Grok vs ChatGPT vs Claude: Prompting Compared (2026)

TL;DR: Grok vs ChatGPT vs Claude comes down to modes, live data and defaults. As of September 16, 2026, ChatGPT picks reasoning with a Thinking slider, Claude Opus 5 writes longer answers unless told otherwise, and Grok 4.6 always reasons and relies on search tools for current events. Below are ten raw test prompts to compare them yourself.

Grok vs ChatGPT vs Claude: what actually differs when you prompt them?

The biggest differences aren't in how you phrase a request. They're in what sits around it: which model the app picks, whether it looks things up, how long it answers by default, and how much it reasons before replying. Get those right and the same well-built prompt works in all three.

Here is the side-by-side, taken from each vendor's own pages on September 16, 2026:

Two caveats. The Grok modes come from the mode list grok.com served to a logged-out browser, where Expert is described as "Thinks hard · Grok 4.6" and Heavy as "Team of Experts · Grok 4.6". That's app configuration, not a spec sheet, and it can change without notice. And ChatGPT's own help page is careful about plans: Free and Go users "do not have access to GPT-5.6 Sol."

For the head-to-head of just two of these, see our older comparison of ChatGPT vs Claude prompts. This post adds Grok and the September 2026 lineup.

How do you choose the right model inside each app?

Use the app's own control, not the prompt. A common mistake in three-way comparisons: people type "think carefully" and assume they've changed the model.

ChatGPT. OpenAI's help article says: "Asking ChatGPT to “think harder” or “think deeply” won’t automatically switch to a different thinking level." On Plus and Pro plans, Instant stays on Instant. To get more reasoning, move the Thinking slider to Medium, High or Extra High, which OpenAI describes as standard, extended and "The highest thinking level available with GPT-5.6 Sol." Free and Go users get a separate Think option, which OpenAI says "uses GPT-5.6 Luna, not GPT-5.6 Sol."

Grok. The app offers Auto, Fast, Expert and Heavy. Auto is described as "Chooses Fast or Expert". If you compare Grok against ChatGPT on High, compare it on Expert, not Auto, or you're not testing like for like. Our guide to Grok's modes covers the consumer options in more depth.

Claude. Model and effort choices depend on your plan and on whether you have the newer Claude experience, and we couldn't verify a single current picker description for every plan. On the API, Anthropic's models overview recommends starting with Claude Opus 5 "for most workloads" and using Fable 5.1 for "demanding reasoning and long-horizon agentic work".

Whatever you pick, write it down with your results. A comparison without the mode and plan attached is not reproducible.

How does live information differ between the three?

All three can reach current information, but the switch works differently and it changes how you should prompt.

• ChatGPT "may search the web automatically when your question would benefit from current information", according to OpenAI, and you can force it with the Search tool or by typing /. OpenAI lists web search as available on every plan, including for people who are not signed in.
• Claude has a web search toggle under the "+" menu, according to Anthropic's help centre, which also notes: "If you have the new Claude experience, there's no web search toggle. Claude searches the web when it helps." Every searched response includes citations.
• Grok, on xAI's API, has no live data by default. xAI's models page says Grok "has no knowledge of current events or data beyond what was present in its training data" and tells developers to enable Web Search or X Search. X Search can filter by up to 20 handles and by date range.

We couldn't load X's own help pages about Grok on X today, so we're not describing exactly when the consumer app decides to search. The prompting lesson holds either way: if a question depends on today's information, say so explicitly and name the source type you want. "Search recent X posts from the last 48 hours" is a much better Grok prompt than "what are people saying?"

For API users, note a pricing change xAI has announced: from September 21, 2026, X Search moves from per-call billing to per-post and per-profile billing, according to its X Search page. Our older guide to prompting Grok with real-time data covers the search patterns.

How does each model respond to the same instructions?

