AI Model Comparison: Test ChatGPT, Claude, Gemini

Choose the assistant that handles your own task best at a price and usage level you can accept. If you are deciding among ChatGPT, Claude, and Gemini, give each the same input, score the outputs against requirements written in advance, and keep a record of the exact models and plans you used.

No prompts were run through the products for this article, so there are no scores or winners to report. Checked November 1, 2026: the product and plan details below come from the official pages linked beside them. Those pages can change, and the model shown in your account may differ from the one shown here.

Which AI products are you actually comparing?

Start by separating an assistant from a model. ChatGPT is a product through which OpenAI makes different models and tools available. Claude is Anthropic's assistant, with model access that varies by plan. Google's Gemini app likewise offers model choices that depend on account and plan. A result labeled only "ChatGPT versus Claude versus Gemini" leaves out the detail someone would need to repeat it. Record the model name shown in each interface, whether an automatic selection mode was active, the plan, and the date of the run.

One catalog entry in the proposed comparison needs a correction before testing. Gemini Omni is Google's video creation and editing model family, not the general chat assistant to put opposite ChatGPT and Claude on a text task. Google's May 19, 2026 introduction describes Gemini Omni Flash producing video from references and editing video through conversation. For the text example in this article, use the Google Gemini assistant listing and identify the model available in the Gemini app. Keep Omni for a separate video comparison with other video tools, not for this test.

ChatGPT

ChatGPT is a sensible candidate if your task involves drafting, revising, or working with an uploaded file. OpenAI's pricing page lists a Free plan with text chat and limited access to uploads and other tools. Its Plus help page lists Plus at $20 per month, billed monthly, with broader model and tool access; API use is billed separately. The pricing page lists GPT-6 family options on some plans, but access is plan-specific. Use the model picker in your account to capture the exact choice. Plus can still have message caps that vary with demand, so it is the wrong paid choice if your required workload exceeds the allowance you can actually obtain.

OpenAI lists access on the web and supported apps, while its pricing comparison names web, iOS, and Android access. A browser test is easiest to document because you can note the same workflow for each candidate. If you use a file upload for one assistant, check that the other candidates can accept the same file under the plans you have. Otherwise, paste the same source text into all three. Do not compare an upload-assisted answer with an answer that received only your short question.

Claude

Claude is another candidate for a text and document task. Anthropic's pricing page lists Free at $0 and Pro at $20 per month when billed monthly, or $200 billed up front for an annual subscription. It lists chat on web, desktop, and mobile for both. The plan comparison shows Sonnet and Haiku on Free, with Opus among the Pro options. It does not justify treating "Claude" as one fixed model: write down the model selected for your run. Free has limits that apply within five-hour and weekly periods, and paid usage varies with the model, conversation length, and files. If those limits interrupt your normal work, a good single response may still make the free plan the wrong pick.

The distinction between a subscription and a particular model matters here. A Pro subscriber may be able to choose an option that a Free user cannot. Comparing that Pro response with free responses elsewhere answers a legitimate buying question, but it is a comparison of purchasable experiences, not a controlled comparison of equally priced models. Label it accordingly. If you want to decide whether a paid upgrade earns its cost, run a separate free-plan round rather than combining the two rounds into one score.

Google Gemini, rather than Gemini Omni

Use Google Gemini for the third text response. Google's US Google AI plans page lists Google AI Plus at $4.99 per month and Google AI Pro at $19.99 per month; the Pro listing includes access to Gemini models in the Gemini app and higher access to some features. The Gemini help page describes access without an AI subscription and an October 2026 change: for affected personal accounts without a plan, Auto chooses a model for each prompt, with most prompts going to Flash-Lite. If that is your setting, record "Auto" rather than claiming you tested a specific Gemini model. The drawback is practical: an automatic selection can make the exact model harder to hold constant across runs.

Google says Gemini app usage is based on factors including prompt complexity, features, and chat length. The allowance refreshes every five hours until a weekly limit is reached. Availability can depend on country, age, and account type; Google's help page distinguishes personal accounts from work or school accounts. A listed US subscription price is not a promise about another country's checkout price. Gemini is a poor fit for this comparison if the model or feature you need cannot be selected on your account. Check access first rather than substituting an undisclosed model.

