Before you start
ChatGPT, Gemini, Claude, Meta AI, DeepSeek, Grok: everyone is talking about AI chatbots, and each company says its own is the best. This guide explains what they actually are, who makes them, what each is good and bad at, and how to pick one for what you need. It takes about 15 minutes.
It is written for everyday users in India, not engineers. AI moves fast, so model names and prices below are correct as of 11 October 2026 and will change. The advice on how to choose and how to stay safe will not.
Nobody paid for a place in this guide, and it recommends no single product.
What an LLM is, in one minute
An LLM (large language model) is a computer program trained on a huge amount of text to predict which words should come next. Do that well enough and it can answer questions, write, summarise, translate and write code.
Three things to know before comparing them:
- The chatbot is not the model. ChatGPT is an app; GPT-6 Astra is one of the models behind it. One app can switch between several models, and one model can sit inside many apps.
- "Open" and "closed" models. A closed model is used only through the company's app or service. An open-weight model can be downloaded and run on your own computer or server, so your data never has to leave it. Open-weight usually needs technical skill and a powerful machine.
- Every LLM can be confidently wrong. They produce text that sounds right. That's called a hallucination, and no model on this list is free of it.
The rule: An AI answer is a first draft, not a fact. Check anything that matters: money, health, law, or news.
The main families, side by side
| Family | Made by | Open or closed | Best known for |
|---|---|---|---|
| GPT (ChatGPT) | OpenAI, USA | Closed | The most widely used chatbot; all-rounder |
| Gemini | Google, USA | Closed | Built into Android, Gmail, Docs and Search |
| Claude | Anthropic, USA | Closed | Long documents, careful writing, coding |
| Llama / Meta AI | Meta, USA | Llama open-weight; Meta AI closed | Meta AI inside WhatsApp and Instagram |
| Mistral | Mistral AI, France | Mostly open-weight | European, open models |
| DeepSeek | DeepSeek, China | Open-weight | Strong reasoning at very low cost |
| Qwen | Alibaba, China | Many open-weight | A big range of open models in many sizes |
| Grok | xAI, USA | Closed | Built into X (formerly Twitter) |
| Sarvam | Sarvam AI, India | Open-weight | All 22 scheduled Indian languages |
The families in detail
GPT and ChatGPT (OpenAI)
Latest: GPT-6 Astra, released 3 September 2026, after the GPT-5.6 family in July.
- Good at: almost everything, at a good level. It has the biggest range of features (voice, images, file analysis, apps) and the most tutorials and help online.
- Watch out: the newest models usually reach paid plans first; the free plan typically uses smaller ones. Being the most popular also makes it the most copied: scammers run fake "ChatGPT" apps and websites.
- Good for: people who want one familiar app for many jobs.
Gemini (Google)
Latest: Gemini 4 Argon, announced 30 September 2026. At first it is available only to selected testers and security partners, then to Google's top paid plan and paying developers. Most people use the Gemini 3.8 family.
- Good at: working with what you already have in Google: Gmail, Docs, Drive, Android and Search. It handles images and very long inputs well.
- Watch out: Google renames and reshuffles models often, so the name you read about may not be the one in your app.
- Good for: Android and Google Workspace users.
Claude (Anthropic)
Latest: Claude Fable 5.1 (1 September 2026) for the hardest work, and Claude Opus 5.5 (22 September 2026), which Anthropic says matches it on most tasks at a lower price. Sonnet and Haiku are the faster, cheaper tiers.
- Good at: reading and writing long documents, coding, and multi-step work. Its writing tends to be careful and clear.
- Watch out: the most capable models (Fable) are on paid plans only.
- Good for: students, writers, analysts and programmers working with long material.
Llama and Meta AI (Meta)
Latest open model: Llama 4 (April 2025). In April 2026 Meta stopped the Llama line and moved its own assistant, Meta AI, to a new closed model called Muse Spark.
- Good at: Meta AI is free and already inside WhatsApp, Instagram and Facebook, so it is the AI many Indians use first. Llama 4 can still be downloaded and run privately.