The same prompt gets tuned differently because each vendor documents different defaults. Three worth knowing before you compare anything:

Length. Anthropic says Claude Opus 5's "default user-facing responses run longer than prior Opus models'" and that lowering effort doesn't reliably shorten them. If you want a short Claude answer, ask for one. OpenAI says GPT-6 Astra "tends to use lists, tables and Markdown to make responses scannable." xAI doesn't document a default style for Grok 4.6.

Asking vs doing. OpenAI says Astra is "more likely to ask for clarification where earlier models would make assumptions." Anthropic says Opus 5 "can also expand the scope of a task". A scoring sheet that rewards "did the task" will penalise one and a sheet that rewards "stayed in scope" will penalise the other. Decide which you care about before you start.

Reasoning. xAI's docs say Grok 4.6 reasoning "cannot be disabled" and defaults to high. On ChatGPT, reasoning depends on your slider. On Claude, Anthropic's API guidance is to control cost through effort rather than by turning thinking off. That means a "quick question" test is not neutral: Grok may be reasoning while ChatGPT on Instant is not.

For why this matters so much, our explainer on reasoning models vs chat models covers the prompting side, and our prompt differences by model reference table lists the API-level differences in one place.

Ten test prompts to compare Grok, ChatGPT and Claude

Here are ten prompts, published raw. Each one tests a specific documented difference. Run each in a fresh chat, on the mode you'd normally use, with memory and custom instructions switched off if your plan allows it.

We have not published the three models' answers. Outputs change with every model update, plan and mode, and we don't spend production API credits on blog experiments. The prompts are the reusable part.

1. Default length (tests verbosity)

Explain what an API rate limit is to a new product manager.

2. Explicit length control

Explain what an API rate limit is to a new product manager in exactly three sentences. No lists, no headings.

3. Current events without instructions (tests search behaviour)

What changed in AI model pricing this week?

4. Current events with an explicit source instruction

Search the web for announcements from the last 7 days about AI model API pricing. List each change with the vendor, the date, and a link to the vendor's own page. If you cannot search, say so instead of answering from memory.

5. Ambiguous task (tests asking vs acting)

Clean up this list for me:
apples, Apples, banana , cherry, cherry, Date

6. Scope discipline

Fix only the spelling in this sentence. Do not change wording, punctuation or style:
"Their going to recieve the the package tomorow, weather or not it rains."

7. Structured output

Return a JSON object with the keys "city", "country" and "population_estimate" for the three largest cities in Japan. Return only the JSON, no explanation.

8. Reasoning under a constraint

A meeting must include Ana, Ben and Chloe. Ana is free 9-11 and 14-16. Ben is free 10-12 and 15-17. Chloe is free 10:30-15:30. Find every 30-minute slot that works for all three, and show how you checked.

9. Honesty about uncertainty

What was the exact revenue of a private company called "Northwind Pottery Ltd" last year? If you don't know, say so and explain how I could find out.

10. Tone transfer

Rewrite this for a customer who is upset, keeping it under 80 words and without apologising more than once:
"Your refund was delayed because our payment provider had an outage. It will arrive in 3-5 business days."

Prompt 9 uses a made-up company on purpose. A good answer admits it can't find the figure; a weak one invents a number.

What should you watch for on each prompt?

Each pair of prompts isolates one documented behaviour. Here's what the vendors' own pages suggest you'll see, so you know what to look for rather than what to expect.

Prompts 1 and 2 (length). The gap between these two answers is the interesting number, not either answer alone. Anthropic's Opus 5 guidance says a short, explicit conciseness instruction is effective, so Claude's answer to prompt 2 should tighten sharply. OpenAI's guidance says Astra leans toward lists and formatting, so watch whether the "No lists, no headings" line in prompt 2 is followed on ChatGPT's Pro option.

Prompts 3 and 4 (search). Prompt 3 is vague on purpose. ChatGPT may search on its own, Claude may search on its own in the newer experience, and Grok's behaviour depends on whether search is available on your surface. Prompt 4 removes the guesswork. If any model answers prompt 4 without links, or doesn't say it couldn't search, mark it down.

