AI news, 19 October 2026

DeepSeek ships a faster, cheaper Flash model and opens its web search API

DeepSeek released DeepSeek-V4.1-Flash-1017, an update to last month's Flash model, live on its website, app and API. The company says it scores higher on benchmarks, finishes tasks faster and gives shorter answers than before; those are DeepSeek's own figures. In the same thread it opened its Search and Fetch APIs to everyone, which let AI agents look things up on the live web, at $0.75 per 1,000 calls each. Calls made through DeepSeek's tools inside DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on 31 October.

Why it mattersIf you run agents or chatbots that need current information, this is a cheap way to add web search. Test it on a real task before the free period ends on 31 October, and compare the answers with the model you use today.

Sourcesx.com post ↗ · x.com post ↗

• DeepSeek

Anthropic puts 5% of its shares into a public benefit fund

Anthropic, the company behind Claude, is moving 5% of its shares (subject to stockholder approval) into a new Anthropic Public Benefit Fund, with 2% more possible for every 50% rise in the company's value above a baseline it hasn't disclosed. It says the fund could make large donations or direct payments to groups affected by AI, such as workers whose jobs change, and won't be used for lobbying or political giving. Counting the founders' and staff pledges, Anthropic says about 20% of its shares are now set aside for public benefit.

Why it mattersIt's a signal that the big labs expect AI to shake up jobs and want to be seen preparing for it. If AI is changing roles in your team, this is the conversation to have early and openly.

Sourcesx.com post ↗ · anthropic.com ↗

Big brands are starting to block AI agents

Greg Isenberg (@gregisenberg) pointed out that Delta, Chase and Amazon are blocking AI agents from their sites, and sketched how this probably plays out: blocking first, then an arms race, then a market where companies compete for agents with better prices, faster checkout and SEO built for agents. Reporting backs up the trend: Delta says it has no deal letting third-party AI agents book on its platforms, Amazon has blocked some AI shopping agents, and banks are restricting agents from customer accounts. The stages are his prediction, not a fact.

Why it mattersIf your customers start sending agents to buy, book or compare for them, decide now whether your site welcomes them. And if you rely on agents to work on third-party sites, expect some of those flows to break.

Sourcesx.com post ↗ · techcrunch.com ↗

Letting AI agents work inside your ad accounts, safely

A builder thread rounded up the official ways agents can now plug into ad platforms. Google's open-source Google Ads connector only reads data (accounts, fields and reports), so it's safe for reporting. Meta's Ads connector and command-line tool, in open beta since April, can also build campaigns, and new campaigns start paused so nothing goes live until a person switches it on. The suggested split: Google for reporting, Meta for reporting plus drafts, and a human on the switch.

Why it mattersYou can have an agent pull weekly ad reports or draft test campaigns for you, while keeping spend decisions with a person. Start with read-only reporting.

Sourcesx.com post ↗

Inbox Zero Mail: an open-source alternative to Superhuman

Inbox Zero, an AI email assistant that's been around for three years, launched its own email client and was the top Product Hunt launch on 18 October. It puts several Gmail and Outlook accounts in one fast, keyboard-first inbox on web, Mac, Windows and iPhone, with split inbox, bulk unsubscribe and private team comments. The client is free; the AI that sorts every email and drafts replies in your voice is $20 a month, and you can self-host it.

Why it mattersA cheaper way to get Superhuman-style email speed with AI drafts, and the open source code means you're not locked in.

Sourcesx.com post ↗ · getinboxzero.com ↗

• Inbox Zero
• Superhuman (Grammarly suite)

shim-cli: stop your coding agent seeing your secrets

A free, open-source add-on for Claude Code, Codex, Copilot CLI and VS Code. In Claude Code it hides API keys and personal data in what the agent reads before the model sees it, and a watch mode shows what each session sent and cost. No account and no tracking.

Why it mattersIf you or a freelancer build with AI coding agents, this is a simple safety net against keys in a .env file ending up in an AI conversation.

Sourcesx.com post ↗ · getshim.tech ↗

• shim-cli

Move step-by-step jobs out of CLAUDE.md and into skills

A popular post walked through a tip from Anthropic's own Claude Code docs: keep CLAUDE.md for short facts and turn every procedure into a skill, a folder with a SKILL.md file. Only the skill's one-line description sits in memory until the job comes up, so long checklists cost almost nothing until they're used. The docs confirm skills can also run on a cheaper model in their own separate session and switch on only for certain folders.

Why it mattersShorter instructions mean faster, cheaper and more reliable agent runs. Ask Claude Code to audit your CLAUDE.md, split facts from procedures, and show you the changes before it edits anything.

Sourcesx.com post ↗ · code.claude.com ↗

When your agent gets something wrong, save the reason as a skill

A Grok Bot fan account shared a rule from @mattyp: when an agent breaks, don't just fix it. Ask it to explain step by step why it went wrong, find the exact step that failed, and save the fix as a skill so every future run checks for it. The useful caveat: read the skill before you keep it, because a wrong explanation saved as a skill repeats the mistake every time.

Why it mattersThis turns one-off fixes into lasting improvements, which is how an assistant gets better at your business over weeks instead of making the same error again.

Sourcesx.com post ↗

The viral "Andrew Ng graph engineering course" is really DeepLearning.AI's 2024 LangGraph course

Verdict: an excellent hands-on agent course if you code in Python (intermediate level, coding required); non-technical owners should watch just Andrew Ng's ten-minute intro.

Sourcesx.com post ↗Partly fact Hype or Fact? checkSteps to try, cost and caveat ↗

Stripe's Head of Design on 'Zombie UI'

Verdict: a sharp idea to share, not a tutorial.

Sourcesx.com post ↗Steps to try, cost and caveat ↗

Check which of your pages Google's AI already shows

Google Search Console now has a Generative AI performance report (open Performance, then Generative AI) that shows how often your pages appear in AI Overviews and AI Mode, broken down by page, country and device. Look at which pages get impressions and which don't: if only your blog posts show up and your service, pricing or comparison pages never do, those pages are the gap to fix with clear answers and specifics. It counts impressions only, not clicks, so check your analytics for visits too, and recheck the report monthly to see if the fixes moved the numbers.

That's it for today.

Sourcessupport.google.com ↗