Best Large Language Models (LLMs) in 2026: A Practical Comparison

The Best LLMs Right Now: Model Tables

Every model worth knowing about in October 2026, split into closed (paid) models and open-weight models you can download. There are two context columns, because what developers get through the API and what you get in the chat app are often very different numbers. More on that below the tables.

Closed-Source (Paid) Models

Open-Source Models

These are free to download and can be self-hosted, which matters for privacy-sensitive work. Check the licence before you build on one: MiMo V2.6 Pro and DeepSeek V4.1-Flash are MIT and Muse Glimmer is Apache 2.0, but Kimi K3, GLM 5.3, Qwen 3.8-Max and MiniMax M3 use custom licences with conditions for large companies.

Two context windows, not one

Almost every comparison table you’ll find online quotes a single context window figure per model. That figure is the API number, and for most readers it’s the wrong one. What you get through a developer API and what you get inside the consumer chat app are often very different numbers, so the tables above list both.

The gap is widest at OpenAI. GPT-6 Astra and GPT-6.1 Sol accept over a million tokens through the API, but ChatGPT caps the conversation by plan and mode: 54K on Plus for everyday (Instant) chats and 256K when you use a reasoning model, rising to 128K and 400K on Pro. Google is more consistent: 1M in the API and 1M in the Gemini app on a paid plan, though only 32K free and 128K on Plus. Anthropic used to be the strict case, capping chat at 200K while the API ran to 1M, but that’s no longer true: Fable 5.1, Opus 5.5, Sonnet 5.5 and the new Haiku 5.5 all give you the full 1M window on any paid Claude plan, in chat as well as in Claude Code. Older Claude models are still capped at 500K or 200K in chat.

For open models the chat figure depends entirely on who is hosting them, which is why those rows say so.

If You Only Have 30 Seconds

Most people don’t need to read this whole page. Here’s what I tell clients who just want an answer:

• For writing — Claude. It follows nuanced instructions better than anything else I’ve tested. It doesn’t add filler, it doesn’t ignore half your brief, and the tone control is noticeably better.
• For coding — Claude again. I’ve watched developers switch mid-session after seeing the difference. It holds context across long files and catches edge cases the others miss.
• For hands-on professional work (spreadsheets, presentations, data, emails, calendar) — Claude Cowork or ChatGPT’s Codex. This is the category most people haven’t caught up with yet. Instead of just chatting, they work inside a folder on your desktop and drive your browser — so they actually build the spreadsheet, edit the deck, or clear the inbox rather than telling you how. For getting real work done across your files and apps, these two take the top prize.
• For everyday questions and quick research — Gemini or Grok. Gemini is wired into Google’s index, so it’s fast and current. Grok pulls from X/Twitter in real-time, which is useful for anything news-related.
• For deep research — ChatGPT’s deep research mode or Perplexity. Both produce proper cited reports. ChatGPT goes deeper; Perplexity is faster and easier.
• For image and video generation — ChatGPT or Gemini. Claude and Perplexity don’t do this at all.
• If you’re already in Microsoft 365 — Copilot is built into Word, Excel, Outlook, and Teams, so the integration is unmatched. But be aware: Microsoft doesn’t have a frontier model. The underlying AI is behind Claude, Gemini, and even Grok on the leaderboards right now. And tools like Claude can actually work with Microsoft apps in some ways better than Copilot does, through MCP and Cowork. So the integration advantage is narrowing.
• If you want free — Gemini gives you the most on the free tier. Google’s been generous. DeepSeek V4.1-Flash is also free and surprisingly capable for reasoning tasks. And Claude’s free plan now includes Sonnet 5.5, which is a serious model.
• If you want one subscription to try everything — Start with ChatLLMs. It gives you access to all the frontier models (GPT-6, Claude, Gemini, Grok) and the best open-source ones in a single interface for $10/month. It’s the best way to figure out which LLM suits you before committing to one ecosystem. I recommend it to most of my clients as a starting point.

Once you’ve found the model you prefer, move to that platform’s native subscription. The native apps tend to perform slightly better with their own models, and you get the full ecosystem: integrations, agents, desktop apps, the lot.

What the feature table doesn’t tell you

A checkmark means the feature exists. It doesn’t mean it’s good.

ChatGPT has voice chat and it’s the best — genuinely conversational, low-latency, handles interruptions. Claude has voice chat too, but it’s newer and less polished. Perplexity has it, but it’s basically text-to-speech over search results.