How do you choose a fair task and scoring rubric?

Pick something you will actually ask an assistant to do again. The example here is an editorial triage task: turn a short, fictional support note into a customer reply and an internal handoff. It is narrow enough to inspect line by line, yet demanding enough to expose differences in accuracy, restraint, and format. It does not require web browsing or a large file. If your real work is code review, contract analysis, or image interpretation, replace the task and define checks that matter there. A writing-task result should never be presented as evidence that its winner is better at coding.

Write the expected facts before you open any assistant. For this illustrative input, the note says: "A customer says the export button returned an error on Tuesday. They retried once and the download worked. They ask whether any data was lost. The note contains no error code, account identifier, or evidence about data loss." These are invented test materials, not a real customer case. A responsible reply can acknowledge the failed attempt and successful retry, but cannot promise that no data was lost. A useful internal handoff should identify what needs checking instead of declaring an incident resolved.

Set the scoring rule before seeing any output. One workable rubric awards up to four points for factual fidelity, up to three for following the requested format, up to two for usefulness to the customer and support team, and one for brevity without lost meaning. A response earns the full factual score only if it preserves the failed attempt and successful retry and avoids an unsupported claim about data loss. Give zero for factual fidelity if it invents an error cause or assures the customer their data is safe without evidence. That hard stop matters more than a polished tone.

Define what each smaller score means in a note you can reuse. For format, full credit means the response has the two requested parts and no extra material; partial credit means both parts are present but one is hard to use. For usefulness, full credit means the customer gets an appropriate next step and the handoff tells support what to verify. For brevity, reward an answer that says enough without repeating the note. The point values are an editorial choice for this example, not an industry standard. Change them if another error would be more costly in your work.

Decide whether a single failure disqualifies a candidate. In a customer-facing workflow, an invented assurance may be unacceptable even if the total score is high. State that rule before testing. Otherwise, you may find yourself forgiving the mistake in the response whose style you prefer. Keep a copy of the input, rubric, and scoring decisions together. A second reviewer can then see whether the grade follows the rule rather than trusting a winner label.

What prompt should you run on each assistant?

Use this identical prompt in a new conversation with each candidate: "You are helping a support agent draft a reply. Source note: A customer says the export button returned an error on Tuesday. They retried once and the download worked. They ask whether any data was lost. The note contains no error code, account identifier, or evidence about data loss. Write two parts. Part 1: a customer reply of at most 100 words. Acknowledge what happened, do not claim to know whether data was lost, and say what the team needs to check. Part 2: an internal handoff of at most 60 words stating the known facts and the unanswered questions. Use only the source note. Do not invent a cause, a resolution, a timeline, or an account detail."

The prompt makes the source boundary visible. It also gives you several checks that do not depend on whether you like a particular writing style. Save it exactly as submitted, including punctuation and word limits. If an interface silently adds a workspace instruction, project context, connected app, or prior memory, note that condition. A fresh chat reduces carryover from earlier prompts, but it does not by itself prove that every account setting is identical.

For the main round, do not ask follow-up questions or repair an answer before scoring it. A follow-up may be useful in real work, but it changes the task. If you want to measure how well each assistant responds to correction, design a second round with the same correction prompt for all candidates. Keep first-pass and revised scores separate. The ability to fix a mistake is valuable; it does not erase the mistake in a first draft sent without review.

Run the prompts as close together as practical and record the calendar date, plan, platform, visible model setting, any thinking or effort setting, and whether browsing, memory, or connected tools were enabled. For Gemini, note whether Auto selected the model. For Claude and ChatGPT, write down the precise label the interface gives you rather than inferring it from a pricing page. Record refusals, interruptions, and limit messages too. An interrupted run is not a zero-quality response, and quietly rerunning only that candidate would make the comparison harder to interpret.

How should you compare the outputs?

Copy each unedited answer into a private comparison document and hide the product names while you score, if you can. Read the source note first. Mark every statement in the customer reply and handoff as supported, appropriately uncertain, or invented. Check the word limits after checking factual fidelity. A short reply that falsely says "your data is intact" should not win because it sounds reassuring. A cautious reply that never gives the customer a next step should lose usefulness points.

