GPT-6 Astra vs Claude vs Gemini: Which AI Assistant for Everyday Work? (September 2026)

If you only want the answer: for everyday writing, long documents and careful thinking, Claude Pro is the strongest $20 plan in late September 2026. ChatGPT Plus is the better all-rounder if you make images or want an agent that does jobs across your apps. Gemini’s Google AI Pro is the best value if your life already runs on Gmail, Docs and Drive. The free tiers of all three are good enough for light use.

The catch is the headline model. GPT-6 Astra is OpenAI’s new flagship, but on the $20 Plus plan you only get it inside ChatGPT Work and Codex, not in normal chat. This guide explains what each plan really gives you, which one fits which task, how to switch off training on your chats, and how to test all three yourself in one afternoon. Expect changes within weeks: OpenAI’s DevDay is on September 29 and Google says Gemini 4 is coming “as soon as possible”.

What changed in September 2026

This has been an unusually busy month, which is why older comparisons are already out of date.

• September 2: Google released Gemini 3.8 Flash, and it went live in the Gemini app for Google AI Pro and Ultra subscribers (Google blog). Anthropic released Claude Fable 5.1, a premium model.
• September 3: OpenAI launched GPT-6 Astra.
• September 10: OpenAI paused new sign-ups and upgrades to the $200 ChatGPT Pro plan (OpenAI Help Center). Google launched a Gemini app for Windows.
• September 22: Anthropic released Claude Opus 5.5 and raised five-hour usage limits on paid plans (Anthropic). OpenAI released the smaller GPT-6 Sol and Luna models the same day.
• September 24: Google DeepMind’s chief said Gemini 4 is in post-training and should ship “as soon as possible”, hopefully well before the end of 2026 (9to5Google).

Two more things are close. OpenAI’s DevDay is on September 29, and a leaked $500 “Pro Max” ChatGPT plan may appear there. OpenAI has not confirmed it (TestingCatalog). So treat this guide as a snapshot. The method for choosing will last longer than the verdicts.

The naming mess: which model you actually get

Each company now has a family of models, and the name on the news is rarely the name on your screen. A “model” is the AI brain behind the chat box. The app is just the window.

Here is what each paid tier actually runs in normal chat, according to the companies’ own pages.

The ChatGPT rows come from OpenAI’s help page on GPT-5.6 and GPT-6 Pro, which says GPT-6 Pro is available “for Pro $100, Pro $200, Business and Enterprise” (OpenAI Help Center). OpenAI’s launch post had said Astra would reach all Plus users, and Plus subscribers are asking about the gap on OpenAI’s community forum.

The Claude rows come from Anthropic’s pricing page (Claude pricing). Claude Fable 5.1 is Anthropic’s premium tier. On Pro it is not part of your plan limits, so you pay extra credits to use it. Max plans can spend up to half their weekly limit on it. For everyday work, Opus 5.5 is the model that matters on Pro.

On Google’s side, “Gemini Pro” still means 3.1 Pro, a model first released in February. The newest Gemini model is 3.8 Flash. “Flash” normally means Google’s faster, cheaper line, but 3.8 Flash scores higher than 3.1 Pro on the leaderboards covered below.

What $0, $20 and $100 actually buy

All prices are US dollars per month, as of September 2026, before tax. Local prices differ in many countries, so check the checkout page where you live.

Sources for the table: ChatGPT Go’s $8 US price and the ad test come from OpenAI’s launch post, which also says Plus and higher stay ad-free (OpenAI). OpenAI’s help centre says Plus is billed monthly, with no annual option. The Max prices come from Anthropic’s help centre, because the pricing page just says “from $100” (Claude Help Center). Gemini prices and features come from Google’s plan page (Google Gemini subscriptions).

Two practical points. First, Claude is the only one of the three whose official pricing page shows an annual discount on its $20 plan: $200 a year instead of $240. Second, Google’s $19.99 plan is a bundle. If you already pay Google for extra storage, AI Pro may replace that bill rather than add to it.

