for ChatGPT, ChatGPT Business, Claude, Google AI, Google Workspace and Mistral; OpenAI on GPT-6 in ChatGPT; OpenAI and Tek.no on ads in Europe.

• Writing, offers and contracts: ChatGPT Plus or Claude Pro. OpenAI says GPT-6 Astra is its best model at following your own templates, and Anthropic built Opus 5.5 for long tasks in coding and office work.
• Coding: paid ChatGPT plans include Codex and paid Claude plans include Claude Code, tools that write and fix code.
• You live in Gmail and Google Docs: Google Workspace, with Gemini built in.
• A very hard problem now and then: start with a $20 plan. OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra, and Anthropic says Opus 5.5 works at Fable 5.1’s level on most tasks.
• The same small task, all day long, like sorting email or answering simple chat questions: that’s a job for an automation, a setup that does the task by itself on a low-cost model such as GPT-6 Luna, Claude Haiku 5.5 or Gemini 3.8 Flash.
• A European provider: Mistral trained Large 4 in Europe and runs it in its own European data centres. It’s still a test version, so try it before you rely on it.

Sources, checked 13 October 2026: Google Workspace prices; OpenAI on GPT-6 Astra and GPT-6.1 Sol; Anthropic on Opus 5.5 and Haiku 5.5; Google on Gemini 3.8 Flash; Mistral Large 4.

Through the API you pay per million tokens, the small pieces of text a model reads and writes, with one price for what goes in and one for what comes out.

Each maker’s benchmarks favour its own models, so test on your own prompts. Mistral promises open weights for Large 4 by the end of October, so you could run it on your own servers. Anthropic’s developer platform stores data in the US only for now.

Sources, checked 13 October 2026: API prices from OpenAI, Google and Mistral; Anthropic’s model overview and pages on Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5; GPT-6 Sol and Luna; OpenAI’s model guide; Anthropic on data residency.

• Personal plans (ChatGPT Free, Go, Plus and Pro, and Claude Free, Pro and Max) can use your chats to train the AI until you turn this off in the settings.
• The personal Gemini app uses your activity to improve Google’s AI, and people at Google read a sample of chats. You control this under Gemini Apps Activity.
• Business plans: OpenAI says it doesn’t train on ChatGPT Business data, and Anthropic doesn’t train on Claude Team content by default. Google says Workspace content isn’t used to train its AI outside your company without permission. Mistral lets you switch training off.
• Data that must stay in Europe: OpenAI offers this on ChatGPT Enterprise, not on Business, and Google lets you choose where data is stored on Workspace Enterprise. If you need it, get it in writing before you buy.

On free and personal plans, keep client names, contracts and figures out of your chats.

Sources, checked 13 October 2026: the price pages for ChatGPT, ChatGPT Business, Claude, Google Workspace and Mistral; the privacy pages for Gemini Apps and Google Workspace.

1. Take three real tasks from last week: a tricky email, a long document to summarise, a bug.
2. Remove names and figures before you paste anything into a free plan.
3. Give two or three tools the same instructions.
4. Score each answer from 1 to 5: is it right, does it sound like you, how much did you fix?
5. Pay for the winner, and turn off training on your chats the day you sign up.

A chat plan helps one person at a time. AI that answers your customers or takes over work your team repeats is set up differently: one task, tested on real data for one to two weeks, which is how our AI pilot works.