Google Gemini vs ChatGPT: which one belongs in your working day

Picture two people. One writes in Google Docs, lives in Gmail and keeps every file in Drive. The other bounces between Excel, Slack, a code editor and a dozen browser tabs. The first should pay Google. The second should pay OpenAI.

That's the decision for most people, and it has less to do with the models than you'd expect. Google shipped Gemini 4 Argon to AI Pro and Ultra subscribers on 13 October 2026; OpenAI's GPT-6 family arrived in September. On the broadest independent index the two flagships now score the same. What separates them is where they sit: Gemini inside Google's apps, ChatGPT as a standalone workbench with a coding agent attached.

Prices from gemini.google and learn.chatgpt.com, checked 22 October 2026. UK readers: Google lists AI Pro at £18.99 and AI Plus at £4.49.

The honest counterpoint: if you don't already use Google's apps, the Workspace integration is worth nothing to you, and ChatGPT's breadth wins comfortably.

Two indexes, two field studies, and no clean winner.

Artificial Analysis Intelligence Index, a blend of ten evaluations: Gemini 4 Argon (high) and GPT-6 Astra (max) both score 53. Neither leads; Claude Opus 5.5 is top at 58, which we cover in Claude vs ChatGPT.

AA-Omniscience, 6,000 knowledge questions that reward right answers, penalise confident wrong ones and don't penalise declining: GPT-6 Astra scores 44, second overall behind Claude Opus 5.5. Argon isn't in the top three (Artificial Analysis, checked 22 October 2026). On recalling facts without a document to hand, OpenAI's flagship is ahead.

A caution about the same board's hallucination column. Its lowest rates belong to small models such as MiniCPM5-1B, at 1%, which mostly decline to answer. A low error rate can simply mean a model that says very little, so read it next to accuracy, never alone.

Summarising a document, the newest models aren't measured yet. Their predecessors were close: OpenAI's GPT-5.5 added unsupported facts to 9.3% of summaries and Gemini 3.1 Pro (preview) to 10.4% (Vectara, updated 16 October 2026).

Field tests of earlier versions were rough on both. In a European Broadcasting Union and BBC study of news answers, 45% of responses from four assistants had at least one significant issue; for Gemini the figure was 76% (EBU/BBC, October 2025). Columbia's Tow Center asked eight AI search tools to identify the source of news excerpts: ChatGPT Search got 134 of 200 wrong and never declined, and Gemini gave a completely correct answer once (Tow Center, March 2025). Both studies predate Argon and GPT-6.

So: on knowledge, ChatGPT's flagship has a small lead; on overall capability, a tie; and the field tests of older versions were harsher on Gemini. None of those gaps is big enough to matter more than checking the answer in front of you. More in AI hallucination rates.

Checking the answer in front of you is the part no index does. Count does it for Claude answers today, tagging each claim as a citation, an assumption, model knowledge or a hallucination.

At around $20 the two plans buy different things.

Google AI Pro ($19.99) gives you Gemini 4 Argon, four times the free usage, Gemini in Gmail, Docs and Vids, 1,000 Flow credits, Antigravity, YouTube Premium Lite and 5 TB of Google storage. If you were already paying Google for storage, part of the subscription is money you'd spend anyway. Limits are compute-based and refresh "every 5 hours until you reach your weekly limit"; hit one and you can carry on with the lighter Flash-Lite model (Google support).

ChatGPT Plus ($20) gives you the Codex coding agent on web, CLI, IDE and iOS, image generation, voice, GPT-6.1 Sol and GPT-6 Luna, and credits to stretch your usage. OpenAI publishes five-hour estimates rather than caps: 15–160 messages with GPT-6.1 Sol and 350–3,000 with Luna on Plus, for its Work and Codex surfaces (learn.chatgpt.com).

A warning on that last phrase. OpenAI's own pages disagree on whether the GPT-6 models appear in ordinary Chat or only in Work and Codex. Open the model picker on your plan before you pay for a flagship you might not see.

The cheap tiers diverge more. Google AI Plus at $4.99 gets you the Flash model, double the free usage, video generation and a 128K context. Google's page also shows a stray line suggesting Argon on Plus; we wouldn't buy Plus counting on it. ChatGPT Go at $8 matches the free plan's model access in Codex.

At the top, Google AI Ultra costs $99.99 (five times Pro's usage) or $199.99 (twenty times). ChatGPT Pro comes at $100, $200 and $500, and OpenAI says Pro plans currently have no five-hour limit. Pay for those only if you hit the ceiling every day.

If you won't pay at all, Gemini is the more generous app. Its free tier includes Deep Research, image generation, Canvas and Gemini Live, running on Flash-Lite (gemini.google).

The catch is the context window. Free Gemini holds 32K tokens, against 1M on AI Pro (Google support). That's enough for a long email thread and nowhere near enough for a full contract or a year of board papers. Paste in more and the model works from part of the document while answering as if it had read all of it. That's how a confident summary ends up missing the clause that mattered.

ChatGPT's free plan is harder to pin down this month. OpenAI's pricing page confirms GPT-6 Luna in the desktop app's Codex and Work surfaces, "subject to rollout". Third-party reports say the free Chat default moved to GPT-6 Luna from 8 October; OpenAI's own pages didn't confirm that when we checked.

For occasional questions, free Gemini is the better deal. For anything long, neither free tier is the right tool.

Both can write and run code on your data, and both vendors would rather you let them. Google's own prompting guide says to use code execution "whenever the model needs to perform any kind of arithmetic, counting, or calculation" (Gemini API docs). A model doing sums from memory is the riskiest thing it does.

Gemini's edge is proximity. On AI Pro it sits inside Gmail and Docs, and the October release added a menu that connects more than 40 apps (Gemini release notes, 13 October 2026). If the spreadsheet is already in Drive, there's no exporting and re-uploading.

ChatGPT's edge is the workbench around the code. It runs Python on uploads, and OpenAI's help centre lists a 512 MB per-file limit, with spreadsheets practically capped at around 50 MB depending on row size. Its apps connect to outside tools over MCP, and Codex is there when the analysis turns into a script.

An illustrative test for either: give it a quarter of sales data and ask for revenue growth by region. Then ask which column it used, whether it compared calendar or fiscal quarters, and what it did with refunds. A good answer names those choices before you ask. Most won't, and the number will look equally confident either way.

Benchmarks tell you how models do on someone else's questions. You need to know how they do on yours. Pick a recent piece of work where you know the right answer, and give Gemini and ChatGPT identical prompts and files.

A two-minute check

1. Ask for something recent. A figure from the last month tests whether each one searched or answered from memory. Ask: "Which of these statements came from a search result, and which from what you already knew?"
2. Open two sources each. Find the sentence the answer leans on. Is it on the page, or is the page merely on the same topic?
3. Make it show the calculation. For any number, ask for the code or formula. Check the inputs, not just the output.
4. Ask what it doesn't know. "What would change your answer?" A model that can name what it's unsure of is worth more than one that can't.
5. Ask twice. A fact that changes between runs hasn't been established.

Whichever wins, the habit matters more than the pick. That's the job Count does: connect it to ChatGPT, or any assistant that supports MCP, and it splits an answer into claims and shows the reasoning chain behind each one, so you can see which rest on a source and which on nothing at all.

We've written about why polished answers get checked least in the confidence bubble.

Is Gemini better than ChatGPT?

Which is more accurate, Gemini or ChatGPT?

Is Google AI Pro worth it compared with ChatGPT Plus?

Can I use Gemini for free?

Is Gemini 4 Argon available in the Gemini app?