ChatGPT vs Claude vs Gemini vs Grok vs DeepSeek vs Perplexity vs Manus

People usually arrive at this comparison with one question: which AI assistant should I pay for? The honest answer is that the seven products no longer compete on one clean axis. Across the group, the available surfaces span chat, search, files, media tools, coding, and different degrees of task execution. Choosing only by the model name misses the product you will actually use.

The documentation review for this revision was completed July 17, 2026. There is no HUMAI benchmark behind it, and no year-long test or private performance dataset is claimed. Product access, regional availability, prices, and usage limits can change quickly, so the links point to the current official pages rather than freezing every quota into the article.

The practical answer is simple. Start with the job that has to be done, then check the account boundary. Source research depends on inspectable citations; repository work depends on access to files, commands, and tests. A chat subscription and the same company's API are separate purchases. An agent that can click through websites also introduces a different permission problem from an assistant that only drafts text.

AI assistant comparison table

Those descriptions are starting points, not scores. The right product can change between two tasks on the same afternoon.

First separate the app, the model, and the API

Many bad comparisons mix three different things. The app is the interface, storage, tools, connectors, and plan limits you interact with. The model is one component selected by that app. The API is a developer product with its own authentication, terms, billing, and data handling.

A ChatGPT subscription does not turn into OpenAI API credit. OpenAI documents separate billing systems for ChatGPT and the API platform. Claude Code can be included with certain Claude subscriptions or use separate Anthropic API billing, as described in the Claude Code setup guide. Google AI plans bundle consumer features and storage, while the Gemini API has its own pricing. The same separation appears in the official materials for xAI's API, the DeepSeek platform, Perplexity's Sonar API, and the Manus API.

This matters at checkout. Paying for a polished app may buy research mode, larger file allowances, memory, or an agent interface without buying programmatic access. Paying an API bill may buy model calls without giving anyone a consumer workspace.

ChatGPT: the broad workspace option

ChatGPT covers a broad set of everyday jobs without asking a user to leave the product. The current ChatGPT plan page groups chat, search, voice, file uploads, data analysis, vision, projects, memory, image creation, research, and coding access across its tiers. That list describes the product's surface area; it does not show that one underlying model wins every prompt.

OpenAI now explains three distinct work modes. Chat is for conversation and quick questions. Work handles longer research and finished materials. Codex is the dedicated software development surface. The official ChatGPT Work and Codex guide is more useful than a model leaderboard because it shows where the product expects each task to happen.

For web research, Deep Research can use the web, uploaded files, and enabled apps, proposes a research plan, and returns a report with citations. Projects keep files, instructions, and conversations together. Both surfaces remain subject to the account's plan, settings, connected sources, and usage limits.

The tradeoff is product sprawl. Two users on different plans, devices, regions, or staged rollouts may not see the same tools. Before paying, open the live plan comparison and confirm the exact feature, local price, and usage policy that matters to your workflow.

Claude: document work with a serious coding branch

Claude's consumer product is easiest to understand as a writing and analysis workspace that has grown into research, files, projects, and tool use. The current Claude pricing page lists web search, memory, file creation and execution, connectors, extended thinking, projects, Research, and access to Claude Code at different tiers.

Anthropic's own guide separates three modes clearly. Web search is for a quick answer that needs current information. Extended thinking is for difficult reasoning that does not necessarily require the latest web. Research performs a longer tool-driven investigation and returns citations. That distinction appears in Anthropic's mode selection guide.

The coding decision should be made around Claude Code, not around whether a chat response can print a code block. Claude Code operates in a terminal and works with a repository, commands, and developer tooling. It is a different working surface from a browser conversation, even when a subscription provides access to both.

Gemini: the Google ecosystem and media route

Gemini's advantage is not a single answer score. It is the number of Google surfaces that can participate in a task. The Gemini Apps help center documents connected apps, Gems, file and GitHub inputs, Canvas, research, image and video creation, music tools, Audio Overviews, and newer agent features.

Gemini Deep Research uses Google Search by default and can add sources such as Gmail, Drive, uploaded files, and NotebookLM when the account and settings allow it. This is useful when the source material already lives in Google products. It also means the connector scope has to be understood before the task starts.

The Google AI plan page combines assistant access with Google storage, Workspace features, and eligible image, video, or music tools. The exact bundle depends on the viewer's country and account, which makes this a product-and-storage decision rather than a pure model purchase.

