ChatGPT vs Claude vs Gemini — 2026 Comparison

OpenAI, Anthropic and Google each sell a chat assistant and a developer API built on their own family of large language models. ChatGPT, Claude and Gemini look alike from the outside: a text box, a reply, a way to attach files. Underneath, they differ in what their models accept and produce, which tools they can call, where you can run them, and what happens to your conversations afterwards.

Most comparisons of the three rank them, usually on the author's impressions or on a benchmark that was outdated a few months later. This one does something narrower and more durable. It sets out what each vendor documents about its own current models and policies as of October 2026, explains what those properties mean in practice, and then looks at writing, the most common everyday use, as a set of considerations you can test for yourself rather than a verdict. Prices are left out deliberately because they change frequently; our AI chatbot pricing overview tracks them separately.

---

The Three Families in October 2026

Model names change several times a year, so treat the names below as a snapshot and check the linked pages for the current list.

OpenAI (ChatGPT)

OpenAI's API models page lists three current GPT-6 generation models: GPT-6 Astra, described as the most capable and the recommended starting point for complex reasoning and coding; GPT-6.1 Sol, described as near-Astra performance at lower cost; and GPT-6 Luna, the most efficient option for high-volume work. All three are documented with a context window of about 1.05 million tokens, a maximum output of 128,000 tokens, text and image input, text output, and built-in tools for function calling, web search, file search and computer use. All support adjustable reasoning effort. OpenAI's image generation is offered through separate image models rather than these text models.

ChatGPT is the consumer and business app that runs on these models. Which model a given ChatGPT plan uses, and how many messages it allows, is set by plan and changes over time.

Anthropic (Claude)

Anthropic's models overview lists four current models: Claude Fable 5.1, for demanding reasoning and long-horizon agentic work; Claude Opus 5.5, which Anthropic suggests as the starting point for most workloads; Claude Sonnet 5.5, described as the best combination of speed and intelligence; and Claude Haiku 5.5, for high-volume, latency-sensitive tasks such as classification and extraction. All four are documented with a 1 million token context window, a 128,000-token maximum output on the standard Messages API, text and image input, text output, tool use, and "adaptive" thinking that the model scales according to an effort setting. Anthropic lists a reliable knowledge cutoff of June 2026 for all four.

The page also documents server-side web search and code execution tools that capable models accept, and notes that Anthropic converts its context figures into rough word counts: 1 million tokens is about 555,000 English words on the current tokenizer.

Google (Gemini)

Google's Gemini API models page names Gemini 4 Argon as its flagship reasoning model, alongside Flash and Flash-Lite models for speed and cost, Live models for real-time voice, text-to-speech models, the Nano Banana image-generation models, Veo for video generation, and Lyria for music. The Gemini 4 Argon model page, updated in October 2026, documents an input limit of 1,048,576 tokens, a default output limit of 65,536 tokens (extendable with a "long decode" option), and input of text, images, video, audio and PDF, with text output. Documented capabilities include search grounding, Google Maps grounding, URL context, code execution, function calling, structured outputs, file search, computer use (in preview) and adjustable thinking levels.

The Gemini app is Google's consumer assistant, and Gemini features are also built into Google Workspace products.

---

Documented Properties Side by Side

Every cell below comes from the vendor's own documentation for its flagship text model as of October 2026. "Not documented" means the page did not state it, not that the feature is absent.

Property	OpenAI GPT-6 Astra	Anthropic Claude Opus 5.5	Google Gemini 4 Argon
Context window	~1.05M tokens	1M tokens	1,048,576 tokens (shared with output)
Maximum output	128K tokens	128K tokens (up to 300K on the batch API with a beta header)	65,536 by default; longer with long decode
Input types	Text, image	Text, image	Text, image, video, audio, PDF
Output types (this model)	Text	Text	Text
Image generation in the family	Yes, via separate image models	No	Yes, Nano Banana models
Video generation in the family	Not documented on the models page	No	Yes, Veo models
Built-in web search	Yes	Yes (server tool)	Yes (Google Search grounding)
Code execution	Not listed on the models page	Yes (server tool)	Yes
Computer use	Yes	Not listed in the overview table	Yes (preview)
Reasoning control	Reasoning effort levels	Adaptive thinking with effort levels	Thinking levels: low, medium, high
Third-party cloud platforms	Not covered on the models page	Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS	Not covered on the model page

Three things stand out. First, raw context size has converged: all three flagships now take roughly a million tokens, so "which one can read the longest document" is no longer the differentiator it was in 2024. Second, Google's family is the broadest in media: its flagship reads audio and video directly, and sibling models generate images, video and music. Third, Anthropic's models are text-out only but are documented on more third-party clouds, which matters for organisations that must keep data inside an existing AWS, Google Cloud or Azure account.

