AI News Today — October 30, 2026: Mistral Open-Sources a Trillion-Parameter Model, Amazon Commits $220 Billion to AI, and Reddit Walls Off the Open Web

The big picture

October 29 was a day about who controls AI — and who pays for it. Mistral handed the keys to a frontier model to anyone who wants them, releasing the full weights of its trillion-parameter flagship for download. Amazon went the other direction on scale: roughly $220 billion of spending this year, mostly to build AI infrastructure it insists will make AWS a $1 trillion business. And Reddit showed what AI is doing to everyone in between — as AI search answers questions without sending clicks, one of the open web's biggest sites is retreating behind a login wall to protect the data AI companies want.

What happened today

Mistral open-sourced a trillion-parameter model built for cybersecurity

What happened: Mistral released the weights for Mistral Large 4 — its 1-trillion-parameter, natively multimodal flagship, nicknamed "Le Chonk" — on Hugging Face and through its API. Anyone, including enterprises and governments, can now download and self-host it. Mistral positions the model for cybersecurity work: detecting, analyzing and responding to threats across code and IT infrastructure. The company says it performs comparably to China's GLM-5.3, previously rated among the best open-weight models, while using 47% fewer tokens per task, and that it tops the Artificial Analysis Coding Agent Index among open models. Those benchmark figures are Mistral's own; independent validation is still pending.

Who it matters to: Security teams, governments and any organization that cannot send sensitive data to someone else's cloud. Mistral's pitch, in CEO Arthur Mensch's framing, is AI as "a tool for empowerment, not dominance" — capability that, in VP of Science Pierre Stock's words, "can't be cut off by definition because the model is open-weighted."

What you can do: If your work involves sensitive code or infrastructure, a downloadable frontier-class model is now a real option to evaluate. Start with a scoped pilot — code scanning or incident triage — and benchmark it on your own tasks rather than trusting vendor numbers.

Learn more: TechCrunch's coverage of the Mistral Large 4 release

Amazon is spending $220 billion this year — and says AWS becomes a $1 trillion business

What happened: On Amazon's quarterly earnings call Thursday, CEO Andy Jassy confirmed the company still expects 2026 cash capital spending of approximately $220 billion — the majority for AWS — and expects to spend "meaningfully more" in 2027. AWS revenue grew 39% year over year, its sixth straight quarter of steady growth, to a $184 billion annualized run rate. Jassy forecast growth north of 40% in both 2027 and 2028 and called AWS a future trillion-dollar-revenue business "in the next five to 10 years." Investors were not sold: Amazon shares fell about 6% in after-hours trading as analysts pressed him on returns.

Who it matters to: Everyone who buys cloud or AI services. This is the supply side of the AI economy being built in public — the capacity behind the models and agents the rest of this briefing runs on. It also sets the stakes: spending at this level only pays back if enterprise AI demand keeps compounding.

What you can do: If your business depends on AWS or any hyperscaler, plan for continued price and capacity competition in your favor — but read the investor skepticism as a reminder to keep your own AI spending tied to measurable returns, not momentum.

Learn more: TechCrunch on Jassy's $1 trillion AWS call

Reddit is retreating behind a login as AI search eats its traffic

What happened: On its third-quarter earnings call, Reddit announced a shift to a primarily "logged-in experience." The reason is in its own numbers: search referrals dropped by more than 1 million per week on average in October, U.S. daily active users fell from 53.2 million to 50.7 million, and the company warned daily users could decline by a low-double-digit percentage next quarter if the trend holds. The wall also protects Reddit's data for AI licensing deals — non-advertising revenue, which includes licensing, grew 35% to $48 million. Overall revenue still rose 53% to $895 million, beating expectations. Tests start in select non-U.S. markets this quarter; there is no global timeline yet.

Who it matters to: Publishers, marketers and anyone whose traffic comes from search. Reddit is the canary: when AI answers replace clicks, even a 130-million-daily-user platform starts closing the door. Logged-in users are also worth more to advertisers, so the retreat pays for itself twice.

What you can do: Audit how much of your traffic depends on being found by AI summaries and search referrals. Audiences you own — email lists, communities, logged-in products — are becoming the only reliable distribution.

Learn more: TechCrunch on Reddit's logged-in shift

An OpenAI partner just bought ADT in a $15.3 billion AI bet

What happened: Thrive Holdings — the holding-company spinout of Thrive Capital, which counts OpenAI as both partner and investor — is acquiring home-security company ADT for $10 per share, about $7.6 billion in equity and $15.3 billion in total enterprise value. ADT becomes the foundation of Thrive's smart-home security business. As a partner, OpenAI supplies models, technology and forward-deployed engineers to Thrive's 75-plus portfolio companies. Neither company has yet announced specific AI security products.

Who it matters to: The smart-home market, and anyone watching how AI labs expand. This is the "AI revitalizes an old industry" model at its largest scale yet — a 150-year-old security company as the test bed, with the lab embedded inside.

What you can do: If you use ADT or a rival, expect an AI-feature race — Apple and Amazon both pushed AI home-security products this season. Hold off on long contracts until the new owners show what they're actually shipping.