Google also documents Gemini Spark, an agent for workflows, schedules, connected apps, and browser or computer use. Availability is restricted by plan, region, and account requirements. Do not assume that seeing Gemini chat means Spark is included.

Grok: web and X context with xAI media tools

Grok is available through xAI's own web and mobile apps and through X. The current Grok product overview documents chat, web access, image and video generation through Imagine, voice, file uploads, and connectors. The xAI pricing page separates free and paid Grok plans from team and API offerings.

The X surface deserves its own sentence because its context and controls differ. X says Grok can search public X posts and the web. The Grok on X help page also explains X-side personalization, training controls, and history deletion. xAI's own privacy policy states that it covers Grok on grok.com and the mobile app, not Grok supplied through X.

Live public discussion on X, voice, and xAI's media tools are Grok's distinct decision factors. X posts still need to be traced to a primary source, and anonymous posts should not be promoted to authority just because the product can retrieve them quickly.

DeepSeek: accessible chat, separate API, distinct data boundary

DeepSeek presents a free consumer entry point on its official site and a separate developer platform. That makes it easy to try without treating the trial as evidence that it will fit a production workflow. The API documentation covers developer access, while API prices and model aliases can change independently of the consumer app.

The public product material reviewed for this update centers on chat access and API use. It does not justify treating DeepSeek as a drop-in substitute for a persistent project workspace, a citation-first research product, or a browser agent. If one of those surfaces is required, verify it in the current app before making the plan decision.

DeepSeek also has a clear data-residency consideration. Its privacy policy says the consumer service collects prompts, uploads, chat history, and technical data, may use information to improve and train services, provides a training opt-out, and stores personal data it collects and processes in the People's Republic of China. That is not a verdict that the product is universally safe or unsafe. It is a documented boundary that some organizations cannot accept for client files, personal data, source code, or regulated work.

Perplexity: research and citations first

Search and citations sit at the center of Perplexity's consumer product. A normal question starts from web search and returns a sourced answer. Research mode runs a longer search and analysis process, then produces a report that can be exported or shared. Model selection is secondary in that mode because the product selects the research stack.

The current plan family includes Standard, Pro, Max, Education, and enterprise options. The official plan guide separates consumer subscriptions, enterprise workspaces, and the Sonar API. It also explains that API access is not a bundle of the web app's premium features.

Perplexity has expanded beyond answer pages. Paid plans can create documents, presentations, spreadsheets, and simple apps. The asset creation guide describes file preview, editing history, downloads, sharing, and export options. Citations are specifically documented for generated documents, so users should inspect the evidence behavior in every other output format rather than assume it is identical.

Manus: delegate a deliverable, not just a response

Manus is the outlier because its product starts with execution. The Cloud Browser documentation describes a browser that can visit sites, click controls, fill forms, extract data, and work inside authenticated services. Users can watch the run and take over when verification or judgment is required. Manus can also work with code environments, generate slides and websites, analyze data, schedule tasks, and connect to external tools.

Manus documents a deliverable surface that includes files, reports, dashboards, slide decks, websites, and completed web workflows. Its data analysis guide lists reports, dashboards, webpages, and slides as output formats. Wide Research handles many similar items in parallel, which is a different job from one deep conversation.

The billing model is another difference. Manus uses credits based on the resources consumed by a task, including model work, virtual machines, and third-party services. The plan documentation recommends monitoring task estimates and usage rather than assuming every prompt costs the same.

Execution power raises the cost of a mistake. Give the agent only the accounts and permissions needed for the task. Use take-over points before sending, publishing, purchasing, deleting, or editing an authoritative system. A browser agent should not receive broad access merely because the prompt sounds routine.

Choose by the job, not by the launch headline

For a general assistant

Start with ChatGPT, Claude, or Gemini. ChatGPT offers a broad single-workspace menu. Claude has a clear document and coding shape. Gemini is attractive when Google services already hold the context. Grok and DeepSeek can also handle general chat, but choose them because their specific access, media, cost, or data boundary fits, not because a one-prompt comparison declared a winner.

For current-web research and citations

Perplexity's product is explicitly organized around research and cited answers, so it belongs on this shortlist. ChatGPT, Claude, and Gemini also have documented research modes with citations, while Grok adds X context. The evaluation question is not how many links appear. Open each citation and check whether it supports the nearby claim, whether the page is primary, and whether its date fits the question.