Usage limits in plain English

None of the three publishes a simple “X messages a day” number for its $20 plan. Here is what they do say.

• ChatGPT: Free and Go have unlimited everyday text chat, with tighter limits on tools like images and uploads. Plus has “higher model limits” but no public number. The only hard numbers are for the Pro model: 50 messages a week on Pro $100 and 200 a week on Pro $200.
• Claude: usage works in five-hour sessions. When you hit the cap, you wait for the next reset. Paid plans also have a weekly cap. Max buys 5 or 20 times Pro’s per-session allowance. Anthropic raised the five-hour limits on September 22 but did not publish the new figures.
• Gemini: limits refresh every five hours until you reach a weekly limit. AI Plus is 2x the free limits and AI Pro is 4x (Gemini Apps Help).

The other limit that matters is context. That is how much text the assistant can hold in mind at once, measured in tokens (roughly three quarters of a word each). Anthropic’s help centre lists 1 million tokens in chat for Opus 5.5 on paid plans. Google AI Pro also has 1 million, but AI Plus has 128,000 and Free has 32,000. ChatGPT Plus has 54,000 in its fast Instant mode and 256,000 when it reasons, and Pro raises that to 128,000 and 400,000 (ChatGPT pricing).

In practice, 1 million tokens is roughly a few long books. 54,000 is roughly a long report. If you paste whole contracts, transcripts or research folders, that gap shows up quickly.

What independent rankings say (and where they disagree)

Each company publishes benchmark wins for its own model. Those are sales material. Two independent boards are more useful, and on GPT-6 Astra they disagree sharply.

Artificial Analysis runs a battery of tests and combines them into one score (Artificial Analysis). LMArena shows people two anonymous answers and asks which is better. Its text board, updated September 25, puts GPT-6 Astra below OpenAI’s own older GPT-5.6 Sol (LMArena).

What that means for you: Astra looks strong on hard reasoning tests, but voters did not prefer its everyday answers. Claude Opus 5.5 leads both boards. Gemini’s newest Flash model does well with voters but trails on the test battery.

Leaderboards measure average quality, not which assistant suits your emails or your writing voice. That is why this guide includes a way to test them yourself.

Task-by-task verdicts

This table is our judgement. It combines the plan features above, the independent rankings and published reviews. It is not the result of our own benchmark, and a different reader could reasonably weigh things differently.

Email and everyday writing

Most of what normal people do with an assistant is writing: replies, requests, complaints, cover letters, social posts. For this, the model that people prefer in blind votes matters more than maths scores. Claude Opus 5.5 leads that vote.

One published head-to-head points the same way. Tom’s Guide tested Claude Opus 5.5 against ChatGPT on GPT-6 Sol with five everyday prompts, including rewriting a corporate announcement and planning a birthday party. Claude won four of five. The reviewer found ChatGPT “less fluffy” and faster, and it won the personal organisation task (Tom’s Guide). That is one reviewer and five prompts, so treat it as a signal, not proof.

Writing in your own voice is a known weak spot for Astra. Matt Shumer, who had early access, called Astra very reliable on multi-step newsletter work, but wrote that getting it “to consistently absorb my writing preferences has been less successful” (Matt Shumer).

Gemini’s advantage is location, not quality. On paid plans it drafts inside Gmail and Docs, so you skip the copy and paste.

Research and fact-finding

All three can search the web and all three have a “deep research” mode. That mode spends several minutes reading dozens of pages and writes a cited report. For a normal person comparing, say, phone plans or course platforms, any of them will do the job.

The real risk is the same everywhere: confident answers with wrong details. Prices, dates and rules are where assistants slip most. Always click through to the source for any number you will act on.

Gemini has one useful extra. Gemini Notebook (the tool that used to be called NotebookLM) lets you load your own sources and ask questions only about them. It was renamed on July 16, 2026, remains a separate product, and gained the ability to run code for data analysis (Google blog).

