Large Language Models, in plain words.
What an LLM actually is, the major ones to know in 2026, and how to pick the right one for what you are doing.
Over a billion people now use a Large Language Model every month, often without realising it. The smart compose in your inbox. The summary above a Google search. The chat helper your child uses for homework. Different products, same family of technology underneath.
In the next fourteen minutes you will see, in plain words: what an LLM actually is, how they learn, the seven major LLMs to know in 2026, the difference between open-source and closed, why prices have dropped 100× in 18 months, what is new this year, and how to pick the right one for what you are doing.
What an LLM actually is
The acronym sounds technical. The idea behind it is simple.
An LLM is a program that has been trained to predict the next word in a sentence. Then the next. Then the next. Billions of times in a row, faster than you can blink. That is what it does when you chat with it: it is choosing one word at a time, based on everything that came before in the conversation, plus everything it ever read during training.
What makes it surprising is how much it has read. Before talking to you, a modern LLM has absorbed a substantial fraction of the public internet, most public-domain books, most of Wikipedia, most code on GitHub, most discussion forums. So when it predicts what to say next, those predictions are grounded in a huge stretch of human writing.
That is the whole trick. A statistical machine that learned what sentences usually look like, on a scale no one person could match.
How they learn, in three steps
Every modern LLM goes through three training stages. In plain English:
1. Pre-training. The model reads a vast slice of the internet, public-domain books, code repositories, and forums. It learns the statistical shape of language. At this point, the model can produce text that sounds right but might say almost anything.
2. Fine-tuning. Human trainers show the model examples of good answers and bad answers for typical questions. The model adjusts itself to give helpful answers, in the format people expect.
3. Reinforcement Learning from Human Feedback (RLHF). Humans rate hundreds of thousands of answer pairs ("which of these two is better?"). The model learns to prefer answers that are helpful, harmless, and honest.
Think of it like this: pre-training is school, fine-tuning is internship, RLHF is polite-society finishing school. By the end, the model is well-read, useful, and reasonably well-behaved.
The newest models in 2026 add a fourth stage called reasoning: the model is taught to take a moment to think before answering, exploring possible responses internally. We come back to this below.
The big LLMs to know in 2026
There are dozens of LLMs in the world. These seven families cover what most people are using, or hearing about, right now — three closed/proprietary, four open-source.
ChatGPT · OpenAI's GPT family
Made by OpenAI. The brand that launched this wave in November 2022. The biggest user base by far.
What it is best at: general-purpose writing and brainstorming, free voice mode, free image generation. Custom GPTs in the GPT Store let other people share specialised assistants.
Hidden feature worth trying: Custom Instructions. Set your role, tone, and preferences once under Settings → Personalization → Customize ChatGPT, and every chat afterwards respects them. No more re-explaining yourself.
I'm a [your role]. Reply in under 150 words. Include one concrete example. Use British English. Don't apologise or hedge.
Open it at chatgpt.com
Claude · Anthropic
Made by Anthropic. Often praised for clearer writing and stronger reasoning on long documents.
What it is best at: long documents (one of the longest context windows available), careful writing, and code. The free tier handles book-length context.
Hidden feature worth trying: Artifacts. When Claude writes a document, code, or web page, it appears in an editable side panel you can download. Much cleaner than scrolling through a long chat reply.
Paste a long article. Then ask: "Make me a one-page brief with the 3 main arguments, the 3 strongest counter-arguments not mentioned in the piece, and 5 questions a sceptical reader would ask. Put it in an Artifact."
Open it at claude.ai
Gemini · Google
Made by Google. Tightly integrated into Gmail, Drive, Docs, and Calendar.
What it is best at: long context (2 million tokens at the top of the line, the most of any LLM today, roughly two thousand pages of text in a single conversation) and answering questions about your actual Google data.
Hidden feature worth trying: Deep Research. Click the Deep Research button before sending a question and Gemini does multi-step web research, then writes a long, sourced report. Great for "compare X vs Y vs Z" questions.
Compare the three most-recommended beginner podcast microphones under $200. List pros, cons, and price. Cite sources.
Open it at gemini.google.com
Llama · Meta (open source)
Made by Meta. The flagship of the open-source LLM world.
What it is best at: being free to download and run. Anyone can use Llama on their own computer, modify it, or build products on top of it, including commercial use. Llama already powers thousands of independent products you may have used without realising.
Hidden feature worth trying: running it on your own laptop with a free app called Ollama. No data leaves your machine, which makes it suitable for sensitive documents.