For long documents and persistent context

Compare Claude projects, ChatGPT projects, Gemini with Drive or NotebookLM sources, and Perplexity projects. Use one representative document set and inspect upload limits, supported formats, citation behavior, retrieval from the middle of a file, export quality, and what happens when a source is replaced. A large advertised context window does not prove that the app will retrieve the right paragraph.

For software development

Compare coding surfaces at the repository level. ChatGPT's Codex and Claude Code are explicit development products. Gemini's GitHub import can help with repository context, while Google's dedicated developer tools are separate products. Grok, DeepSeek, and Perplexity can discuss or analyze code, but that does not make their chat apps equivalent to an agent that can inspect files, run tests, and prepare a patch. Manus can operate a code environment and browser, which is useful for delivery work but requires tighter permission controls.

For image, audio, and video work

The current Gemini and Grok documentation describes explicit native image, audio, or video surfaces. ChatGPT includes image and voice features across its plan family. Perplexity offers image and video generation on eligible plans and can turn research into office files. Manus combines media generation with deliverable creation. Claude's official plan material centers on reasoning, writing, files, and coding, so confirm any required media feature in the current account.

For agentic tasks

Begin with the action boundary. Manus Cloud Browser, Gemini Spark, ChatGPT's agent and Work surfaces, and Perplexity's browser or computer features can move beyond drafting. Ask what the agent can click, which accounts it can reach, whether actions are logged, where confirmation is required, and how to revoke access. The more a tool can do, the less useful a generic model ranking becomes.

Data controls can change the answer

Do not use the logo as a privacy policy. Check the exact product, account type, settings, connector, and contract. Consumer chat, a business workspace, and an API from the same company can follow different rules.

• ChatGPT: OpenAI's Data Controls FAQ explains the consumer training opt-out and Temporary Chat. Its enterprise privacy page says business-service and API data is not used for model training by default unless the customer explicitly opts in to share it.
• Claude: Anthropic documents consumer choices in its model training privacy article. Its commercial data guidance says it does not train on commercial customer data by default.
• Gemini: Google's Gemini Apps Privacy Hub covers Keep Activity, Temporary Chat, connected apps, deletion, and the difference between personal and Workspace accounts.
• Grok: xAI's consumer FAQ explains training controls, Private Chat, and deletion for xAI's Grok apps. Grok on X has separate X settings and policies.
• DeepSeek: the consumer privacy policy documents training use, opt-out, collected content, and PRC storage. Review it before entering sensitive or regulated material.
• Perplexity: the plan guide says consumer users can opt out of AI training, while enterprise plans state that organization data is not used for model training. API terms are separate from consumer subscriptions.
• Manus: individual tasks are private by default, but the task visibility guide says a Team Owner can access team session data. Review the privacy policy, team roles, connector scope, and retention before granting account access.

For confidential work, apply least privilege. Remove personal identifiers where possible, upload only the pages needed, disable training where the product offers that control, and use the appropriate business or API contract when consumer terms do not fit. Never place passwords, recovery codes, private keys, or payment credentials in a prompt.

How to make the plan decision without inventing a benchmark

Use three tasks drawn from the work you already do. One should be a normal mixed request, one should require current sources, and one should use the files or repository that make the purchase worthwhile. If an agent is being considered, add a fourth task that stops before an external side effect.

1. Write the acceptance rule first. Define the required output, sources, file format, maximum review time, and actions the tool must not take.
2. Use the real product surface. Test research in Research mode, coding in the coding workspace, and browser work in the agent. Do not compare a basic chat answer with a multi-step research report.
3. Inspect the evidence. Open citations, verify document references, run generated code, and review every proposed external action.
4. Record friction as well as output. Note missing connectors, upload failures, plan gates, export cleanup, permission prompts, and how much human correction was required.
5. Check the live plan page. Confirm local price, supported country, team controls, current limits, and whether API usage is billed separately.

Many readers do not need seven subscriptions. A sensible stack is often one general workspace plus one specialist only when a recurring task justifies it. That might be ChatGPT or Claude plus Perplexity for source-heavy research, Gemini alone for a Google-centered workflow, or a general assistant plus Manus for controlled execution. The combination should follow the work, not the number of products in a comparison headline.

How to make the final cut

Use the opening table to reduce the field to two candidates, then run the representative tasks above in the exact product surfaces you would pay for. Keep the one that meets the acceptance rule with less correction, fewer permission surprises, and a data boundary the work can tolerate. A permanent seven-product winner is neither necessary nor credible.