AI News Today — November 4, 2026: Anthropic Can Prove Which Model Answered You, Muse AI Hits 10 Million Users, and Compute Eats 58% of Anthropic's Revenue

The big picture

Today's AI news is about proof and price. Anthropic open-sourced a way to cryptographically prove which model actually produced an answer — aimed squarely at providers quietly swapping in cheaper models. Meta put hard numbers on its personal agent for the first time: Muse passed 10 million setups in about six weeks, and shopping is already its leading use. And Anthropic's own updated IPO filing shows the cost underneath all of it — compute consumed 58% of its revenue last quarter, up from 48%, because demand forced it into expensive short-term deals. Trust, traction, and the bill for both: that is the whole day in one line.

What happened today

Anthropic's FlashProve can prove which model produced an output

What happened: Anthropic published FlashProve, an open-source system for what researchers call "provable inference" — a cryptographic proof, essentially a mathematical guarantee, that a given output came from a specific model, TechCrunch reports. Labs would publish a "fingerprint" of each model they release, and outputs could later be matched to it and verified cryptographically, without revealing the model's weights. Anthropic released a detailed technical specification and working code under an open-source license on GitHub. The limits are stated plainly in the announcement itself: the research is ongoing, this is "intermediate progress, not a completed research agenda," and there are no plans to deploy it at Anthropic or any other lab yet.

Who it matters to: Anyone paying for a frontier model through an API — and anyone reselling one. The failure mode FlashProve targets is real and hard to catch today: a provider or router quietly serving a cheaper model behind the endpoint you paid for, or swapping in a stronger model to inflate benchmark results. A verifiable fingerprint turns "we used the model you bought" from a contract promise into something checkable.

What you can do: Nothing to install yet — this is a research release, not a product. The practical takeaway for now is contractual: if your business depends on a specific model, ask your provider what guarantee you actually have that you are getting it, and note that the honest answer today is usually "none you can verify." Watch whether any lab commits to publishing model fingerprints; that is the step that would make FlashProve usable.

Learn more: TechCrunch on Anthropic's FlashProve model-identity system

Muse AI: more than 10 million users in six weeks, and shopping is the leading use

What happened: Meta's third-quarter earnings call on November 3 gave Muse its first official traction numbers, Zacks reports. CEO Mark Zuckerberg said more than 10 million users in the United States and Canada established Muse personal agents within approximately six weeks of launch, with roughly 25% using the service daily. Shopping has emerged as Muse's leading use case, with more than one-quarter of users engaging in commerce and average purchases exceeding $200, including travel transactions. Zuckerberg outlined monetization through transaction fees, premium subscriptions and enterprise services, said he expects compute efficiency to improve more than tenfold over the next year, and pointed to plans to serve over 1 billion users. CFO Susan Li acknowledged significant Muse serving costs and described the efficiency work — smaller models, better GPU utilization, off-peak capacity — meant to bring them down. The numbers sit inside a strong quarter for Meta overall: Q3 revenue of $64.6 billion, up 26% year over year, with full-year 2026 capital expenditure guidance narrowed to $135 billion–$145 billion. Separately — and unconfirmed by Meta — Marketplace Pulse founder Juozas Kaziukėnas reported on X on November 3 that Muse is getting native shopping carts it renders itself, with Instacart as the first connector; treat that as a reported sighting, not an announcement.

Who it matters to: Anyone selling online, and anyone watching whether personal agents become a real shopping channel. Ten million setups in six weeks with a quarter of users buying things — at an average above $200 — is the first hard evidence that agent-led commerce is more than a demo. It also reframes yesterday's story: the banks and retailers blocking Muse are negotiating with a channel that is already moving money, and Meta is now saying monetization will come sooner than its usual playbook.

What you can do: If you sell products or services, check how your business appears to an agent, not just to a search engine: can your site be read, quoted and transacted with by an automated buyer, and what approval steps stand between an agent's recommendation and a completed purchase? If you use Muse, the serving-cost comments are the signal to watch — free, generous agent usage is exactly what gets repriced once a lab starts optimizing it.

Learn more: Zacks on Meta's Q3 call and the Muse growth numbers

Anthropic's compute bill ate 58% of its revenue — and the share is rising

What happened: Anthropic's updated IPO filing, extending its financial disclosures through September 30, shows a company growing fast and paying more to do it, TechCrunch reports. Revenue reached $17 billion in the quarter, up from $11.5 billion the quarter before, but the operating loss grew alongside it. Compute took up 58% of revenue, up from 48% the previous quarter, and total compute spending hit $10 billion for the quarter — including, as the filing notes, $5.0 billion of compute under short-term commitments, which "currently costs significantly more than comparable compute procured under our multi-year commitments." In plain terms: demand grew faster than Anthropic's planned capacity, so it scrambled for expensive stopgap compute. The filing says it expects compute and infrastructure to "remain our largest operating expense for the foreseeable future."

