Claude, ChatGPT or Gemini: which assistant is worth paying for in 2026

Updated 30 October 2026: Since 7 October GPT-6 has been the default model in ChatGPT’s Chat on Plus and above, and on 29 October OpenAI cancelled the lower Pro 200 allowance for new subscribers. We have updated the model table, plan details, FAQ and sources; our pick is unchanged.

Claude, ChatGPT and Gemini all want about $20 a month from you, and in September 2026 each has a brand-new flagship to justify it. Anthropic has Claude Fable 5.1 and Opus 5.5, plus Sonnet 5.5 since 28 September; OpenAI has GPT-6 Astra, Sol and Luna, plus GPT-6.1 Sol since 29 September; and Google has Gemini 4 Argon alongside its Pro and Flash models. We went through the vendors’ own pages to work out what a personal plan costs, which model you really get for the money, how much context you get, what file, voice and app support comes with it, and what each company does with your chats unless you tell it otherwise. We checked docs and prices on 28 September 2026 and rechecked them on 30 September for OpenAI’s 29 September announcements, Anthropic’s Sonnet 5.5 and Google’s AI Plus price. For picks beyond these three, see the best AI chatbot for each job.

The models on offer right now

All three companies shipped their biggest models of the year in the last four weeks, so it matters which version you’re looking at. Anthropic announced Claude Fable 5.1 at the start of September (the system card is dated 1 September 2026), Claude Opus 5.5 on 22 September and Claude Sonnet 5.5 on 28 September, which replaces June’s Sonnet 5 as its current Sonnet. The Free plan runs “Sonnet”, and Anthropic’s plan page doesn’t say which version. OpenAI launched GPT-6 Astra on 3 September and followed it with the cheaper GPT-6 Sol and Luna on 22 September and GPT-6.1 Sol on 29 September, while the older GPT-5.6 models still power the free tier. Google released Gemini 3.8 Flash on 2 September, its third Flash model in six weeks. In the Gemini app, AI Pro and Ultra now include Gemini 4 Argon alongside the Pro and Flash models.

Model	Context / max output	API price per 1M tokens (in / out)	Knowledge cutoff	Consumer plan
Claude Fable 5.1Anthropic · Sep 2026	1M / 128K	$10 / $50	Jun 2026	Pro and Max, 50 % of weekly limit
Claude Opus 5.5Anthropic · 22 Sep 2026	1M / 128K	$4 / $20	Jun 2026	Pro $20 and up
Claude Sonnet 5.5Anthropic · 28 Sep 2026	1M / 128K	$2 / $10	Jun 2026	Paid plans; Free version not stated
GPT-6 AstraOpenAI · 3 Sep 2026	1.05M / 128K	$10 / $50	Apr 2026	Plus $20 and up; Astra Pro on Pro
GPT-6 SolOpenAI · 22 Sep 2026	1.05M / 128K	$2 / $10	Apr 2026	Plus and up; default in Chat since 7 Oct
GPT-6.1 SolOpenAI · 29 Sep 2026	Not yet confirmed	$2 / $10	Not yet confirmed	Plus and up, in Work and Codex only
Gemini 3.8 FlashGoogle · 2 Sep 2026	1,048,576 / 65,536	$0.75 / $3.75 (intro, to 31 Dec 2026)	Not published	AI Pro $19.99 and up
Gemini 3.1 ProGoogle · 19 Feb 2026, preview	1,048,576 / 65,536	Not on the model page	Not published	Not on the free tier

At the top end, pricing has converged. Claude Fable 5.1 and GPT-6 Astra both cost $10 in and $50 out per million tokens, and Claude Opus 5.5, GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5 all sit between $2 and $4 in. Google is the exception and is plainly competing on cost. Gemini 3.8 Flash’s introductory API price is a fraction of the others’. Whether any of that matters depends on whether you pay per token or per month. Most people pay per month, so that’s where we’ll start.

Prices

The prices below are the US dollar list prices from each company’s own pages. Viewed from Europe, the same pages show euro prices (ChatGPT Plus is €23, Google AI Pro is €21.99), so expect to pay a little more if you’re billed in euros. ChatGPT Plus has no annual option; Claude Pro drops to $17 a month if you pay $200 up front.