Our verdict is a tie. Pick on ecosystem and price.

Long documents, PDFs and meeting notes

This is where context size stops being a spec sheet number. Say you paste a 90-minute meeting transcript and a 40-page contract into one chat. At around 60,000 to 80,000 tokens, that is already past ChatGPT Plus’s 54,000 token Instant window. Reasoning mode goes to 256,000, but it is slower.

Claude Pro and Google AI Pro both hold 1 million tokens. In practice both can take a folder of reports and keep track of all of them.

For meeting summaries, the useful prompt is the same everywhere: ask for decisions, owners, deadlines and open questions, each with a quote from the transcript. Quotes let you check the summary against what was actually said.

Verdict: Claude, with Gemini very close. If your documents live in Google Drive, Gemini Notebook’s source-only answers are a strong reason to pick Google instead.

Spreadsheets and data

All three can read an uploaded spreadsheet, run code on it, and give you totals, charts and a cleaned file back. For a small business sales sheet or a household budget, the differences are small.

What differs is where the work happens. Gemini works inside Google Sheets on paid plans, and Google says 3.8 Flash is available there. Claude can create and edit Sheets through its Google Workspace connectors, and asks before each change by default. Claude Pro also lists Microsoft 365 among its integrations. ChatGPT Work, launched in July, produces spreadsheets and connects to Google Drive and SharePoint.

Whichever you use, ask it to show the formula or code, then spot-check two or three numbers by hand. Assistants sometimes skip rows or misread dates.

Images and video

Claude does not generate images. It can build diagrams, simple graphics and slides through Claude Design and Artifacts, but not photos or illustrations. If images are a regular part of your week, Claude cannot be your only tool.

ChatGPT released Images 2.5 on September 8, for users “across all tiers”, with faster generation and more precise edits (OpenAI). Gemini uses Google’s Nano Banana models. An April TechRadar test of the previous versions found ChatGPT’s images more realistic, while calling Google’s “more than good enough in most circumstances” (TechRadar). We found no independent test of Images 2.5 against Google’s current model yet.

Video is simpler. OpenAI shut down Sora, so ChatGPT has no video generation. Gemini’s paid plans make 10-second clips with sound using Gemini Omni 1.1 Flash, and every clip carries Google’s invisible SynthID watermark (Google Gemini). For serious video work, dedicated tools are still stronger. Our comparison of Kling, Seedance and Veo covers that.

Small coding jobs and automations

“Coding” for most readers means small things: a script that renames 300 files, a formula that cleans a spreadsheet, a simple landing page, a tool that pulls data from a website.

Claude Pro includes Claude Code and Cowork. Cowork is Anthropic’s agent for file and desktop tasks, and Claude Code writes and runs code for you. ChatGPT Plus includes Codex, and this is where Plus users actually get GPT-6 Astra. Google AI Pro includes Jules, Google’s coding agent.

Anthropic’s launch page claims Opus 5.5 matches Astra on one coding benchmark “at about 40% of the cost”. That is a vendor claim. The independent Artificial Analysis index also ranks Opus 5.5 above Astra. For small jobs, both are more than capable.

If you want to see what Astra can build from one prompt, our GPT-6 Astra games walkthrough shows it in action.

Side-hustle tasks: scripts, listings and product copy

If you earn with AI, this is where the assistant pays for itself or doesn’t. Typical jobs:

• YouTube scripts, hooks and titles.
• Etsy or Gumroad product descriptions and tags.
• Outlines and chapter drafts for ebooks or courses.
• Customer replies and short sales emails.

All of these are writing tasks, so our pick is Claude, with ChatGPT close behind. Two warnings apply to all three.

First, generic output is a real risk. YouTube has tightened its rules on mass-produced, repetitive content, and marketplace buyers skim past bland listings. Feed the assistant your own notes, examples and opinions, then edit hard. Our guide to YouTube hooks that stop the scroll shows what that editing looks like.