- Watch out: no new open Llama models are coming, and Llama's licence has conditions, so it is not "open source" in the strict sense.
- Good for: quick questions inside apps you already use.
Mistral (France)
Latest: Mistral Large 4, announced 6 October 2026, a very large model. At launch it was available online only; Mistral says the downloadable weights follow later in October.
- Good at: strong open models that organisations can run on their own servers, with a European approach to data.
- Watch out: a smaller consumer app and community than the American giants.
- Good for: businesses that want an open model and control over where their data lives.
DeepSeek (China)
Latest: DeepSeek V4 and V4-Pro (2026), and V4.1-Flash (September 2026). The weights are open under the permissive MIT licence.
- Good at: reasoning, maths and coding, at a very low price, and free to download.
- Watch out: its own app and website store data on servers in China. Several governments have restricted it on official devices. Its answers on politically sensitive topics are filtered. The downloaded open model, run on your own machine, does not send data anywhere.
- Good for: developers who run it themselves. Think carefully before typing anything personal into its app.
Qwen (Alibaba, China)
Latest: Qwen3.8-Max (August 2026), plus many open models, including a 27-billion-parameter model under the Apache 2.0 licence.
- Good at: a wide choice of open models, from ones small enough for a laptop to very large ones, and good multilingual ability.
- Watch out: licences differ from model to model, so check before using one in a product. The hosted service is run from China.
- Good for: developers choosing an open model of the right size.
Grok (xAI)
Latest: Grok 4.7, released 21 September 2026.
- Good at: questions about what's happening right now, because it can draw on posts on X.
- Watch out: what's trending on X is not the same as what's true, and Grok can repeat it.
- Good for: following live conversations, with extra checking.
Sarvam (India)
Latest: Sarvam-30B and Sarvam-105B (February 2026), built in Bengaluru from scratch for Indian languages and released openly under Apache 2.0. Its chat app is called Indus.
- Good at: all 22 scheduled Indian languages, including speech. It's built and hosted in India.
- Watch out: it's newer and smaller, with fewer features and a smaller community than the global leaders.
- Good for: working in Hindi, Tamil, Bengali, Marathi and other Indian languages.
How to choose
Ask what you need, not which model topped a chart this week:
- Everyday questions and writing: any of the big chatbots will do. Use the one already on your phone, and try two for a week before paying for either.
- Indian languages: try Sarvam's Indus alongside Gemini and ChatGPT in your language, and compare the answers.
- Long documents, reports, study material: Claude and Gemini handle very long inputs well.
- Coding: all the leaders are strong. Developers often use Claude, GPT or an open model like DeepSeek or Qwen inside their code editor.
- Privacy matters most: run an open-weight model on your own computer, so nothing leaves it.
- Price: the free tiers are generous enough for most people. Pay only when you hit a limit you actually feel.
About benchmarks: companies announce scores on tests like "SWE" or "MMLU", usually ones they chose themselves. Rankings change every few weeks. Treat them as advertising, not proof.
Staying safe with any AI chatbot
- Never type passwords, OTPs, Aadhaar, PAN or bank details, or your company's confidential documents, into any chatbot. India's Finance Ministry told its own staff in 2025 not to use AI tools for official documents for exactly this reason.
- Check the settings. Most free chatbots may use your conversations to improve their models unless you switch that off. Look under Settings → Data or Privacy.
- Download only the real app, from the official app store and the company's own name. Fake "AI" apps that steal logins and money are common.
- Don't trust a screenshot of an AI answer as proof of anything. It's easy to fake.
- Watch for AI scams: cloned voices, deepfake videos of celebrities "recommending" an investment, and fake "AI trading bots" promising guaranteed returns. Our Deepfake & Investment Scams guide explains how they work.
Key words
- LLM: large language model, the AI behind a chatbot.
- Model vs app: the model is the AI itself; the app is how you use it.
- Open-weight: a model you can download and run yourself.
- Hallucination: a confident answer that is false.
- Context window: how much text a model can read at once.
- Token: a chunk of text, roughly ¾ of an English word, used to measure length and price.