The plan structures have ended up almost identical. Each has a free tier, a $20 tier and a power tier at $100 and $200 (ChatGPT has added a third, Pro 500, at $500), and two of the three have a cheap tier under $10 (ChatGPT Go at $8 and Google AI Plus, which Google cut from $7.99 to $4.99 in June while doubling its storage to 400 GB; Anthropic has nothing at that price). Google’s bundle is the odd one out because it’s a Google One plan. AI Pro comes with 5 TB of storage and Gemini inside Gmail, Docs and Vids, which is a real saving if you’d pay for the storage anyway. ChatGPT Go is the only plan of the lot that “may include ads”, according to OpenAI’s own pricing page.

What your $20 actually buys

This is where the three differ most, and where the marketing is least clear. On Claude Free you get Sonnet and Haiku, though Anthropic’s plan page doesn’t say which versions. Claude Pro includes the new Sonnet 5.5 and adds Opus 5.5 at no extra charge and Fable 5.1, Anthropic’s top model, at “50 % of weekly limits”. Max multiplies usage by 5x or 20x. So a Claude Pro subscriber can pick the company’s best model from the menu in any conversation. Nobody else offers that at $20.

ChatGPT’s tiers are more layered. Free gets you “unlimited text chats with GPT-5.6 Luna”, the previous generation’s small model, plus limited uploads, voice and image generation. Plus adds “advanced reasoning models with GPT-6”, and OpenAI’s Astra launch post says Astra itself is rolling out to “all ChatGPT Plus, Pro, Business, and Enterprise users” within existing allowances. Pro, from $100, adds “Pro reasoning powered by GPT-6 Astra” and 5x the usage. The new GPT-6 Sol and Luna models launched only in ChatGPT’s Work and Codex surfaces, and Free and Go users get Luna only in the desktop app. Since 7 October, though, GPT-6 has been the default in the ordinary Chat view on Plus and above, with GPT-6 Sol behind the Instant and thinking levels, and GPT-5.6 Sol can still be picked from the model menu. GPT-6.1 Sol, added on 29 September, follows the same pattern. Plus, Pro, Business, Enterprise and Edu users get it in Work and Codex, it is not yet in Chat, and it isn’t listed for Free or Go.

Gemini is the simplest. Google’s plan page lists the Flash-Lite model on the free tier, and AI Plus at $4.99 moves you up to the Flash model. AI Pro at $19.99 raises the usage limits, takes the context window to 1 million tokens and adds Gemini 4 Argon alongside the Pro and Flash models. Ultra, at $100 or $200, adds Deep Think, Google’s slow-reasoning mode, and early access to Gemini Spark.

Context windows and file limits

All the flagship models now take about a million tokens in the API, roughly 555,000 words on Anthropic’s tokenizer. What you get inside the consumer apps is a different matter, and only two of the three publish it. OpenAI’s pricing page is the most explicit. The “GPT Instant total context window” is 27K on Free, 54K on Go and Plus and 128K on Pro, while the reasoning models get 256K on Go and Plus and 400K on Pro, with the caveat that system instructions and memories eat into that space. Google’s support page gives 32K tokens with no subscription, 128K on AI Plus and 1 million on Pro and Ultra. Anthropic doesn’t publish a per-plan context figure for claude.ai at all; the models themselves take 1M, and usage is metered in five-hour sessions instead.

On files, Anthropic publishes the clearest limits: up to 20 files per chat at 500 MB each, images up to 8,000 by 8,000 pixels, and PDFs up to 1,000 pages, with the first 100 pages read visually and the rest as text. OpenAI describes uploads as “limited” on Free and Go and “expanded” on Plus and gives no numbers. Google’s 3.8 Flash and 3.1 Pro accept text, images, video, audio and PDF natively in the API, the broadest input list of the three, but the consumer help pages don’t say what the file limits are per plan.

Apps, voice and extras

All three run on the web, on iOS and Android, and on desktop. Claude’s voice mode is in beta on every plan including Free, on mobile, desktop and web, and it can use connected tools such as Gmail and Calendar mid-conversation. ChatGPT lists voice as “limited” on Free and Go and “expanded” on Plus, with unlimited voice on Pro. Gemini Live is built into the Gemini app and has been merged with ordinary chats since May; Google shipped a native Windows app on 10 September and fn-key dictation on macOS in July.