Download Ollama from ollama.com. Open Terminal (Mac) or PowerShell (Windows). Type: ollama run llama3.3. Wait for the download. Have a conversation entirely offline. Total cost: $0.
Open it at llama.com or run locally via ollama.com
DeepSeek · open source, China
Made by DeepSeek AI in Hangzhou. DeepSeek V3 was released in December 2024, reportedly trained for under $6 million in compute, and reached parity with the top closed models on most benchmarks. Both their main chat model and their reasoning model (R1) are open-source.
What it is best at: step-by-step reasoning shown out loud. When DeepSeek-R1 thinks through a problem, it lets you watch the internal monologue, useful when you want to understand how it reached an answer or check that it did.
Hidden feature worth trying: the DeepThink toggle in DeepSeek's chat shows you the full reasoning trace before the answer. Helpful when you are trying to learn how to think through a problem yourself, or when you suspect the answer might be wrong.
I have to choose between two job offers. Help me think through it, but show me your reasoning step-by-step before giving your conclusion. Offer A pays more but is further away. Offer B is closer but less interesting work.
Open it at chat.deepseek.com
Kimi K2 · Moonshot AI (open source, China)
Made by Moonshot AI in Beijing. Kimi K2 was released in mid-2025 and has been the fastest-growing open-source model by mindshare in the months since. A trillion-parameter mixture-of-experts model with a strong agentic focus.
What it is best at: agentic tool use (browsing, file reading, multi-step actions) and very long context. K2 was designed from the start to take multi-step actions rather than just answer questions, and to handle book-length material in a single conversation.
Hidden feature worth trying: Kimi's free tier handles long PDFs and document folders that would cost real money on closed services. You can drop a 50-page report or a set of contracts in and chat about them without per-token charges.
Paste a long report. Ask: "Identify the three claims with the weakest evidence, the three that are well-supported, and any contradictions between them. Quote the relevant lines."
Open it at kimi.ai
Qwen · Alibaba (open source, China)
Made by Alibaba's Qwen team. The Qwen 3 series, released in 2025, has been consistently at or near the top of open-source leaderboards. The most-downloaded open-source LLM family on Hugging Face in 2025-2026.
What it is best at: non-English languages (Chinese, Japanese, Korean, Arabic, many more) and high-volume production use. Often the open model deployed inside business products that need a strong but cheap LLM.
Hidden feature worth trying: Qwen offers a vision variant (Qwen3-VL) that reads images, documents, charts, and tables directly. Useful for screenshots of forms, receipts, and data.
Upload a screenshot of a chart from a report. Ask: "Read every value off this chart, write them as a markdown table, and tell me which trend is the strongest."
Open it at qwen.ai
Open source vs closed: what it actually means
When people say an LLM is "open source", they usually mean the model's weights (the numerical settings that make it work) are published. Anyone can download them, run them on their own computer, modify them, or build products on top of them.
The opposite is closed (sometimes called "proprietary"). The model lives on someone else's servers. You talk to it through their app or API. You cannot inspect it, modify it, or run it offline. ChatGPT, Claude, and Gemini all work this way.
What this means for you:
Use closed models if you want the easiest, often best experience and you do not mind your prompts going to a third-party server. This covers most everyday use.
Use open models if you need offline use (no internet), you have strict privacy requirements (legal documents, medical records, internal company code), or you want to learn how LLMs actually work by tinkering.
It is not either-or. Many people use ChatGPT for everyday writing and a local Llama or DeepSeek for sensitive material.
The cost story in 2026
The biggest surprise of the past 18 months is not that LLMs got more capable. It is how much cheaper they got.
Approximate API pricing per million output tokens, today:
• Frontier closed (GPT-5, Claude Opus, Gemini 2.5 Pro): $10 to $75
• Mid-tier (Gemini Flash, Claude Haiku, GPT-5 mini): $0.40 to $5
• Open source via cheap providers (Llama, DeepSeek, Kimi, Qwen): $0.15 to $2.50
The cheapest options are roughly 100 times cheaper than the most expensive for the same task. Quality varies, but the quality gap is much smaller than the price gap.
Two reasons for the gap. First, open model weights are free, so any cloud provider can serve them, and competition between providers drives the price down. Second, the open labs (DeepSeek, Moonshot, Alibaba) have proven you can train competitive models on far less compute than US labs assumed was necessary. DeepSeek V3 was reportedly trained for under $6 million in late 2024, against the rumoured $100 million spent on GPT-4 in 2023. The training-cost gap has reshaped the field's economics.
