Technology

New AI Models Released in October 2026: GPT-6.1 Terra, Claude Haiku 5.5, Gemini 4 Argon and More

October 2026 brought a flood of new AI models from OpenAI, Google, Anthropic and DeepSeek. Here's what actually launched, explained in plain English, and what it changes for you if you use ChatGPT, Gemini or Claude.

If you tried to keep up with AI news in October, you probably gave up halfway through. One release tracker counted more than twenty new AI models from seventeen different companies in a single month. Most of them don't matter to you. A few do, and one quietly changed what free Gemini users get.

Here's what actually happened, in plain English.


First, What Is an "AI Model"?

When you chat with ChatGPT, Gemini or Claude, the app you're typing into is just the front door. The thing actually reading your question and writing the answer is the model, the "brain" behind it. Companies keep training new, better brains and swapping them in, which is why the same app can suddenly feel smarter (or occasionally worse) from one week to the next.

Most companies now make two kinds:

  • Big models are the smartest and best at hard problems, but slower and more expensive to run.
  • Small, fast models are a bit less clever, but much quicker and cheaper. Free plans and everyday features usually run on these.

That second group is where October's real story is. The small, cheap models got genuinely good: good enough for most of what normal people use AI for, like writing emails, summarising articles, explaining things and planning trips.


OpenAI (the ChatGPT Company): GPT-6.1 Sol and GPT-6.1 Terra

OpenAI released two models in quick succession.

GPT-6.1 Sol was announced on 29 September at OpenAI's developer event. Think of it as the "very smart but sensibly priced" option. OpenAI says it comes close to its top model at about a fifth of the cost to run. Launch coverage says it's rolling out to paying ChatGPT users first.

GPT-6.1 Terra followed on 26 October. It's smaller and cheaper again, and it's built to be good at "looking" at your screen and images and at everyday questions. Right now it's mostly being offered to companies that build apps on top of OpenAI. OpenAI hasn't said clearly whether, or when, it will become what free ChatGPT users get.

You may also see a story about OpenAI holding back an even bigger model over safety concerns. That comes from a single report, and OpenAI hasn't confirmed it, so I'd treat it as a rumour for now.

What it means for you: if you pay for ChatGPT, answers may feel a bit sharper this month. If you're on the free plan, nothing has visibly changed yet.

ChatGPT itself is still the easiest place to start if you've never used an AI chatbot. The free plan is genuinely useful for everyday writing and questions, and the paid plans mainly buy you the smarter models and higher daily limits. If you want announcements straight from the source rather than from headlines, OpenAI posts every launch on its news page.


Google: A Change That Affects Free Gemini Users

This is the one most worth knowing about, because it may already have affected you.

From 9 October, Google reshuffled which Gemini models you get depending on your plan (for personal Google accounts):

  • Free users now get Flash-Lite, Google's lightest model.
  • Google AI Plus subscribers keep the mid-level Flash model but are losing access to the top Pro model. Google is emailing the exact dates to affected users.

If Gemini's answers have felt a little thinner or more basic lately, that's probably why. It's not your imagination.

Google also announced Gemini 4 Argon on 30 September, its next top-tier model. You can't use it yet, though. It's limited to a small group of approved cybersecurity experts for now, with paying customers promised access "next" and no dates given.

What it means for you: if you rely on Gemini for work or study on the free plan, try asking the same question in another free chatbot and compare. You may find you get better answers elsewhere this month.

Gemini is free to use with any Google account. Its biggest strength is how well it fits into the Google apps you already use, like Gmail, Docs and Search. The paid Google AI plans unlock the stronger models and extra storage. Google announces its AI changes on The Keyword blog, which is worth a glance whenever Gemini suddenly behaves differently.


Anthropic (the Claude Company): Claude Haiku 5.5

Anthropic released Claude Haiku 5.5 on 7 October. Haiku is Anthropic's small, fast model: the "quick answers" member of the family, sitting below the bigger Sonnet 5.5 and Opus 5.5. This is the first Haiku upgrade in about a year.