Same with “everyday answers.” Gemini and Grok both get the star, but for different reasons. Gemini is fast and pulls from Google’s search index, so it’s great for factual lookups. Grok is plugged into X, so it’s better for “what’s happening right now” questions.

The features that matter most to professionals I work with? Writing quality, coding ability, and deep research. Everything else is nice to have.

What Are AI Agents? (And Why They’re on This List)

You’ll see “Agents” in the feature comparison table at the top of the page. It’s worth explaining what this means, because it’s the biggest shift in how these tools work right now.

An AI agent is a system that can use tools, make decisions, and keep working towards a goal without you supervising every step. Instead of you typing a prompt, reading the answer, typing another prompt, it goes off and does the work: browsing websites, writing code, creating files, calling APIs.

The simple version: perception → reasoning → action → repeat.

You give it a goal. It figures out the steps. It executes them. It keeps going until it’s done or gets stuck.

This matters because most white-collar work happens in browser apps — Salesforce, Jira, Google Docs, email. As soon as agents can operate those tools as well as (or better than) a human, the economics of that work change fundamentally. We’re not there yet for everything, but we’re getting closer every month.

Here’s what agent capabilities look like across platforms right now:

Manus is the most agent-forward. It opens a browser, uses a terminal, builds files, and completes multi-step tasks autonomously. In sessions, most people are blown away the first time they see an LLM using a terminal. It’s still rough around the edges, but it’s improving fast.

ChatGPT has Work mode and Codex, which run longer tasks across your files and apps, plus custom GPTs. Claude has computer use and Cowork mode. Gemini has agent-like features through extensions and Google Workspace integration. Copilot has agents built into Microsoft 365 that can act across Word, Excel, Outlook, and Teams. Grok is experimenting with multi-agent approaches where specialised agents collaborate before giving you an answer.

This is moving fast. I expect this section to look very different in six months.

What Each Platform Is Actually Best At

ChatGPT (OpenAI)

The Swiss Army knife. It does everything — text, images, voice, video, code, research, desktop control. No other platform matches it for breadth. It’s still the best for voice, deep research, and image generation. If you want one tool that covers 80% of use cases, this is a solid all-rounder.

Here’s the thing people don’t realise: ChatGPT’s models are falling behind. On the Arena leaderboard, GPT-6 Astra is not the top model. Claude and Gemini have overtaken it on writing, coding, and reasoning benchmarks. But because it’s ChatGPT — because it’s the name everyone knows — it’s still the most popular. Brand recognition is doing a lot of heavy lifting right now.

There’s also a gap between the headline spec and what you actually get. GPT-6 Astra has a 1.05M-token context window through the API, and that number is what gets quoted everywhere. Inside ChatGPT it is far smaller: 54K on Plus for everyday chats, or 256K if you switch to a reasoning model, and 128K or 400K on Pro. If you are on Plus and you paste in a long report expecting the model to hold all of it, it often won’t. This is the single most common surprise I see in training sessions, and it’s why the model tables split the two figures.

Writing quality is the other weak point. ChatGPT has a recognisable “voice” that’s hard to shake. It loves bullet points, bold text, and phrases like “Here are some key considerations.” In training sessions, I can usually spot ChatGPT output within two sentences.

The bright spot in OpenAI’s ecosystem is Codex. It started life as a coding agent, but it has grown into something that works much like Claude’s Cowork — it operates inside a folder on your machine, runs tasks on your desktop, and drives your browser to get things done. If you live in the ChatGPT world and you want an agent that actually does the work rather than just describing it, Codex is the piece worth paying attention to. It’s the strongest argument for staying in OpenAI’s ecosystem right now.

ChatGPT is also splitting in two. Regular Chat is where you ask questions; ChatGPT Work is where the agentic models live. OpenAI’s newest model, GPT-6.1 Sol, which gets close to Astra at a fifth of the price, is only available in Work and Codex for now. If you’re on Plus and wondering why you can’t find it in the model picker, that’s why.

Claude (Anthropic)

The best model you can actually use right now. Claude holds six of the top seven places on the Arena text leaderboard, with Opus 5.5 in second. The only model above it is Google’s Gemini 4 Argon, which hardly anyone can access yet. The new 5.5 family also changed the maths: Opus 5.5 does Fable-level work for less than Opus 5 used to cost, and Sonnet 5.5, which is on the free plan, beats Opus 5.5 on Anthropic’s own coding benchmark. It follows complex instructions more faithfully than anything else I’ve tested. It doesn’t add unnecessary qualifiers, it doesn’t pad, and it respects tone.