Second, check facts. Assistants happily invent statistics that sound right. If a number goes in a video or listing, find the source yourself.

Ecosystem fit: Google, Microsoft, Apple or none

The apps you already use can matter more than which model is 5 points ahead on a leaderboard.

You live in Google (Gmail, Docs, Drive, Android). Gemini has the obvious edge. AI Plus and AI Pro put it inside Google apps, and AI Pro adds 5 TB of storage. Claude also connects to Gmail, Calendar, Drive, Docs, Sheets and Slides, and can draft or send email if you approve each action (Claude Help Center). So Claude is a reasonable choice for Google users who care most about writing quality.

You live in Microsoft (Outlook, Word, Excel, Teams). None of the three is built in. ChatGPT Work connects to Teams and SharePoint. Claude Pro lists Microsoft 365 among its integrations. Microsoft’s own option is Microsoft 365 Premium at $19.99 a month or $199.99 a year, which puts Copilot inside Word, Excel and Outlook. If you mostly work inside Office files, compare that plan with ChatGPT and Claude before buying.

You use Apple devices. All three have iPhone apps and work on a Mac. We did not find a reason in current official material to pick one over another for Apple users alone, so choose on tasks.

No strong ecosystem. Choose purely on your tasks, which is what the tool below does.

Privacy: training on your chats and how to switch it off

“Training” means the company may use your conversations to improve future models. On consumer plans, the defaults differ.

ChatGPT: OpenAI says it “may use your content to train our models” on individual plans. One catch: if you give a thumbs up or down, that whole conversation may be used even after you opt out (OpenAI Help Center).

Claude: since late 2025, Free, Pro and Max users choose whether chats are used. If you allow it, chats are kept for five years. If not, 30 days (Anthropic). Existing users made this choice in a pop-up in 2025, so check the toggle rather than assume.

Gemini: Keep Activity is on by default if you are 18 or over (Gemini Apps Help). Google uses that activity to improve its services, “including training generative AI models”. A subset of chats is read by human reviewers and kept for up to three years. Google says files you add as Gemini Notebook sources are not used directly for training (Gemini Apps Privacy Hub).

Pay for one or two?

For most people, one paid plan plus two free tiers is the right setup. The free tiers are genuinely useful now. You can use Gemini free for a quick image while paying for Claude, or keep ChatGPT free for a second opinion.

Paying for two makes sense in three situations:

• You need images and top writing. Claude for writing, plus ChatGPT or Gemini for images. The free tiers may cover light image use, so try that first.
• You keep hitting limits. Two $20 plans cost $40 and double your headroom, which is cheaper than any $100 plan. You also get a backup when one service is slow or down.
• Your ecosystem and your best model differ. Say you live in Gmail but prefer Claude’s writing. Google AI Pro for the Gmail integration and storage, Claude Pro for drafting.

Paying for two is waste when both do the same job for you. If you use both mainly for chat and writing, you are paying twice for one job. Our subscription audit checklist helps spot that kind of overlap.

The $100 plans are for people who hit caps daily. Start at $20 and upgrade only after hitting limits for two or three weeks running.

Worked monthly cost examples

These are examples, not recommendations for any real person. Prices are US list prices as of September 2026, before tax.

Example 1: the careful writer. Mostly emails, articles and long PDFs. Claude Pro billed yearly: $200 a year, about $16.67 a month. Images from Gemini’s free tier when needed. Yearly cost: $200.

Example 2: the Google household. Gmail, Docs, Google Photos, an Android phone. Google AI Pro at $19.99 a month is $239.88 a year, and it includes 5 TB of storage. If that replaces a storage plan you already pay for, the real extra cost is lower.

Example 3: the faceless YouTube creator. Scripts, thumbnails and short clips. Claude Pro monthly ($20) plus Google AI Pro ($19.99) for images and video comes to $39.99 a month, or $479.88 a year. Switching Claude to yearly billing brings the year to $439.88.