Away from the chat box, each vendor is pulling in its own direction. Claude Pro bundles Claude Code and the newer Claude Design, Slides, Docs and Science tools, plus extensions for Chrome and Microsoft 365. ChatGPT Plus adds projects, scheduled tasks, custom GPTs, expanded Codex usage and the ChatGPT Work desktop app, and Pro adds unlimited image generation. Pro is also getting dots, OpenAI’s always-on agents powered by GPT-6 Astra, rolling out gradually from 29 September, though Pro access excludes the European Economic Area, Switzerland and the UK at launch. Google’s advantage is that Gemini is already inside Gmail, Docs, Vids and Search on a paid plan, and 3.8 Flash is embedded in Sheets and in AI Mode in Search as well as the app. If you spend your day in Google Workspace, that integration is the product you’re buying. If you spend it in a terminal, Claude Code or Codex is.

What each one does with your chats

Most comparisons skip this part, yet it’s where the three companies take the most different approaches. OpenAI’s pricing page says plainly, for Free, Go, Plus and Pro, that “content is used to train our models”, with an opt-out. The switch is called “Improve the model for everyone” under Data controls, and temporary chats are excluded and kept for up to 30 days. Google’s Gemini Apps Activity works the same way. With “Keep Activity” on, chats are used to improve Google’s models, a subset is read by human reviewers, and reviewed chats are kept for up to three years. Turn it off and future chats aren’t used for training and are held for 72 hours. Activity auto-deletes after 18 months by default.

Anthropic takes a different approach and asks first. Since it updated its consumer terms, new users choose at sign-up whether their chats may be used to improve Claude, and they can change their answer in Privacy Settings. Say yes and retention runs to five years; say no and Anthropic keeps chats for 30 days. Incognito chats are never used for training either way. On every service, business and enterprise workspaces are excluded from training by default, so if this matters to you and your employer is paying, get the work plan.

The benchmark claims

We haven’t run a head-to-head test of the three for this guide, so the only numbers here are the companies’ own. Each vendor picks the tests it wins, so read them with that in mind. OpenAI’s GPT-6 Astra launch table gives Astra 57.9 percent on Terminal-Bench 4.0 against 55.8 percent for Claude Fable 5.1 and 19.1 percent for Gemini 3.8 Flash, and 96.0 percent on GPQA Diamond against 93.7 and 95.3 percent. The same table, to OpenAI’s credit, shows Fable 5.1 ahead on Humanity’s Last Exam with tools, 65.0 to 57.2 percent, and on the Artificial Analysis Intelligence Index, 65.7 to 61.2. OpenAI’s GPT-6.1 Sol post, from 29 September, claims the new model matches Astra on DeepSWE v1.1 at roughly a fifth of the cost and scores 2.2 points above Claude Opus 5.5 at medium effort on AutomationBench at about a third of the cost, while Astra stays highest on Terminal-Bench Science at 68.1 percent.

Anthropic’s own Fable 5.1 page reports 55.8 percent on Terminal-Bench 4.0, 52.6 percent on Terminal-Bench-Science and 65.0 percent on Humanity’s Last Exam with tools, so at least the two companies agree on Anthropic’s numbers. Anthropic’s Sonnet 5.5 post, from 28 September, claims 70.6 percent on Terminal-Bench 4.0 for Sonnet 5.5 and 66.4 percent for Opus 5.5, both above the Astra and Fable 5.1 figures, though vendors run these tests their own way and Anthropic itself says Opus 5.5 “remains clearly stronger at complex, open-ended work”. Google’s 3.8 Flash post claims 54.9 percent on HLE-Verified and 47.2 percent on CWE-Bench patching, but sells the model on cost rather than raw score. Our reading of all this is that the top models sit within a few points of each other on most published tests, that Astra and Anthropic’s models lead on agentic coding.

Reasons to pick each one

Gemini’s case is simpler, so it gets a paragraph rather than a list. Pick it if you live in Google Workspace, if 5 TB of storage was on your shopping list anyway, or if you need native video and audio input. Its weaknesses are the ones the table shows. Google doesn’t publish knowledge cutoffs, and the free tier only gets Flash-Lite. If none of the three suits you, our guide to ChatGPT alternatives covers other assistants and what each does better.

Verdict

Frequently asked questions

Sources and method