For most people, Haiku matters less as something you pick from a menu and more as the engine inside apps and tools built on Claude. Faster, cheaper engines usually mean those apps can afford to do more for you, and respond quicker.

If you want to try Claude yourself, claude.ai has a free plan. In my experience it's especially good at two things: writing that sounds natural rather than robotic, and working through long documents you paste or upload. The paid plans, listed on Anthropic's pricing page, mostly raise your usage limits and add the bigger models. Anthropic posts its launches on its news page.


DeepSeek: Quiet Improvements

DeepSeek, the Chinese AI company that made headlines for building strong models cheaply, launched V4.1 Flash back on 10 September. Its main new trick is understanding images, not just text. In October it pushed out smaller updates to that model rather than anything brand new. A bigger "Pro" version is expected but hasn't arrived yet.

You can use DeepSeek for free in its web chat. The answers are surprisingly strong for a free tool, especially for maths and step-by-step reasoning. One thing worth knowing before you paste in anything personal: it's run by a Chinese company, so think about what you share with it, as you would with any free online service.


Also Released in October

Release trackers also logged these. I couldn't check every date against each company's own announcement, so treat the exact dates as approximate:

Mistral Large 4 (6 October) is the latest top model from France's best-known AI company. You can try Mistral's models in its own chatbot, Le Chat, which has a free plan and is a nice option if you prefer a European company.

Meta Muse Spark Open (15 October) is a new model from Meta, Facebook's parent company. As the name suggests, it's listed as an "open" release, the kind developers can typically download and run themselves. Meta shares its AI work at ai.meta.com.

Kimi K3.1 Pro (23 October) comes from China's Moonshot AI, whose Kimi assistant is free to try and is known for handling very long documents.

Google Genie 4 (27 October) is an experimental model from Google DeepMind that generates interactive, game-like worlds. It's fascinating to watch, but it's not something you'll use day to day yet.


October 2026 at a Glance

Date What launched Company In one line
29 Sep GPT-6.1 Sol OpenAI Near-top smarts at lower cost; paid ChatGPT first
30 Sep Gemini 4 Argon Google Next top model; not available to the public yet
7 Oct Claude Haiku 5.5 Anthropic Faster, cheaper Claude for apps and quick tasks
9 Oct Gemini plan changes Google Free users moved to Flash-Lite
26 Oct GPT-6.1 Terra OpenAI Small and cheap; mostly for app makers for now
Oct DeepSeek V4.1 Flash updates DeepSeek Minor improvements

So Which AI Should You Actually Use?

Honestly, the best one is whichever gives you good answers for the things you ask about. Benchmarks and launch-day charts don't tell you that. Here's the simple approach I suggest to friends:

  1. Don't chase every launch. New models arrive every week now. Most of them change very little for regular users.
  2. Run your own quick test. Pick three or four real questions you'd normally ask, like a work email, something you're studying, or a recipe tweak. Ask the same questions in two or three free chatbots and compare. Ten minutes of this tells you more than any review.
  3. Check what your free plan actually gives you. As Gemini just showed, companies change this quietly. Most apps show the model name somewhere in the chat window or settings.
  4. Only pay when you keep hitting limits. If the free version regularly runs out or can't handle what you need, that's when a subscription earns its money, not before.

For Developers: the Short Version

If you build with these models, October's headline is cost. GPT-6.1 Terra is reported at $1 per million input tokens and $5 per million output tokens. Claude Haiku 5.5 (claude-haiku-5-5) and DeepSeek V4.1 Flash (deepseek-flash) sit in the same budget class. Test the cheap tier first, send only the hard cases to a bigger model, and pin model versions in production. Always confirm current prices on the providers' own pages: OpenAI API pricing, Anthropic pricing and the DeepSeek API docs.

I'll do a November roundup next month. If something big drops before then, I'll update this page.