Who it matters to: Anyone building a budget on top of frontier-model APIs. The lab serving you is buying capacity at premium spot rates to keep up with demand — that pressure flows downhill eventually, as pricing changes, tighter limits, or both. It also explains yesterday's Grok Bot story from the other side: Musk cut limits citing unsustainable serving costs; Anthropic's filing shows what those costs look like in a securities document.

What you can do: Treat today's API prices and limits as provisional, because the labs' own filings do. If a workflow is core to your business, know what it costs you per task at current prices, and what you would switch to — a smaller model, a rival lab, a cached pattern — if the price moved. The users caught out by repricing are the ones who never measured their usage while it was cheap.

Learn more: TechCrunch on Anthropic's compute economics in the updated filing

Fal doubles its valuation to $16 billion as inference demand surges

What happened: Fal, the San Francisco startup that hosts and runs AI video, image and audio models, raised $1.3 billion in a Series F led by Bessemer Venture Partners, with participation from Nvidia, Sound Ventures and T. Rowe Price, valuing the company at $16 billion — double its $8 billion valuation in April, TechCrunch reports. Fal announced its Series E and F rounds at the same time, a combined $1.65 billion raised across both. The round comes amid surging demand for inference — running already-trained models to generate outputs — particularly from customers relying on open-source models. TechCrunch notes rival inference providers including Baseten, Fireworks AI and Modal are also growing and raising rapidly, while gross margins across the category remain thin because of the cost of buying or leasing compute.

Who it matters to: Developers and media teams who generate images, video or audio through hosted models. The money is concentrating in the layer between you and the GPU: whoever runs the model cheapest and fastest increasingly sets your price and your latency. Fal's $800 million in annualized revenue earlier this year, per Bloomberg's reporting cited by TechCrunch, says this layer is already a real business, not a wrapper.

What you can do: If you generate media through an inference platform, you are probably not locked in — the same open models run on several hosts. Benchmark your two or three most-used models on a second provider for price and speed before volume makes switching painful; thin margins in this category mean prices will keep moving, in both directions.

Learn more: TechCrunch on Fal's $1.3B raise and $16B valuation

Apple Intelligence begins developer testing in China today, with Alibaba's Qwen inside

What happened: Apple's long-delayed AI rollout in China reached its test phase today, November 4: the first Apple Intelligence features for mainland China opened to developer testing in the latest iOS beta, ahead of a public launch with iOS 27.3 in December, according to Apple's China announcement of October 31 as reported by the Bangkok Post and cross-checked in ByteNib's summary of the launch details. The first wave covers Live Translation, Writing Tools, Visual Intelligence, Clean Up in Photos, Genmoji and notification summaries. The structural news is the partnership: Alibaba's Qwen model is integrated into Siri, Writing Tools and Visual Intelligence for text and image understanding — users opt in, and Alibaba says requests and responses will not be used to train its models, per reports of the joint announcement. The rebuilt Siri AI shown at WWDC 2026 is not part of this phase. Apple only secured Chinese regulatory approval in July, after roughly two years of delay since Apple Intelligence first launched in the U.S. in autumn 2024.

Who it matters to: Anyone selling to, building for, or competing in the Chinese market — and anyone tracking how frontier AI actually crosses regulatory borders. The pattern is now explicit: a U.S. platform plus a domestic model partner, with data-handling commitments as part of the deal. It is the template every other Western lab will be measured against in regulated markets.

What you can do: Nothing to download unless you are a developer in China on the beta. The transferable lesson is for product planning: if your AI product depends on a single model provider, a market can force a second one on you. Architect for model substitution now — ironically, the same problem Anthropic's FlashProve story above is about proving.

Learn more: ByteNib on Apple Intelligence reaching China with Alibaba's Qwen

Gemini Spark: no new Spark announcement — the Gems freeze is now 13 days out

What happened: Our daily priority check found no new Gemini Spark announcement. Spark remains folded into the Gemini app as Task mode — the Chat | Task switch, event-triggered multistep tasks across connected apps and the web, up to 15 tasks at once, for Google AI Pro and Ultra subscribers with a personal account, Keep Activity on and Personal Intelligence opted in, everywhere Gemini Apps are supported except the EEA, the UK, Switzerland and Nigeria. The deadline underneath it moved one day closer and is now the story: personal-account Gems become read-only on November 17 — 13 days from today — and are permanently deleted on December 17. Google will auto-convert Gems to Skills, and the Skills format is the portable SKILL.md file Gemini now shares with Claude, but auto-conversion drops attachments over the 100 MB per-Skill cap, largest first, and some formats — legacy Office documents and archives among them — do not carry over at all. One adjacent Gemini change did land yesterday: Google will turn its AI tools off by default for K-12 students next year unless school administrators explicitly opt in, covering Gemini, NotebookLM and Gemini in Classroom — but schools that restrict AI altogether will also lose Google Search, because Google says Search and its AI features are now inseparable, TechCrunch reports.

