AI News, October 2026: Cheaper Models, Agents That Keep Working, Better Product Images

The first week of October brought more AI announcements than anyone running a business has time to read. Most of them change nothing for you this month. A few do.

This is the short version, sorted by usefulness rather than by noise. The figures are the companies’ own claims unless we say otherwise.

1. Good AI got cheaper and faster

Two releases point the same way.

OpenAI released GPT-6.1 Sol, a lower-cost model that the company says comes close to its top model, Astra, on several tests at about one-fifth of the price. It is available in ChatGPT Work, in Codex and through the API, and it is aimed at coding and everyday office work.

Anthropic released Claude Sonnet 5.5, which it describes as more than 30% faster than Sonnet 5 and up to 30% cheaper per task, with better results on bug fixes, documents, slides and spreadsheets.

Why this matters to a small company: the cost of an assistant on your website, of drafted replies or of four hundred product descriptions is mostly the cost of the model underneath. When the mid-priced models become this capable, the work that used to justify the most expensive model no longer needs it.

What does not change: a cheaper model still invents an answer if you do not give it your facts.

2. Agents that keep working after you close the laptop

OpenAI launched “dots”: agents built on GPT-6 Astra that work continuously toward a goal you set, run on their own cloud computers and connect to your applications. They are being offered first to eligible Pro, Business Premium and Enterprise customers.

Google announced Gemini 4 Argon, built for heavy coding, business and cybersecurity work, and able to produce up to a million tokens in one run — roughly a long book of output from a single request. Access starts with vetted security teams while Google tests its safeguards.

This is the direction of the whole industry: from a chat window that answers, to a program that does a job for hours. It is also where the risk moves. An assistant that gives a wrong answer wastes a minute. An agent that acts on a wrong assumption for six hours inside your store or your accounting does real damage.

If you try one, start it on work that is easy to check and easy to undo — sorting enquiries, preparing drafts, building reports. Not issuing refunds.

3. Voice that sounds human, in your customer’s language

ElevenLabs released Eleven v4, with more than 90 languages, several speakers in one recording and sound effects. A Turbo version is made for live voice agents, with a reported delay of about 150 milliseconds before it starts speaking — short enough to feel like a conversation.

Two small speech-recognition models also appeared: Whistle from Cactus, 16.9 MB, seven languages, running on an ordinary processor; and Phonon-2 from Fermion Research, which transcribes an hour of English audio in about 20 seconds on a MacBook Air.

In practice: product videos voiced in ten languages without a studio, and call or meeting recordings turned into text on your own computer, without sending customer conversations to anyone.

4. Image tools that respect the product

This is the part most relevant to anyone who sells online.

• FLUX 3 Image from Black Forest Labs lets you place elements in boxes on a canvas and edit each one separately. You decide where the product, the text and the background go instead of hoping the prompt is understood.
• Ideogram 4.5 is built to keep an image stable through repeated edits. Change the text or the lighting, and the product’s colour and texture are meant to stay as they were.
• NVIDIA’s SoL-Refiner upscales lower-resolution AI video to 4K in one step and works with different video generators.

The weak point of AI product images has always been drift: by the third edit the bottle has a different label. Tools that hold the product still while everything around it changes are the ones worth watching. The rule from our article on ComfyUI and Qwen-Image 2.1 still applies — a person compares every image with the real item before it is published.

What you can safely ignore

Most of the week’s list was for researchers and engineers: open agent models with hundreds of billions of parameters, software for training even larger ones, faster calculations on Huawei chips, a humanoid robot that plays badminton, a watermark for AI-designed proteins, and a Napoleonic letter decoded after 217 years.

Interesting, and some of it will reach ordinary tools in a year or two. None of it changes what you should do on Monday.

What to do with this

1. If you already pay for AI tools, check whether a cheaper model now does the same job. Prices fell this week; your invoice will not fall on its own.
2. If you are considering an agent, pick one boring, checkable task and watch it for a month before giving it more.
3. If you sell in several countries, voice and translation are now good enough to test on one product video.
4. If product images are your bottleneck, the new editing tools are worth an afternoon — with a person approving the result.

And if none of this matches a problem you have, the right amount of AI for your business this month is the amount you already use.

Based on the AI Search newsletter of 4 October 2026 and the companies’ own announcements.