Example 4: the tight budget. Occasional questions and a few images a week. Google AI Plus at $4.99 a month is $59.88 a year. ChatGPT Go at $8 is $96 a year, but may show ads.

Example 5: the heavy user. Hits limits most days. Claude Max 5x or ChatGPT Pro $100 costs $1,200 a year. ChatGPT’s $100 tier is the cheaper way to get GPT-6 Astra in normal chat, capped at 50 Pro messages a week.

A useful sanity check is cost per use. If you pay $20 a month and use the assistant 60 times, each use costs about 33 cents. If you use it 5 times, each use costs $4, and the free tier would probably do.

Which assistant fits your week

This tool turns the verdicts above into a recommendation. Rate how much of your week goes to each type of task, from 0 (never) to 5 (most days). The weights behind it are our judgement from this article, not a benchmark.

With the default settings (a writing-heavy week, some research and documents, a little of everything else, Google apps and a $20 budget), the tool picks Claude, with Gemini as runner-up. The plan is Claude Pro at $20 a month or $240 a year, and $200 if you pay yearly. Gemini gets close because you picked Google apps. Push images to 4 or 5 and ChatGPT or Gemini takes over. Set budget to "about $40" and you get the two-plan combination instead.

Test all three yourself in one afternoon

Leaderboards and reviews only narrow the field. Your own work decides. This takes about three hours and costs nothing on free tiers, or one month's fee to test paid plans.

1. Open three browser tabs, one for each assistant, all on the same tier (all free, or all paid). Comparing Claude Free with ChatGPT Plus tells you nothing.
2. Paste the same prompt into all three, word for word. Don't rephrase for one of them.
3. Copy each answer into a document labelled A, B and C, in a random order. Score them before you check which is which. This stops brand loyalty from creeping in.
4. Score each answer 1 to 5 on three things: was it correct, could you use it as is, and how many edits would it need.
5. Add up the scores. Look at your two most frequent tasks first, not the grand total.

Here are six prompts to copy. Swap in your own details where marked.

Prompt 1: the awkward email.

"Write a reply to this email. I want to say no politely, suggest a smaller alternative, and keep it under 120 words. Match a friendly but professional tone. Email: 'Hi, could you put together a full 20-page brand guide for our launch next Friday? Budget is tight but we'd love your help.'"

Prompt 2: the document summary. Upload a PDF of 10 to 30 pages that you already know well (a lease, a report, a policy).

"Summarise this document in 5 bullet points for someone deciding whether to sign it. For each point, quote the exact sentence it's based on. Then list anything that looks unusual or risky, and anything you're unsure about."

Prompt 3: the fact check.

"What fees does Etsy charge a US seller for one digital download sold at $10? List each fee, the amount, and a link to the official Etsy page for each." You can check the answer against our Etsy fees guide. Wrong fees or dead links are a clear fail.

Prompt 4: the spreadsheet.

"Here is my sales data. Month, sales in dollars: Jan 420, Feb 380, Mar 510, Apr 460, May 620, Jun 590, Jul 700, Aug 655, Sep 810. Give me the total, the average, the best and worst month, the growth from January to September as a percentage, and a simple chart. Show your working." The correct total is 5,145. Check the rest yourself.

Prompt 5: the plan.

"I have 6 hours a week for a side project, split into weekday evenings (1 hour each, Monday to Thursday) and 2 hours on Saturday. Plan my next 4 weeks to [your goal, for example 'publish my first 4 YouTube Shorts']. Put the hardest tasks in the Saturday slots. Give me a table."

Prompt 6: the product copy.

"Write an Etsy listing for [your product, for example 'a printable weekly meal planner, A4 and US Letter, 12 pages']. Give me 3 title options under 140 characters, a description under 200 words that starts with who it's for, and 13 tags. Don't use the words 'perfect', 'elevate' or 'must-have'."

Prompt 6 includes a banned-words rule on purpose. Following simple instructions exactly is a good test, and a surprising number of answers fail it.

If images matter to you, add a seventh test with a product mockup or thumbnail idea, and run it only in ChatGPT and Gemini.