Who it matters to: Anyone with saved Gems on a personal Google account — in under two weeks the setups feeding your Spark tasks stop being editable, and in six weeks the originals are gone. The schools change matters to a different group the same way: "opt out of AI" increasingly means "opt out of the platform," because the two are being fused at the product level.

What you can do: This is the week to do the backup, not the week of the freeze: open each Gem you depend on, copy its instructions into a plain SKILL.md file you keep yourself, and download any attachments over the 100 MB cap or in unsupported formats. That single file then works in Gemini's new Skills — and, because it is the shared format, in Claude as well. If you administer a school account, start the opt-in conversation now rather than at rollout.

Learn more: TechCrunch on Google's K-12 AI opt-in change and the Search trade-off

Also worth knowing

• Netflix launched an AI Discover tab on TV in the U.S., starting with mood-based recommendations — themes like "something light," "scare me" and "quick watches," matched to clips, available time and taste, TechCrunch reports. Netflix calls it a test kitchen; user feedback decides what reaches the main experience.
• Palantir topped $2 billion in quarterly revenue for the first time — revenue up 96% year over year to $2.31 billion, with U.S. commercial revenue ($984 million) passing U.S. government revenue for the first time. The caution inside the beat: U.S. enterprise contract value fell to $1.46 billion from $2.13 billion the prior quarter, and customers are signing shorter trials, TechCrunch reports.
• Data centers are turning to old jet engines for power. Stuck in years-long lines for grid connections — about seven years for 100 megawatts in Northern Virginia, per the IEA — some operators are running converted aircraft engines rebuilt to burn natural gas; xAI has eight installed for its Colossus 2 site and OpenAI's Abilene Stargate campus uses gas turbines, TechCrunch reports via the New York Times.
• Apple will restrict dating apps and AI companions to users 18 and up. Apps offering dating or romantic features — including AI companions with sexual or romantic content — must carry an 18+ rating, with new content descriptors required starting November 16; apps must also disclose real-world location sharing, TechCrunch reports.
• Ballistic Ventures is raising $490 million for cybersecurity. Ted Schlein's firm filed for a $360 million third flagship fund plus $130 million for Fund B, as AI-driven attacks expand the market — AI-enabled companies accounted for about 72% of U.S. cybersecurity venture deals through May, per a JPMorgan/PitchBook report cited by TechCrunch.
• Enterprise AI spending hit $176.8 billion by September 2026, up 5.4x from $33 billion in 2025, and surveyed companies spent 2.5 times what they had budgeted for the year, according to Menlo Ventures' 2026 State of Enterprise AI report released November 3 and based on nearly 400 U.S. enterprise decision-makers, as summarized in The Agile Brand Guide's November 4 roundup. Read alongside Anthropic's compute filing above, the overspend and the squeeze are the same story told from both ends.

One practical takeaway

Verify what you are actually getting — and price what it actually costs. Today's three lead stories are one habit applied three ways: Anthropic built a proof because "the model you paid for" is currently unverifiable, Meta's agent numbers show a shopping channel forming before most sellers have checked how they appear to it, and Anthropic's filing shows the labs themselves re-pricing under compute pressure. This week, run the two-minute version: confirm which model your key tools are really serving, check your usage and cost per task while prices are still low, and back up the setups — Gems included, 13 days before they freeze — that you cannot afford to lose.

Sources

TechCrunch (Anthropic's FlashProve provable-inference release, November 3; Anthropic's compute economics in the updated S-1/A filing, November 3; Fal's Series E/F and $16B valuation, November 3; Netflix's AI Discover tab, November 3; Google's K-12 AI opt-in and Search trade-off, November 3; Palantir's Q3 results and enterprise contract signal, November 3; data centers using converted jet engines, via the New York Times, November 3; Apple's 18+ dating and AI companion policy, November 3; Ballistic Ventures' $490M fund filings, November 3); Zacks (Meta Q3 earnings call — Muse's 10M+ setups, 25% daily use, shopping as leading use, $64.6B revenue and $135B–$145B capex guidance, November 3); ByteNib and the Bangkok Post (Apple Intelligence developer testing in China starting November 4, iOS 27.3 in December, Alibaba Qwen integration, based on Apple's October 31 China announcement); FourWeekMBA summarizing Marketplace Pulse on X (Muse native shopping carts with Instacart first — reported, not confirmed by Meta, November 3); The Agile Brand Guide's November 4 roundup (Menlo Ventures' 2026 State of Enterprise AI report, November 3); Everpod's and Universal Digital Assistant's Gemini Spark guides, checked against Google's Gemini Apps Help pages (Task mode requirements, availability and the Gems-to-Skills timeline). Summaries are original; follow the links for full reporting.