What is new in 2026
Five shifts reshaping the field this year. Useful to know whether you are a casual user or work in the area.
1. Reasoning models that "think" before answering
OpenAI (o3), Anthropic (Claude Extended Thinking), Google (Gemini Deep Think), and DeepSeek (R1) all introduced reasoning modes in 2025-2026. The model spends seconds to minutes thinking internally before giving its answer.
Reasoning models are dramatically better at maths, coding, logic, and any task where the answer depends on multiple steps. They are slower and cost more. For important questions, the quality jump is real.
2. Massive context windows
Two years ago, a four-page document felt long for an LLM. Today, Claude holds a million tokens of context (roughly 750,000 words, about ten novels). Gemini 2.5 Pro reaches 2 million. You can paste a full codebase, a whole book, or a year of emails and have the model reason across all of it.
3. Native multimodal: text, image, audio, video
The 2026 generation handles every data type natively. Show ChatGPT a photo of a leaky tap and it diagnoses the problem. Play a meeting recording to Claude and it transcribes, summarises, and extracts action items. Send Gemini a 30-minute YouTube video and ask "what are the three best moments". The chat box is no longer the only way in.
4. Real-time voice that feels human
Voice mode in ChatGPT, Gemini Live, and Microsoft Copilot now handles natural back-and-forth, including interruption and emotional tone. You can have a 30-minute walking conversation with an LLM on your phone. Particularly useful for language practice, brainstorming on the move, and accessibility.
5. Agentic AI: models that use tools
The biggest shift. New agents can browse the web, fill forms, write and run code, send emails, all in one session, with you as the conductor. ChatGPT Agents, Claude with tool use, Microsoft Copilot Agent Builder, Google Gemini Gems. Before, you had to copy from one app, paste to the LLM, copy the answer, paste back. Agents do all of that for you. We covered Microsoft Copilot agents in detail in a separate piece.
How to pick one for what you are doing
None of this is gospel. Each LLM does most things well. These are starting points based on what each is currently best at:
• Everyday writing, emails, brainstorming — ChatGPT, Claude, Microsoft Copilot (free web version), or Gemini. All four cover this well. Pick the one you find friendliest.
• Research with cited sources — Perplexity, Gemini's Deep Research, ChatGPT Search, or Microsoft Copilot (which searches Bing). All four show their sources.
• Long documents, contracts, full books — Claude (1M context), Gemini 2.5 Pro (2M context), or Kimi (generous free tier for long PDFs).
• Inside Microsoft 365 apps (Outlook, Teams, Word, Excel) — paid Microsoft Copilot, built on OpenAI's GPT models but living inside the apps you already use.
• Inside Google Workspace (Gmail, Drive, Docs, Calendar) — Gemini.
• Coding — Claude or GPT for general code, DeepSeek or Qwen Coder for cheap-and-strong, or specialised tools like Cursor and GitHub Copilot for in-editor use.
• Reasoning-heavy maths or logic — o3 (in ChatGPT), Claude Extended Thinking, Gemini Deep Think, or DeepSeek R1.
• Agentic tasks (browsing, multi-step actions) — Kimi K2, ChatGPT Agents, Claude Computer Use, or Microsoft Copilot Agent Builder.
• Non-English content (Chinese, Japanese, Korean, Arabic) — Qwen, Kimi, DeepSeek, or Gemini.
• Sensitive documents, offline use — Llama, Qwen, or DeepSeek running locally through Ollama.
• Building cheap at scale — open-source models (DeepSeek, Kimi, Qwen, Llama) via low-cost providers, often 10 to 100× cheaper than frontier closed models.
• Image generation — ChatGPT (DALL-E built-in), Gemini Imagen, or Microsoft Copilot (also uses DALL-E).
• Voice conversations — ChatGPT Voice, Gemini Live, or Microsoft Copilot voice.
If you only learn one thing from this article, learn this: there is no "best" LLM. There is "the right LLM for this task". The cost of switching is two minutes and zero money — try the one you can already sign in to first.
Where to start
Pick one task you do at least once a week, something where good writing or clearer thinking would save you real time. Take it to whichever LLM you can sign into fastest. Spend ten minutes. Notice what worked, notice what did not.
Try one today. The rest will follow.
Which LLM did you try first? What surprised you about its answer? Drop a comment below, I read every one.