Common mistakes

• Buying Plus for GPT-6 Astra in chat. On Plus, Astra is only in ChatGPT Work and Codex. Normal chat runs GPT-5.6 Sol.
• Trying to sign up for ChatGPT Pro $200. New sign-ups have been paused since September 10. The $100 tier is open.
• Assuming Claude Pro includes Fable. It doesn't. Fable costs extra credits on Pro.
• Comparing free against paid. Claude Free doesn't have Opus. Gemini Free has a 32,000 token context. Test like for like.
• Judging on one prompt. One great answer is luck. Six prompts, run twice, is evidence.
• Leaving training on with client data. Two of the three train by default. Switch it off before you paste anything private.
• Paying yearly right before a launch window. With DevDay on September 29 and Gemini 4 expected soon, monthly billing keeps you flexible this autumn. Claude's annual discount saves $40, which is not much if you end up switching.
• Stacking three $20 plans. $60 a month is $720 a year. If you use one heavily and the others rarely, you are paying for tools you don't need. Check your real use each quarter with our free subscription audit tool.

FAQ

Is GPT-6 Astra better than Claude Opus 5.5?

On the evidence so far, not for everyday work. Opus 5.5 leads both Artificial Analysis and LMArena. Astra ranks 26th on LMArena's human-preference board. Astra has strong reviews for multi-step agent work, though, so if you mainly want an assistant that carries out jobs across your apps, test ChatGPT Work.

Can I use GPT-6 Astra for free?

Not according to OpenAI's help pages. Free and Go default to GPT-5.6 Luna. Plus users get Astra in ChatGPT Work and Codex. In normal chat, Astra powers the "GPT-6 Pro" option on the $100 and $200 Pro plans.

Which free tier is best?

For writing, Claude Free, which runs Sonnet and Haiku. For images and quick research, Gemini Free or ChatGPT Free. A month on all three free tiers is a cheap way to find your favourite before paying.

Does Claude make images?

No. It can make diagrams, slides and simple graphics through Claude Design and Artifacts, but not photos or illustrations. Use ChatGPT or Gemini for those.

Is Gemini Notebook the same as NotebookLM?

Yes. Google renamed NotebookLM to Gemini Notebook on July 16, 2026. It is still a separate product and syncs with the Gemini app.

Should I wait for Gemini 4 or OpenAI's DevDay?

Don't wait to start. Do wait before committing to a yearly plan. Monthly billing on any of the three lets you switch in a few weeks if something better lands.

Which is the most private?

On default settings, Claude, because it asks you to choose. ChatGPT and Gemini can train on your chats unless you switch it off, and Google also has some chats read by human reviewers. Once you turn training off, the gap narrows.

Is ChatGPT Go worth $8?

If you use ChatGPT daily and keep hitting free limits on uploads or images, maybe. OpenAI says it may show ads, and it still runs GPT-5.6 Luna. At $4.99, Google AI Plus is cheaper and adds video and storage.

Last checked September 26, 2026. Plans, models and limits for all three assistants changed several times this month, and OpenAI's DevDay (September 29) and Gemini 4 may change them again within weeks. Sources: OpenAI Help Center (ChatGPT plans, GPT-6 Pro, data controls); OpenAI (ChatGPT Go, Images 2.5); ChatGPT pricing page (context windows); OpenAI Community (Plus access to Astra); Claude pricing page and Claude Help Center (plans, Max, Fable, context, Google Workspace connectors); Anthropic (Opus 5.5, consumer terms); Google Gemini subscriptions page and Gemini Apps Help (plans, limits, privacy, activity); Google blog (Gemini 3.8 Flash, Gemini Notebook); 9to5Google (Gemini 4); Artificial Analysis and LMArena (rankings); Tom's Guide, Matt Shumer and TechRadar (reviews); The Next Web (ChatGPT Work); TestingCatalog (DevDay); Microsoft (Microsoft 365 plans).