What a context window does and does not tell you

A context window is the maximum amount of text (measured in tokens, pieces of words) that a model can consider at once, including your instructions, attached documents, the conversation so far and its own reply. Our context window explainer covers the mechanics. Two cautions apply when comparing them:

• Fitting is not the same as using well. A model can accept a million tokens and still miss a detail buried in the middle. Vendors' own prompting guides suggest techniques to help: Anthropic recommends placing long documents at the top of the prompt, above the question, and asking the model to quote the relevant passages before answering; Google's Gemini guidance similarly advises putting long context first and instructions at the end.
• Tokens are not words, and tokenizers differ. Anthropic's own page notes that its newer tokenizer fits about 555,000 words into a million tokens, against about 750,000 on earlier models. The same document can use a different number of tokens on each vendor's models.

For how these systems work internally, see How Large Language Models Work; for why "thinking" settings change answers, see Test-Time Compute Explained.

---

Tools: Search, Code and Actions

Modern assistants are less a single model than a model plus tools. The tools available change what each product is good at more than small differences in model quality.

Web search. All three vendors document a built-in search capability for their APIs. Google's is grounded in Google Search, and the Gemini prompting guide explicitly recommends it for recent or obscure facts. Search-grounded answers are easier to check because they come with sources, but they are only as good as the pages retrieved. Multi-step research features that run many searches and write a report are covered in our deep research tools explainer.

Code execution. Anthropic and Google document a code execution tool, in which the model writes and runs code in a sandbox. Google's prompting guide recommends it for arithmetic and counting, tasks where language models working from text alone are unreliable. For data analysis, an assistant that runs code and shows its working is far more checkable than one that reports a number.

Function calling and computer use. All three let developers describe their own functions so the model can request them, which is the basis for building agents and automations. OpenAI and Google also document computer-use tools that operate a graphical interface. Our explainer on function calling vs tool use covers the distinction.

File and document search. OpenAI and Google document file-search tools that retrieve passages from uploaded documents, the managed form of retrieval-augmented generation. If you need to work over a large private document collection, rather than a single long file, that retrieval layer matters more than the context window; see What Is RAG?

---

Data Handling and Privacy

What happens to your conversations depends far more on which product and plan you use than on which company made the model. The documented defaults differ between consumer apps and business or API use.

	Consumer app	Business plans and API
OpenAI	ChatGPT has an "Improve the model for everyone" setting under Settings > Data Controls; switching it off applies to the whole account, per OpenAI's help centre	API data is not used to train OpenAI models unless you opt in; abuse-monitoring logs are kept up to 30 days by default; Zero Data Retention is available to eligible customers, per OpenAI's API data controls page
Anthropic	Model Improvement is a user-controlled setting on Free, Pro and Max; Incognito chats are not used for training even if it is on; feedback is kept up to 5 years, de-linked from the user ID, per Anthropic's privacy centre	By default, inputs and outputs from commercial products, including the API, are not used for training, unless you submit feedback or opt in, per Anthropic's commercial policy page
Google	With Keep Activity on, Gemini app chats are saved and used to improve Google services, including training; a subset is reviewed by humans and kept up to 3 years; activity auto-deletes after 18 months by default (3 or 36 months selectable); with Keep Activity off, chats are still held for 72 hours, per Google's Gemini Apps privacy hub	Work and school accounts may be subject to different terms under Google Workspace's generative AI privacy terms

The practical rule follows from the table: do not put confidential company or client material into a personal consumer account and assume business-grade protections. If your organisation has a business plan or API agreement, use that. If data must not leave your own infrastructure at all, an open-weight model run locally is the alternative (see below).

---

Writing With the Three Assistants

Writing is the most common thing people use these assistants for, and the area where online comparisons are least reliable. Tests of the kind "we asked each model for a story and ours was best" have three problems: the models change every few months, a single sample from a model varies from run to run, and the result depends heavily on the prompt. What follows instead are the writing-specific properties you can check and the vendors' own documented advice, so that you can run a fair test on your own material.