Learn more: TechCrunch on the Thrive–ADT deal

OpenAI rewrote its model rulebook around five values

What happened: OpenAI published an updated Model Spec — its public document describing how it wants its models to behave — reorganizing everything around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness comes first: models should understand what users, developers and OpenAI intend and value, not just follow instructions literally. The update adds guidance on context, memory, personalization and oversight, and retires most of the Spec's detailed style guidance. OpenAI acknowledged its current models "do not yet fully reflect" the new Spec and said it is testing training methods to close that gap.

Who it matters to: Anyone building on or relying on OpenAI's models. The Spec is the closest thing to a public contract for how these systems should act when your instructions run out — which, with agents, is often.

What you can do: Read the five values as a checklist for your own AI use: if a tool can't explain how it handles conflicting instructions, uncertainty and side effects, that's a gap to manage, not a detail.

Learn more: OpenAI's Model Spec update in its model release notes

Muse AI: no new announcement — Singapore and the Nvidia partnership stand

What happened: Our daily priority check found no new Muse announcement from Meta on October 29–30. The latest developments remain the ones already covered here: Muse's launch in Singapore — its fourth market, with eight local connectors rolling out progressively — and the expanded Meta–Nvidia partnership to optimize Muse agent inference and training, which Nvidia announced on October 26. Meta's new consumer AI release this week was separate from Muse: Style Sketch, a fashion app (see Also worth knowing).

Who it matters to: Anyone tracking the personal-agent race. Muse's pattern is steady geographic expansion with local partners, while the enterprise tier remains announced but unshipped and unpriced — the gap to watch.

What you can do: If you're outside the U.S., Canada, Japan and Singapore, there is still nothing to sign up for. Inside those markets, expect connectors to switch on gradually rather than all at once.

Learn more: PYMNTS on the Meta–Nvidia Muse partnership

Gemini Spark: no new announcement — Task mode and the November 17 deadline stand

What happened: Our daily priority check found no new Gemini Spark announcement. Spark still lives inside the Gemini app as Task mode — the Chat | Task switch above the prompt box, with a Scheduled page for active, paused and completed tasks — for Google AI Pro and Ultra subscribers. The next dated change is November 17, when Google begins automatically migrating personal-account Gems into the new Skills format (work and school accounts follow in 2027).

Who it matters to: Anyone who built custom Gems or relies on Spark tasks. The rebrand is done; the migration of your saved setups is what's actually coming.

What you can do: Before November 17, inventory your Gems and note which ones matter. After the automatic conversion, open each converted Skill once and confirm it still does what the Gem did.

Learn more: 9to5Google on the Gemini Task redesign

Also worth knowing

• Wabi raised $55 million at a $1.25 billion valuation. The personal agent from Replika founder Eugenia Kuyda — it builds mini apps from prompts for things like meal planning and date nights — repositioned itself as an AI messenger last month; IVP led the round, with a16z and the Amazon Alexa Fund among the participants. TechCrunch
• Granola made dictation free and unlimited. The meeting-notes app now lets Mac users press Fn and speak into any app, with cleanup and formatting — a direct move on Wispr, which added meeting notes of its own earlier this year. iOS and Windows versions are coming. TechCrunch
• GitHub's Copilot Fusion is generally available. Fusion picks and combines models per task — drafting with one, checking with another — and GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code on Opus 5.5 Max at roughly half the cost per task. GitHub also reports agents now start more than 40% of new pull requests on the platform. Neowin
• Claude for Legal got a gallery of ready-made tools. Anthropic added pre-built artifacts for contract review, cite checks, regulatory horizon scans and more, alongside 60-plus legal connectors — each output links back to its source, and the workflow keeps a lawyer at both ends. Unite.AI
• Meta quietly launched a fashion app. Style Sketch, live in Indonesia, Singapore and Thailand, builds a virtual closet from your Instagram photos and suggests outfits with try-ons matched to your body shape — one of a wave of small AI apps Meta is shipping this year. TechCrunch
• Alexa learned 100-plus new shopping tricks. Ahead of the holidays, Alexa can manage lists and subscriptions, edit delivery instructions and answer Prime billing questions; Amazon says use of its agentic shopping features has grown eightfold this year. TechCrunch
• Deepgram bought the Pipecat maker. Deepgram acquired Daily, the company behind the open-source Pipecat voice-agent framework, tightening its grip on voice AI infrastructure. TechCrunch
• Two more signals from the feed: ex-Palantir executives launched Dream with $17 million from a16z to build AI world models for construction, and Rivian and Volkswagen expanded their joint venture to include an in-car AI assistant.

One practical takeaway

Frontier AI no longer has to live in someone else's cloud. With Mistral Large 4's weights public, "should we self-host?" is now a legitimate question for any team handling sensitive data — and with routers like Copilot Fusion halving agent costs, "which single model should we buy?" is becoming the wrong question entirely. This week, evaluate one workload both ways: self-hosted open weights versus a routed multi-model service. The right answer is now genuinely workload-dependent.

Sources

TechCrunch AI (Mistral Large 4, AWS earnings, Reddit, Thrive–ADT, Style Sketch, Wabi, Granola, Alexa, Deepgram–Daily), OpenAI model release notes and Reuters (Model Spec), PYMNTS (Meta–Nvidia Muse partnership), Neowin and GitHub (Copilot Fusion), Unite.AI (Claude for Legal), 9to5Google (Gemini Task redesign). Summaries are original; follow the links for full reporting.