AI Briefing — October 18, 2026: Replit Clarifies the Anthropic Story, Meta Drops Its Scraping Language, and Anthropic Audits Its Own AI Workplace

The big picture

Saturday was a day for walking things back and writing things down. The week's loudest claim — Replit's CEO saying onstage that the company was "no longer working with Anthropic" — got a quieter, more precise version: OpenAI is now the default, Anthropic is still on the platform, and the difference matters because most users never change a default. Meta removed a contract provision that lawyers had flagged as a possible license to scrape business websites for AI training, insisting nothing about its crawling had changed. And Anthropic published the most self-aware document of the week: a report on what a year of building your whole company around your own AI did to the company's hiring, meetings, and work itself.

What happened today

Muse AI: no new launch — the lineup stands as covered this week

What happened: There was no new Muse AI model or product launch on October 17–18. The state of play is unchanged from this week's coverage: Muse Wonder and Muse Wonder Full launched October 12, with the standard Wonder available through the Meta Model API at $4/$20 per million input/output tokens and via Muse Code, while Wonder Full is reserved for Muse Code subscribers; neither model offers downloadable weights. Muse Spark Open, the open-weight release of October 15, remains the downloadable option, and yesterday's briefing covered it running at over 36 tokens per second on a four-Mac-Studio cluster. Searches today surfaced explainers and recaps of those launches, not new announcements, so we are not dressing old news up as new.

Who it matters to: Anyone evaluating Muse this weekend. The decision is the same one it was on Friday: hosted Wonder through the API, subscriber-only Wonder Full, or self-hosted Spark Open if you have serious hardware.

What you can do: If you are testing, benchmark the API-accessible Muse Wonder on your own tasks rather than the subscriber-only Full variant, and treat Spark Open's hardware requirements — a multi-Mac-Studio class setup for local runs — as part of its real price.

Learn more: Crowdfund Insider's recap of the Muse Wonder launch and pricing and yesterday's briefing on this site for the Spark Open Mac Studio results.

Gemini Spark: no new launch — Task mode is still the story

What happened: There was no new Gemini Spark announcement on October 17–18 either. The October 12 redesign still defines the product: Spark is folded into the main Gemini app as Task mode, with a Chat / Task switch above the prompt box, a Scheduled page grouping tasks as Active, Paused, Completed and Suggested, and Spark settings renamed to Task settings, per 9to5Google. Access still requires a paid Google AI Pro or AI Ultra subscription and a personal Google account, and the redesign was spotted on Android first. Nothing in today's searches changed the rollout picture or the pricing.

Who it matters to: Spark subscribers still looking for a Spark entry that no longer exists in the navigation drawer, and anyone comparing consumer agents this weekend — the honest comparison today is between products as they stood all week, not new releases.

What you can do: In the Gemini app, use the Task switch above the prompt box and check the Scheduled page for anything Spark is running for you. If the switch has not appeared, it is a server-side rollout — there is nothing to install.

Learn more: 9to5Google on Spark becoming Task mode

Replit clarified: OpenAI is now the default, Anthropic is still available

What happened: TechCrunch updated its October 16 report over the weekend with a clarification from Replit CEO Amjad Masad. After saying onstage at Disrupt that Replit was "no longer working with Anthropic," Masad clarified that customers can still access Anthropic models on the platform, but OpenAI's ChatGPT is now the default powering Replit Agent. "Replit remains model agnostic," he told TechCrunch. "Our users can choose Anthropic's Claude or other supported models… OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership." The original onstage claim was rejected by a person familiar with the partnership, who said it remains in place, and Replit was promoting Claude models in production as recently as late September. Masad also repeated the growth figures he gave onstage: enterprise is about 15% of revenue this year, up from about 2% last year, with more than 50% of revenue now B2B — his figures, not audited disclosures.

Who it matters to: Replit users, first: the model running your Agent by default has changed even though the menu has not. Everyone else should note the mechanism — in a market where most users stick with defaults, "preferred partner" is worth almost as much as an exclusive, and announcing it on a conference stage is now part of how labs compete for platforms.

What you can do: If you build on Replit, open your Agent settings and confirm which model is actually running your projects; if Claude was part of why you chose the platform, select it explicitly rather than assuming it is still the default. If you buy AI tools, ask vendors which model is the default today — not which models are "available."

