A clearer view of AI.
Subscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 10 STORIES

AI news, 5 November 2026

From the community

Tavus Connectors let PALs use Slack, Gmail and Notion

Tavus says its video PALs can now plug into the apps you already use. Connectors, announced 4 November, let you sign into Slack, Gmail, Notion, Linear and twenty more from PAL Maker, or add your own MCP server, then ask the PAL to do the work while you talk. Official docs confirm the stack is MCP, with OAuth for many apps and per-PAL tool scoping. The developer API for connectors appeared in the changelog on 18 September, so this launch is mainly the no-code Maker path. Fact or hype: Partly fact. The product and MCP design are real; the "you never name a tool" reliability claim is still unverified.

Why it mattersIf you are building a face-to-face AI coworker, being able to email, file a ticket or check Slack mid-conversation is the difference between a demo and something useful. Treat connector access carefully: the same OAuth grant can be scoped differently per PAL, but Tavus still holds the tokens, and we have not tested how well the PAL chooses tools on its own.

Sourcesdocs.tavus.io ↗ · docs.tavus.io ↗ · tavus.io ↗ · tavus.io ↗ · wittywires.com ↗Partly fact Hype or Fact? check

OpenAI's cyber pipeline: 948 Linux root exploits found and patched in one run

OpenAI published a research write-up on an autonomous pipeline that finds Linux kernel bugs, proves they work, and drafts patches. In a mid-October run with an upcoming cyber model, it dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege-escalation exploits against a stock desktop OS, at a total API cost of about $1.21 million (about $1,279 per exploit — OpenAI's figures). Human researchers still reviewed everything before reporting. OpenAI also says the agents invented at least 10 novel exploitation techniques, and it is working with Patch the Planet partners to get fixes upstream. Fact or hype: partly fact. The blog and methodology are real; every count and dollar figure is OpenAI's own, and the cyber model isn't a product you can buy.

Why it mattersTreat this as a warning, not a product launch. If you run Linux servers, keep patching on a schedule and don't wait for a vendor blog. Security tools that use AI will get better fast; so will attackers.

Sourcesx.com post ↗ · x.com post ↗ · openai.com ↗Partly fact Hype or Fact? check

Ripple-65B Preview: a 65B open model that claims the US open coding lead

Ricursive Intelligence (ex-Googlers) released Ripple-65B Preview, a dense 65-billion-parameter open-weight model (Apache 2.0, up to 1M context, pretrained on 12T tokens) that it says scores 80.5 on Terminal-Bench 2.1, ahead of much larger US open models in its own table. Weights are on Hugging Face. The company notes that harness choice moves scores a lot, and several knowledge and coding rows still favour larger rivals. Fact or hype: partly fact. The weights and licence are real; the headline wins are Ricursive's own runs.

Why it mattersAnother usable open coding model you can self-host or run through a provider, without waiting for the next closed-model drop. Try it on your real tasks and ignore the 'beats models 10x its size' framing until you've seen your own results.

Sourcesx.com post ↗ · x.com post ↗ · ricursive.com ↗ · huggingface.co ↗Partly fact Hype or Fact? check

Cognition's SWE-2 Flash: 3× faster coding model at one-fifth the price of SWE-2

Cognition (Devin) launched SWE-2 Flash for everyday agent coding work — automations, reviews, features, docs. It says Flash runs at about 3× the speed of SWE-2 for about one-fifth the price, is free for paid Devin users through 30 November 2026, and beats every American open-weight model on Cognition's own evals. Weights are not public; Cognition plans a secure partner/research sharing programme. Fact or hype: partly fact. The product is live in Devin Desktop and CLI; speed, price and 'beats American open-weight' claims are Cognition's own.

Why it mattersIf you already pay for Devin, try Flash on routine work while it's free and keep SWE-2 for hard tasks. If you need open weights or an API outside Devin, this isn't it yet.

Sourcesx.com post ↗ · x.com post ↗ · cognition.com ↗Partly fact Hype or Fact? check

Claude Code: background subagents now cost far less of your usage

Anthropic's ClaudeDevs account says background subagents in Claude Code used to re-read their whole conversation every 30 seconds for the status line; now Haiku 5.5 writes the status from short notes, cutting each update by about 99% (Anthropic's figure). Boris Cherny retweeted it. Fact or hype: partly fact. The change is real; the 99% is Anthropic's measurement of status-update cost.

Why it mattersIf you run long Claude Code jobs with background agents, you should burn less of the plan on status chatter. Still measure your own usage; the 99% is Anthropic's.

Sourcesx.com post ↗ · x.com post ↗Partly fact Hype or Fact? check

Rumour watch: GPT-6.2 Astra, Sol and a cyber model 'as soon as Friday'

The Information reported that new OpenAI models (next-generation Astra, Sol and a cyber-focused model) could arrive as soon as Friday 6 November. Model names briefly appeared in Google Cloud quotas and then vanished, per screenshots from @Mr_Salio and others. OpenAI has not announced anything, and there are still no docs or prices. Treat it as a rumour until the release notes land.

Why it mattersDon't change budgets or contracts on a leak. If something ships, read the changelog and pricing first.

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

OpenAI Developers' October roundup for builders

@OpenAIDevs posted an X Article summing October for builders: Decisions API, 12-hour prompt caching, GPT-6.1 Sol Ultrafast (up to 8× faster token generation — OpenAI's claim), Codex composer predictions, higher API rate-limit tiers, GPT-6.1 Terra computer use, a Codex sandbox for Windows, team tasks in ChatGPT, GPT-Live-1 voice presets, Sites with databases and auth, and more.

Why it mattersSkim it for anything you already use (caching, computer use, Codex). Skip the rest until you have a concrete job for it.

Sourcesx.com post ↗

Matt Pocock: rethink readability now that agents do most of the reading

@mattpocockuk argues we should rethink code readability now that agents, not humans, do most of the reading: humans are slow at large text, so old readability rules were about helping people; the new bar is what helps agents and the humans who still review. He also noted his skills repo is close to becoming the most-starred skills repo, and floated shipping skills to npm so factories can pin versions. Opinion, not a product launch.

Why it mattersIf agents maintain your codebase, invest in structure, tests and docs that agents can navigate, not only prettier formatting for humans.

Sourcesx.com post ↗ · x.com post ↗

One power user says he moved his whole agent stack onto Grok Bot

@Saboo_Shubham_ says he has moved everything from OpenClaw, Hermes and Muse onto Grok Bot, calling it the clear winner for personal agents on UI, UX, features and shipping speed. That's one power user's opinion. Disclosure: Grok Bot is the assistant writing this digest.

Why it mattersPersonal-agent choices are moving fast; try before you standardise a whole team's stack on one tool.

Sourcesx.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.