The October 2026 AI model wave, explained for non-engineers

If you tried to follow AI news this month, you probably gave up somewhere between the fourth model launch and the second chart full of benchmark scores. I don't blame you. Between late September and the end of October, every major lab shipped something, and most of the coverage was written for engineers.

This is the version for everyone else. What actually came out, what it changes for a business owner or marketer, and the parts that did not change at all, which matter just as much.

What came out, in one table

Here are the releases I could confirm on each lab's own announcement or documentation, in date order.

Some of these started rolling out slightly before October, but they all reached most people this month, so I count them as one wave.

What changed

1. Small models became cheap enough to use on everything

This is the change I would put first for small businesses. Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens (a token is a chunk of text, roughly three quarters of an English word). That applies to prompts up to 100,000 tokens. The previous Haiku cost $1 and $5. OpenAI said it cut API prices for GPT-6 Sol and Luna by 50% compared with its earlier promotional pricing.

In practice, jobs like sorting every enquiry, tagging every review or drafting a reply to every comment now cost cents a month at small-business volumes. If you dismissed automating something last year because of the cost, look at it again.

2. The everyday models got much better at real tasks

The middle-tier models, the ones most people use daily, improved the most. Anthropic reported that Claude Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, a test of coding tasks, up from 10.3% for Sonnet 5, at the same price. Google says Gemini 4 Argon gives more direct answers with less hedging and handles longer multi-step work, and it can take in up to 1 million tokens in the Gemini API.

You will feel this most when you hand AI a task with several steps (research a topic, draft, check against a brief) rather than a single question.

3. AI that uses a computer moved from demo to product

An agent is an AI that completes a task in many steps by itself, using tools such as a browser or files. Anthropic's Haiku 5.5 announcement added beta support for computer use and browser use in its Python and TypeScript software kits, meaning a developer can have Claude click and type through websites and apps. On Anthropic's OSWorld 2.1 test of that skill, Haiku 5.5 scored 72.4%, against 15.7% for Haiku 4.5.

For a business owner, this is the start of AI that fills in the supplier portal or updates the booking system for you. It still needs a developer to set up, and it still needs checking.

4. Open models caught up at the top

Open-weight models publish the model file so anyone can download and run it. This month brought very large ones. DeepSeek released DeepSeek-V4.1-Flash-1017 under the MIT licence, a 763-billion-parameter model that its Hugging Face page lists with a 1-million-token context. Mistral released the weights of Mistral Large 4, a 1-trillion-parameter model, under its own "Mistral Ecosystem Open-Weight" licence. That licence allows commercial use but requires a separate licence for companies that sell the model as a service. Through Mistral's API it costs $0.70 per million input tokens and $2.10 per million output tokens.

These are far too big for a laptop. What changes for a small business is choice: agencies and software you buy can now build on strong models without depending on a single American lab.

5. Search changed under marketers' feet

Google put Gemini 4 Argon into AI Mode in Search for Google AI Pro and Ultra subscribers on 13 October. That is the AI answer box many of your customers now see before any website link. Its answers come from a stronger model, so they will summarise your category with more confidence. Make sure your website states your prices, services and locations plainly, because that is what the AI reads.

What did not change

The models still make things up

None of these launches removed the need to check facts. Even the best scores in the labs' own announcements are below 100%, so some answers are still wrong, and a wrong answer reads just as confidently as a right one. Treat AI drafts like a new junior's first draft: useful, fast, and checked before it goes out.

Not every model reached the public

OpenAI decided not to release its next Astra model in late September, saying it "didn't quite meet the bar in terms of staying within scope and authorization", according to its head of safety systems, Saachi Jain, speaking to CNBC. So at least one lab is now holding a finished model back because of how it behaves as an agent. It also means a launch you read about on social media may never arrive.

The top tier is still expensive

The cheap end got cheaper, but the most capable models did not get cheap. Claude Opus 5.5 is $4 in and $20 out per million tokens. Gemini 4 Argon's introductory $2 and $10 rises to $4 and $20 on 30 November. Use the top models for the hard jobs and the small ones for the volume.

Most businesses still need help using it

Meta published a YouGov survey of 3,983 small business owners in nine Asia-Pacific markets, including India, on 14 October. 84% said they already use AI and 76% said it had a positive effect on productivity, but 86% said small businesses need more guidance, and cost was the top barrier to training, cited by 51%. The models moved faster than most people's ability to use them, and that gap is where the opportunity sits.

A simple way to respond

Here is how I would turn this wave into action for a typical small business.