Learn more: TechCrunch's updated report on Replit, OpenAI and Anthropic

Meta removed contract language that raised scraping and AI training concerns

What happened: Meta has removed a provision from its commercial terms of use that had declared the company could "use any content from [a business account's] domains, web pages, or URLs to provide, develop, improve, research, train, and maintain the integrity of" Meta products and services, TechCrunch reported, citing The Verge. Legal experts interviewed by Nieman Lab had called the provision remarkable, saying it could give Meta permission to scrape the website of any publisher holding a Meta business account and use the content to train its AI models. A Meta spokesperson had insisted the updated terms did not reflect a change in how Meta crawls the web, and the company said on removing the language: "We've heard you and want to clarify: the crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves. It does not change how Meta uses web crawlers."

Who it matters to: Every business with a Meta business account — which is to say, most businesses that advertise or maintain a presence on Facebook or Instagram. The episode is a reminder that AI training rights are currently being negotiated in terms-of-use updates most account holders never read.

What you can do: If your business has a Meta account, this one ended well, but the habit is worth keeping: skim commercial-terms changes for the words "train" and "content," and if AI training on your site matters to you, state your own terms — a robots.txt AI policy and a site terms clause — rather than relying on a platform's wording staying friendly.

Learn more: TechCrunch on Meta removing the language

Anthropic studied its own workforce — and found fewer juniors, more meetings, more "managing Claudes"

What happened: Anthropic published a report on Friday on how its own workforce and work have changed in the past year, TechCrunch reported. The company says it has made fewer junior hires: junior-level employees in technical departments fell from 21% to 14% of the total since early 2025, and junior technical hires fell more sharply, from 27% to 9% of hires. Technical teams are a smaller share of the overall workforce, down from 61% to 56%, which Anthropic attributes both to business functions scaling and to coding tools letting non-technical employees build software themselves. In an April 2026 survey cited in the report, 27% of employees said Claude had made them more productive and 20% said their job had shifted "from actually doing work to managing Claudes" — one engineer compared it to babysitting. The report says human attention and judgment are becoming a productivity bottleneck as Claude produces a "deluge" of text that needs reading and organizing, and meeting time per employee is up 20% year over year (16% for technical contributors), which the company attributes to the importance of human trust and accountability. Anthropic says output measures — more code per person, more sales closed per rep, more contracts processed — are up, while acknowledging more code does not necessarily mean better code, and saying it has found no indications code quality is declining.

Who it matters to: Anyone making hiring or workflow decisions based on AI productivity claims. This is the most-cited AI lab describing its own experiment on itself: the gains are real by its measures, but so are the second-order effects — a thinner junior pipeline, a workforce whose scarcest resource is review attention, and more meetings, not fewer.

What you can do: Read your own team's version of these numbers before copying the model: if AI output is growing faster than your capacity to review it, your bottleneck has already moved. And if you are early-career, note where Anthropic's remaining human premium sits — judgment, coordination, and accountability — and build visibly toward those.

Learn more: TechCrunch on Anthropic's workforce report

Also worth knowing

• NVIDIA's RTX Spark laptops are on sale. NVIDIA's October 17 blog post says RTX Spark laptops — pairing a Blackwell RTX GPU with a Grace CPU, up to 1 petaflop of AI compute and up to 128GB of unified memory — are now available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,399, aimed at local AI workloads, content creation and gaming. The post is a launch announcement; there is no independent performance testing yet. Summary of NVIDIA's announcement
• Yesterday's rogue-agent checklist still stands. No new incidents were reported today, but nothing has changed about the fix: rotate credentials that ever sat in public code or configs, and tighten cloud roles to least privilege. See the October 17 briefing on this site for the full checklist.

One practical takeaway

Check your defaults — someone else set them this week. Replit's users did not lose Claude; they lost it as the default, which for most people amounts to the same thing. The same pattern runs through the whole briefing: a platform's terms quietly claimed training rights until someone read them, and an AI lab's scarcest resource turned out to be the humans reviewing the output. Spend ten minutes today listing the defaults that govern your work — your coding tool's model, your platforms' content terms, who reviews what your AI produces — and set them deliberately, once, instead of inheriting them by accident.

---

Sources: TechCrunch AI, The Verge via TechCrunch, Nieman Lab via TechCrunch, 9to5Google, Crowdfund Insider, NVIDIA's October 17 announcement as summarized by CryptoNwz. Summaries are original; follow the links for full reporting.