Anthropic is turning into a $100 billion company on the back of this. They’re cleaning up on the enterprise front, and there’s a pattern I keep seeing in sessions: once people start using Claude properly, they don’t go back to anything else. It’s sticky in a way ChatGPT isn’t.

It’s also the strongest coder right now. I’ve run side-by-side tests in sessions — Claude catches things GPT misses, especially in longer codebases.

The other reason Claude has become my daily driver is Cowork. This is the mode that turns Claude from a chatbot into something that actually does the work. Point it at a folder on your desktop and it builds and edits your spreadsheets, presentations, and documents directly; connect it to your browser and it can operate Gmail, your calendar, Linear, and anything else you have open. It’s the closest thing I’ve used to a capable assistant sitting at your machine — and paired with the best model in the world, nothing else touches it for real professional work. This, along with OpenAI’s Codex, is where the most exciting progress is happening right now.

The downside: no image generation, no video, no live camera. If you need those, you’ll need a second tool. But for the actual thinking and working part of AI, nothing else is as good right now.

Gemini (Google)

The dark horse. Gemini has improved more than any other platform in the last year. The 1M context window is real and usable, and unlike ChatGPT you get the full window inside the app on a paid plan — you can drop an entire codebase or a 500-page document into it and get coherent answers.

One caveat on naming: Gemini 3.5 Pro still hasn’t shipped, despite being promised for June, and Google has now jumped ahead to Gemini 4. Gemini 4 Argon, announced on 30 September, went straight to #1 on the Arena text leaderboard, but it’s only available to a small group of cyber-security partners. Google AI Ultra subscribers and paid API customers are next in line, with no date. So the flagship you can actually use is still Gemini 3.1 Pro, from February. What Google has been shipping instead is Flash models — four of them in about a hundred days. The latest, Gemini 3.8 Flash, is genuinely remarkable value: it scored 89.4% on Terminal-bench 2.1 against Claude Opus 5’s 89.1%, at roughly a fifteenth of the price. If you’re running anything at volume, that maths is hard to ignore. Note the promotional API pricing doubles on 1 January 2027.

It’s also the most generous free tier. Google clearly wants market share.

Where it’s ahead of everyone: image and video generation (Veo is impressive), live camera input, and deep integration with Google Workspace. If your team lives in Google Docs and Gmail, Gemini is the natural fit.

Grok (xAI)

The real-time model. Grok’s main advantage is its connection to X/Twitter. For anything happening right now — news, trends, public sentiment — it’s faster than the others because it’s pulling from live social data. It gives you the best window into what humans are actually thinking and talking about right now.

The SuperGrok tier at $30/month is pricier than the competition, and the heavy tier at $300/month is clearly aimed at developers, not regular users.

Grok has also become a serious coding option. Grok 4.6, released in August 2026, was an Opus-class model tuned for coding and agentic work, and Grok 4.7 (21 September) builds on it with a new, larger base model that’s better at long tasks and at checking its own work, at the same low price. That’s why it keeps showing up in tools like Cursor. Its writing still isn’t class-leading, and for most people it remains a supplementary tool rather than a primary one. But for real-time social intelligence, nothing else comes close — and it’s no longer fair to dismiss it on technical ability.

Microsoft Copilot

The enterprise play. Copilot isn’t competing on raw model quality — it’s competing on integration. If your company runs on Microsoft 365, Copilot is embedded directly into Word, Excel, PowerPoint, Outlook, and Teams. No other AI tool can summarise a Teams meeting, draft a follow-up email in Outlook, and update a spreadsheet in Excel without leaving the apps you already use.

The free tier is decent for basic chat. Copilot Pro at $20/month adds priority access and integration with your M365 apps. For businesses, it’s $30/user/month on top of your existing M365 licence.

Here’s what I’m seeing in practice though: the underlying model is falling behind. Microsoft doesn’t have a frontier model of its own — it relies on OpenAI’s GPT under the hood, which as I mentioned above is no longer top of the leaderboards. A lot of companies are being pushed into Copilot by their Microsoft licensing deals, rolling it out to their workforce, and then being disappointed when adoption stays low. People try it, find the output quality lacking compared to Claude or even ChatGPT itself, and quietly go back to whatever they were using before.

The integration is real. But a well-integrated mediocre model is still a mediocre model.

Manus (Manus AI)

The autonomous agent. Manus is different from the others — it’s less of a chatbot and more of a worker. You give it a task (“research these 20 companies and build a spreadsheet comparing their pricing”) and it goes off and does it, using multiple tools: browsing, coding, file creation.