Factual accuracy is a process, not a model choice

All language models can produce fluent, confident statements that are false. Our explainer on why AI hallucinates facts covers why this cannot be fully prompted away. For writing that contains factual claims, the useful questions are about process:

• Can the assistant ground the draft in sources you supply? All three accept long documents. Give the model your source material and ask it to write from that, rather than from its training data.
• Can it show where each claim came from? Anthropic's prompting guide recommends asking the model to extract relevant quotes first, then write from them. That gives you a checklist for verification.
• Does it search, and does it cite? Search-grounded answers carry links you can open. Check that the linked page actually says what the draft claims.
• Is its knowledge recent enough? Anthropic lists a June 2026 knowledge cutoff for its current models; anything after a model's cutoff must come from search or from documents you provide.

No assistant removes the need to check figures, names, dates and quotations against a primary source before publishing.

Tone and voice

Each vendor documents that its models respond strongly to examples and to the form of the prompt itself:

• OpenAI's prompt engineering guide recommends giving high-level tone and behaviour instructions separately from the content (as developer messages or instructions) and including a few input and output examples.
• Anthropic's prompting best practices recommend three to five examples wrapped in tags, explaining the reason behind an instruction, telling the model what to do rather than what not to do, and matching the prompt's own style to the output you want. Removing markdown from your prompt, for example, tends to reduce markdown in the reply.
• Google's Gemini prompting strategies recommend clear constraints, a few consistently formatted examples (warning that too many can cause overfitting), and starting the output yourself so the model continues the pattern.

The common thread is that a paragraph of your own writing, pasted in as an example of the voice you want, does more than any adjective. "Write in a warm, professional tone" means something different to every model; "match the style of these two paragraphs" does not.

Long documents and coherence

With context windows near a million tokens on all three, the limit on long-form writing is rarely what the model can read. It is how much it writes coherently in one go and how well it keeps track of earlier decisions. Documented output limits differ: 128,000 tokens for the OpenAI and Anthropic flagships on standard calls, and 65,536 by default for Gemini 4 Argon. All of these are far longer than a typical article, so in practice the better workflow for long pieces is the one professional writers already use with people: agree an outline, write section by section, and keep a short running document of decisions (terminology, names, the argument of each section) that you paste back in each time.

Editing and rewriting

Rewriting tasks (shorten this, make it plainer, change the register for a different audience) are where all three tend to be most useful and least risky, because you supply the facts and the model only changes the form. The risk is subtle drift: a shortened version that drops a qualifying clause, or a "simplified" sentence that becomes wrong. Compare the edited text with the original for meaning, not just for style. Asking for a list of what was cut or changed alongside the rewrite makes that review faster.

A fair test you can run in an afternoon

Step	What to do	Why
1	Pick three real tasks you do often (for example, a client email, a 1,000-word explainer from notes, a rewrite for a general audience)	Your tasks, not a vendor's demo
2	Write one prompt per task, with your source material and one example of the voice you want	Same inputs for every assistant
3	Run each prompt at least twice on each assistant	Outputs vary between runs
4	Remove the assistant names and read the outputs blind	Brand expectations bias judgement
5	Score each on factual errors against your sources, edits needed before use, and voice match	Measures that matter for publishing
6	Repeat when a vendor releases a new model	Rankings do not survive model updates

This is a small version of what model developers call evaluation; our LLM evaluation explainer describes the full approach, and How to Write AI Prompts That Actually Work covers prompt construction in more depth.

Using more than one assistant

Some writers split a job across assistants, for example brainstorming in one and drafting in another. There is nothing wrong with that, but it is not inherently better. Each handover loses context, and every draft still needs the same fact-check and edit at the end. A single assistant used with good source material and a clear example of the target voice is usually a simpler starting point, with a second one brought in for a specific reason, such as Gemini's direct audio and video input for working from recorded interviews, or a search-grounded assistant for checking recent facts.

Disclosure and responsibility

Whichever assistant you use, the writer or publisher remains responsible for what is published under their name, including errors the assistant introduced. Many publications and platforms have their own rules on disclosing AI assistance; check the policies that apply to you.

---

Open-Weight Alternatives

Not every use needs a hosted assistant. Open-weight models, whose parameters can be downloaded and run on your own hardware, include Meta's Llama family and several of Mistral AI's models, among others. They are useful when data must stay on your own infrastructure, when you need to fine-tune on proprietary material, or when you want predictable behaviour that does not change with a vendor's update. Google also publishes open models under its Gemma name, separate from Gemini.