Prompts 5 and 6 (initiative and scope). Prompt 5 is genuinely ambiguous: should "clean up" remove duplicates, fix case, sort, or all three? OpenAI documents Astra as more likely to ask. A model that asks a sensible question and a model that makes reasonable choices and says what it did can both score well. Prompt 6 is strict: fix the spelling and change nothing else. The homophones ("Their", "weather") and the doubled "the" are deliberate judgement calls, so decide your rule for them before you score.

Prompt 7 (structured output). "Return only the JSON" is the test. Any leading sentence or code-fence commentary counts against it if you plan to parse the result. Population figures vary by source and year, so don't score accuracy here unless you check a named source.

Prompt 8 (reasoning). There is one correct slot set, and you can check it by hand in two minutes. This is where effort settings matter most, so record the mode. Grok always reasons; ChatGPT on Instant may not.

Prompt 9 (uncertainty). Score the refusal to guess, not the research tips. A confident revenue figure for a company we invented is the clearest failure in the set.

Prompt 10 (tone). Count the words and the apologies. It's a small test, but it shows whether a model follows two numeric constraints at once while changing register.

How should you score the results?

Score each answer on a fixed sheet before you read the next model's reply, so the first answer you see doesn't become the benchmark.

Prompt #: ___   Model: ___   Mode/slider: ___   Plan: ___   Date: ___

Followed the length/format instruction?      0 / 1 / 2
Stayed in scope (did not add unrequested)?   0 / 1 / 2
Correct and checkable?                       0 / 1 / 2
Cited sources when asked (prompts 3-4)?      0 / 1 / 2   (n/a otherwise)
Admitted uncertainty where needed (9)?       0 / 1 / 2   (n/a otherwise)
Would I use this answer without editing?     yes / no

Notes:

Some practical rules make the comparison fairer:

1. Match effort levels. ChatGPT on High against Grok on Expert against Claude on a comparable reasoning setting. Don't compare Instant against Expert.
2. Run each prompt twice. If a model gives two very different answers, note it. Consistency is part of the result.
3. Keep the dates. A score from September means little in November.
4. Weight the prompts you actually use. If you never need JSON, prompt 7 shouldn't decide anything.
5. Blind the reading if you can. Paste answers into a document without the model name and score them before you look. It is surprisingly easy to reward the tool you already prefer.

Which one should you use for what?

We won't rank them. Rankings built on a handful of prompts don't survive the next model update, and vendor positioning is the only stable signal we can cite. Here is what each vendor says its model is for, as of September 16, 2026:

• Grok 4.6: xAI says to use it for "everything else, including code," besides its dedicated audio, image and video models, and describes it as "the most intelligent and fastest model we’ve built." Its X Search tool is the distinctive feature if your work depends on X conversations.
• ChatGPT (GPT-5.6 Sol): OpenAI describes Sol as "designed for complex work across coding, knowledge work and research, cybersecurity, science, computer use, and design." GPT-6 Pro is the top option on eligible plans.
• Claude Opus 5: Anthropic describes it as "For complex agentic coding and enterprise work".

Many people end up using two of the three. If that's you, keep your prompts in one place rather than in three chat histories, so the version that works isn't lost. Our guide to Claude prompt templates vs ChatGPT shows what to change when a template moves between them.

Where does Prompt Architects fit?

Prompt Architects works inside ChatGPT, Claude and Grok through our browser extension, and adds structure to what you type: role, task, format and constraints, with platform-specific optimization. It doesn't choose the model or mode for you; that's still the slider or the mode picker in each app.

It's most useful for exactly the comparison above. Save the ten prompts in your library, run them across the three apps, and keep the version that wins with a note of the model and date. The free plan includes 5 prompt enhancements per day, according to our FAQ page.