In sessions, most people are blown away the first time they see Manus open a terminal and start executing commands. It’s one thing to ask a chatbot a question. It’s another to watch it actually do the work — browsing websites, writing scripts, building output files. It’s still quite new, but it’s getting better every month.

The credit-based pricing is the catch. You don’t pay a flat monthly rate for unlimited use — you burn credits per task, and complex tasks burn more. This makes costs unpredictable, which frustrates people.

As soon as agents like Manus become as proficient as a human at operating a terminal and browser, this is where you start seeing real disruption to white-collar work. Lots of companies run their operations through browser apps — Salesforce, Jira, Google Workspace. Now you can have agents orchestrate updates for you.

Perplexity

The research assistant. Perplexity isn’t trying to be everything — it’s built for search. You ask a question, it searches the web, reads the sources, and gives you an answer with citations. Think of it less as a chatbot and more as a research analyst.

The Pro tier at $20/month is worth it if you do serious research. It produces what I call “McKinsey-style reports” — structured, sourced, professional. For quick factual questions, it’s often better than asking ChatGPT because it always cites where the information came from.

Why this matters

Native integrations are becoming more and more important as LLMs get more capable. For the first time, we’re able to do work across multiple SaaS systems and have them updated consistently. That should, in theory, bring up both productivity and knowledge for everyone.

There are two ways this works right now. The first is native integrations, where the platform connects directly to your tools through its own protocol. Gemini talks to Gmail, Calendar, and Drive natively. Copilot does the same across Outlook, Teams, Word, and Excel. These are the fastest and most reliable because the machine is using its own protocol.

The second is through the browser. Claude’s Cowork mode, for example, can use the browser on your machine — so it can operate Gmail, Calendar, Linear, and anything else you have open, directly. Manus does this too: it opens a browser, navigates to Jira, and clicks buttons like a human would. It’s more flexible, but it’s slower and occasionally fragile.

Then there’s MCP (Model Context Protocol), which is Claude’s approach to plugging into almost anything — Slack, Linear, GitHub, custom internal tools. It’s the most flexible option, but it requires some setup.

This is still fairly new. People are struggling to make all the integrations work together reliably. But the trajectory is clear: the LLM that connects best to your existing tools will be the one that saves you the most time.

My honest take on value: You must pay for the premium models. The difference between the free tier and a paid subscription is like the difference between a good intern and a senior employee. You would pay extra for someone who can think better and has more experience every single day of the week. It’s the same here.

ChatLLMs at $10/month is a good place to start — try all the models, find your favourite. But you must dive into a proper ecosystem when you’re ready. The native apps perform better, the integrations go deeper, and you get the full feature set.

If you’re stacking subscriptions, the most common combo I see with clients is Claude Pro + Perplexity Pro — roughly $40/month total. Claude for the actual work (writing, coding, analysis) and Perplexity for research. Add ChatGPT if you need image generation or voice.

Personally, I pay £90/month for Claude Max. I rarely hit the limits and it serves me perfectly well for building my AI training business and software for myself and our clients. For what it gives me, it’s absurdly good value.

Open-Source vs Closed-Source: Why It Matters

A year ago, there was a meaningful gap between open-source and proprietary models. That gap has effectively vanished in 2026. Open models like MiMo V2.6 Pro, DeepSeek V4.1-Flash, GLM 5.3, Kimi K3, Qwen 3.8-Max and MiniMax M3 now match or exceed older closed models on most benchmarks, and the best of them sit level with closed frontier models: on the Artificial Analysis Intelligence Index, Xiaomi’s MiMo V2.6 Pro ties Grok 4.7. The full list is in the open-source table at the top of the page.

The notable defection is Meta. It has stopped developing Llama and replaced it with the closed-weights Muse Spark line (now on version 1.3, with an open release only “on the roadmap”), so the company that arguably started the open-weights movement no longer competes in it at the frontier. Its one concession is Muse Glimmer 30B — an Apache 2.0 model distilled from Spark that runs on a single consumer GPU. The open-source torch has passed to the Chinese labs, and now to a phone maker: Xiaomi’s MiMo is the top open-weight model this month.

Why care? Three reasons: cost (open models are free to use), privacy (you can run them on your own hardware), and control (you can fine-tune them for your specific needs). If you’re a business handling sensitive data — legal firms, healthcare, financial services — the open-source option is worth serious consideration.

Most of the professionals I train won’t self-host anything. That’s fine. But if you’re a developer or you work in a regulated industry, knowing these exist matters.