The trade-offs are operational: you supply the hardware, handle updates and safety filtering yourself, and the largest open models need substantial GPU memory. Our explainers on Mistral AI and running models locally with Ollama cover where to start.

---

How to Choose

Instead of a winner, here are the documented differences most likely to decide the question for a given need.

If your main need is...	The documented property that matters	Check this
Working from audio or video recordings	Native audio and video input	Gemini 4 Argon accepts both; the OpenAI and Anthropic flagships list text and image input
Generating images or video alongside text	Media generation in the same family	Google (Nano Banana, Veo) and OpenAI (separate image models) document it; Anthropic does not
Running inside an existing AWS, Google Cloud or Azure account	Third-party cloud availability	Anthropic documents Bedrock, Google Cloud and Microsoft Foundry IDs
Very long single documents	Context window and output limit	All three near 1M tokens in; output limits differ (128K vs 65,536 default)
Checkable numbers and analysis	Code execution	Documented by Anthropic and Google as built-in tools
Current facts	Search grounding	Documented by all three
Strict privacy for business data	Plan-level data terms	Business and API terms of each vendor, not consumer settings
Writing in your own voice	How each responds to your examples	Run the blind test above on your own material

Most people will be well served by any of the three for everyday tasks. The differences that matter tend to be specific: a data-residency requirement, a need to work from video, a team already living in one company's office suite. Identify yours, check the current documentation, and test on real work.

---

Frequently Asked Questions

Q: Is Claude the same as ChatGPT? A: No. ChatGPT is made by OpenAI and runs on OpenAI's GPT models. Claude is made by Anthropic, a separate company, and runs on Anthropic's own Claude models. Gemini is Google's assistant and model family. They are trained separately and have different features, tools and data policies.

Q: Which one has the largest context window? A: As of October 2026 the flagships are close. OpenAI documents about 1.05 million tokens for its GPT-6 models, Anthropic 1 million for its current Claude models, and Google 1,048,576 for Gemini 4 Argon. Tokenizers differ, so the same document can use a different number of tokens on each.

Q: Which assistant can understand audio and video? A: Among the three flagships, Google documents direct audio, video and PDF input for Gemini 4 Argon. OpenAI's GPT-6 models and Anthropic's current Claude models are documented with text and image input. Each company also offers separate speech tools or apps, so check the product you actually use.

Q: Do these companies train their models on my conversations? A: It depends on the product and plan. OpenAI and Anthropic state that API and commercial data is not used for training by default, while Google says work and school accounts may be subject to different terms. Consumer apps have their own settings: ChatGPT's "Improve the model for everyone" toggle, Claude's Model Improvement setting, and Gemini's Keep Activity setting. Check them in each app.

Q: Which AI is best for writing? A: No ranking survives the next model release, and results depend on your prompts and material. Run the same real writing tasks through each assistant with your own source documents and an example of your voice, compare outputs blind, and count the factual errors and edits each needs before use.

Q: Will any of them stop making factual mistakes? A: Not reliably. All large language models can produce confident false statements. Giving the model source documents, asking it to quote its evidence, using search with citations, and checking every figure against a primary source reduce the risk but do not remove it.

Q: Can I run any of these models on my own computer? A: Not the hosted ChatGPT, Claude or Gemini models. For local use, open-weight families such as Meta's Llama, several Mistral models, and Google's separate Gemma models can be downloaded and run on your own hardware, with trade-offs in setup and capability.

---

Sources

• OpenAI — Models (API documentation). developers.openai.com
• OpenAI — Data controls in the OpenAI platform. developers.openai.com
• OpenAI — Prompt engineering guide. developers.openai.com
• OpenAI Help Centre — ChatGPT data controls. help.openai.com
• Anthropic — Models overview. platform.claude.com
• Anthropic — Prompting best practices. platform.claude.com
• Anthropic Privacy Center — consumer and commercial training policies. privacy.claude.com
• Google — Gemini API models and Gemini 4 Argon model page. ai.google.dev
• Google — Prompt design strategies. ai.google.dev
• Google — Gemini Apps Privacy Hub. support.google.com

---

Related: How Large Language Models Work | AI Hallucinations Explained | How to Use Claude AI