Write a reason beside each score. "Factual fidelity: two of four" is less helpful than "preserved the retry, but said the export completed without confirming what was in the download." If two outputs tie, use your stated priority to break the tie. For this example, factual fidelity takes precedence over style. If both pass that test, choose the one whose reply needs less editing and whose handoff makes the verification work clearer. You can keep a tie if neither offers a meaningful advantage.

One response per product shows what happened on that run, not what will happen every time. If you can afford a repeat, start a new conversation and submit the unchanged prompt again to each candidate. Score those responses separately; do not select the best attempt from one assistant and the first attempt from another. Variation across repeats may matter more to a support team than the highest single score. Do not turn a small set of responses into a percentage success claim.

Keep cost alongside quality rather than folding it into an unexplained overall rating. A free response that meets the safety rule may be enough for occasional drafts. A paid response may be worth its monthly charge if it consistently saves review time on work you do often. The comparison cannot establish that saving without measuring your own editing time and usage. Note the subscription you would actually buy, not the cheapest advertised tier if that tier lacks your chosen model.

How do you choose on budget, team size, platform, and privacy?

Start with budget. Run a free-plan round before paying if the required model and input method are available. ChatGPT's Free plan, Claude Free, and Gemini without a paid AI plan each offer a way to begin, but their access and limits differ. A paid comparison should use the plans you would truly keep: OpenAI lists Plus at $20 per month billed monthly, Anthropic lists Pro at $20 per month billed monthly, and Google's US page lists AI Pro at $19.99 per month. These are subscription prices checked on November 1, 2026, not equal guarantees of usage or model access.

For a team, test with the account type the team would deploy. Anthropic's Team pricing is per seat and includes central billing and administration; an individual Pro score does not establish how a Team workspace will behave. OpenAI's pricing page separates individual and business plans. Google's Gemini limits guidance says its personal-account help page does not describe work or school accounts. If shared administration is required, an individual subscription is the wrong purchase even when its test output is your favorite. Confirm the applicable team terms and rerun the task in the intended workspace before deciding.

Platform can decide the matter before output quality does. Claude's pricing page lists web, desktop, and mobile chat. OpenAI lists ChatGPT access on web and mobile platforms. Google's plan information describes Gemini in Google apps on eligible plans, including Gmail and Docs on AI Pro. If your workflow starts inside Docs, test that workflow rather than assuming a result in a separate chat window tells you how the in-app experience will work. If you mainly work in a browser, hold that platform constant for the first comparison.

For privacy, do not paste a real customer note into three consumer accounts without checking your organization's rules and each product's data settings. The fictional note above avoids personal data while letting you test the writing behavior. OpenAI's data controls guidance explains how a user can choose whether eligible conversations help train models; Anthropic's plan comparison lists model-training controls for individual plans and says Team content is not used for model training by default. Those details are not a blanket approval for confidential material. If your policy requires a managed workspace or forbids external AI services, the consumer-plan test is the wrong basis for procurement.

When should you run a different comparison?

The support-note task is designed to reveal unsupported claims and weak handoffs. It tells you little about searching for sources. If research is your recurring job, give the assistants the same question and judge whether cited pages actually support the answer. Perplexity AI and Kimi.com are relevant additional candidates for a separate search-oriented comparison, based on their directory descriptions. Do not add them halfway through this test or assume their citations are correct without opening the cited material.

A local-model workflow calls for another test again. The directory describes Ollama as a platform for running models locally. That changes the setup and the privacy questions, so a browser-assistant score cannot stand in for it. Likewise, Gemini Omni belongs in a video task where you specify the references, requested edit, and acceptable output. The comparison only becomes useful when every candidate can attempt the same job.

Keep the completed record with your decision: the exact prompt, source input, rubric, date, plan and model labels, unedited outputs, scores with reasons, and any disqualifying mistake. If the preferred answer depends on a model or feature your account cannot keep using, choose the best available alternative or retest after changing plans. The decision is for this task and these conditions. When the task, access, or model changes, run it again.