A clearer view of AI.
THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 11 STORIES

AI news, 20 October 2026

MiniMax M3.1 Flash: a very cheap model for running agents, with one big caveat; The US orders DeepSeek out of federal agencies, including tools built on it; Anthropic, Google and OpenAI are all reportedly lining up new models.

Top stories

MiniMax M3.1 Flash: a very cheap model for running agents, with one big caveat

Chinese lab MiniMax released M3.1 Flash, an everyday model for coding, research and agent work with a 1 million token context window, priced at $0.15 per million input tokens and $0.60 per million output tokens. MiniMax says it runs nearly three times faster than its previous model and cuts the cost of agent tasks by 70%; those are the company's own figures. Independent testers Artificial Analysis put it close to DeepSeek's new Flash model for less money, but also found it is unusually likely to give a wrong answer rather than admit it doesn't know.

Why it mattersIf you run high-volume jobs like sorting, drafting or coding tasks, this could cut your AI bill a lot. Use it where answers are checked (code with tests, drafts a person reviews), not for facts you'll publish or send to customers unchecked.

Sourcesx.com post ↗ · venturebeat.com ↗

The US orders DeepSeek out of federal agencies, including tools built on it

The US Department of Homeland Security published an order, signed in July and made public on 19 October, telling federal civilian agencies to stop buying DeepSeek products and remove them from their systems. It also covers third-party products and services built on DeepSeek models, not just DeepSeek's own app. It doesn't apply to businesses or the public, and @kimmonismus links it to China's own new draft rules on exporting its biggest AI models.

Why it mattersIf you sell software or services to the US government, check whether anything you offer runs on a DeepSeek model, including inside tools you resell. For everyone else, it's a reminder to know which model sits under each AI tool you use, because customers and regulators are starting to ask.

Sourcesx.com post ↗ · federalregister.gov ↗

Anthropic, Google and OpenAI are all reportedly lining up new models

The Information reports that Anthropic is preparing a model with better reasoning and intuition, which, according to one unnamed tester, is better at working out what you really want and, when asked to fix one bug, spots the wider type of problem behind it and asks whether to fix that too. The same reporting says Google is working on Gemini 4 Carbon and OpenAI on a cheaper new Terra model. Anthropic hasn't named or dated anything, and @Mr_Salio, also citing The Information, says some of these may slip to November. Treat all of it as rumour for now.

Why it mattersNothing to switch yet. If you're about to commit to a yearly AI plan or build a big workflow around one model, it may be worth waiting a few weeks to see what lands.

Sourcesx.com post ↗ · x.com post ↗ · x.com post ↗

Big lab updates

OpenAI says its newest models are much better at old business code

OpenAI says its latest models are far better at building and migrating software in older or niche languages such as COBOL, Fortran and Visual Basic, which still run a lot of banking and government systems. On its own tests, GPT-6.1 Sol moved old COBOL programs to a modern language with 86.6% accuracy at low effort and 97.9% at higher effort, up from 5.4% for an earlier model. These are OpenAI's own benchmarks.

Why it mattersIf your business depends on an old system nobody wants to touch, from an accounting package to a custom Visual Basic tool, AI-assisted migration is becoming realistic. Start with a small, well-tested piece and check its behaviour matches the old one before moving anything important.

Sourcesx.com post ↗ · openai.com ↗

Meta explains how it trained its small open agent model, Muse Glimmer

Meta published the technical report for Muse Glimmer, the free 30-billion-parameter open model it released in August under the Apache 2.0 licence. It's built for agents that run on your own Mac or PC with a single consumer graphics card, and Meta says it leads open models of a similar size on tool use and multi-step tasks; that's Meta's own comparison. @kimmonismus picked out the most interesting part: Meta trained the agents inside simulated digital lives, with contacts, weeks of messages, calendars and documents, and threw out practice runs where the agent made things up or claimed success it hadn't checked.

Why it mattersA capable agent model that runs privately on your own computer is useful for sensitive work like client files or finances. You'll need some technical setup, so it's one for a tech-savvy team member or freelancer to test.

Sourcesx.com post ↗ · x.com post ↗ · huggingface.co ↗

New tools

Hy-Memory 2.0: one shared memory for all your AI coding tools

Tencent's Hunyuan team open-sourced Hy-Memory 2.0, a memory system for AI agents. It includes MemHub, an app that gathers your conversation history from tools like Claude Code, Cursor and Codex into one folder on your own computer, so something an agent learned in one tool can be recalled in another. It's free under the Apache 2.0 licence and comes with a small model whose only job is deciding what is worth remembering.

Why it mattersIf your team switches between several AI tools, re-explaining your business every time wastes hours. This is developer-level for now, but the idea of owning your own AI memory locally is worth knowing.

Sourcesx.com post ↗ · github.com ↗

Skills & prompting

A free library of about 50 marketing skills for your AI agent

Marketer Corey Haines's Marketing Skills is a free, open-source (MIT) collection of skills for Claude Code, Codex, Cursor and other agents that follow the Agent Skills standard. It covers conversion copy, SEO audits, AI search optimisation, schema, cold email, ad creative, pricing, churn and more, and every skill first reads one shared file describing your product and audience. A popular post called it version 1.2.0 with 29 skills, but the project page now shows version 2.0 with around 50, so it has grown fast. It also lists paid partner tools, which it clearly labels.

Why it mattersIt's a quick way to give an agent a marketer's checklists instead of writing your own from scratch. Fill in the product file first, then try one skill such as the SEO audit on a single page. As with any skill, skim what it does before you install it, because skills can tell your agent to run commands.

Sourcesx.com post ↗ · github.com ↗

Worth testing

AI search tip

Treat your review profiles and comparison pages as AI search assets

A Foundation and AirOps study of 380,000 AI answers about B2B software found that your own product, comparison and use-case pages are the biggest source AI tools cite, but the rest comes from very different places depending on the engine: Google's AI Overviews and AI Mode lean on YouTube and Reddit, while ChatGPT, Claude and Perplexity rely more on review sites like G2. So this week, refresh your main review profile (G2, Capterra, Trustpilot or Google reviews, whichever your buyers use) and make sure you have a clear comparison or "who it's for" page. To check it worked, ask ChatGPT, Perplexity and Google the same buyer question once a month and note which sources they cite. One caveat: it's a study of software brands, and Foundation has a commercial relationship with AirOps, which sells AI visibility tracking.

That's it for today.

Sourcesfoundationinc.co ↗ · x.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.