AI News Today — November 5, 2026: Uber Says AI Agents Booked Four Rides in Its Busiest Week, Tencent's Hy4 Learns to See and Click, and Moonshot Drops 2.8-Trillion-Parameter Weights

The big picture

Today's AI news is where the agent hype meets a receipt. Uber gave the first hard transaction count for agent commerce — four rides and about 500 food orders booked by AI agents in its busiest week of the year, out of roughly 320 million trips and orders — while Apple is quietly rationing how many AI tokens its own engineers may burn each day. The open-model race went the other way: Tencent gave its flagship eyes and a mouse hand, and Moonshot AI actually shipped 1.56 terabytes of model weights on the day it promised. Setups are easy to announce; transactions, budgets and downloadable files are harder to fake. Today, the hard numbers won.

What happened today

Uber: AI agents booked four rides in Halloween week

What happened: On its third-quarter earnings call on November 4, Uber told analysts that almost nobody is using AI agents to book through its platform yet, TechCrunch reports. During the week of Halloween — its busiest week in October — only four rides were ordered through an AI agent, and CFO Balaji Krishnamurthy put agent-placed Uber Eats orders at about 500, against roughly 320 million trips and orders across Uber and Uber Eats that week. CEO Dara Khosrowshahi called it "very, very early and very limited traction," while noting that booking a ride in Uber's own app takes about 50 seconds — a bar an agent has to beat to be worth using. He also said every major AI system company is talking to Uber, and that he expects agents to become a net positive channel over time. The numbers rhyme with last week's DoorDash disclosure of only "several hundred" agent orders against 993 million quarterly orders.

Who it matters to: Anyone building, selling through, or investing in personal agents. This is the first transaction-level reality check from a platform with no reason to lowball the trend — and it sits in direct tension with the setup numbers labs announce, including Meta's 10 million Muse setups with shopping as the leading use. Both can be true: millions of people have agents, and almost none of them are letting agents buy things yet.

What you can do: If you sell online, treat agent commerce as a channel to prepare for, not a channel that exists at scale today: make sure an agent can read your prices, availability and checkout rules, but keep optimizing the 50-second human path that still produces essentially all of the orders. If you use an agent, notice where it actually saves you time — reordering and research — versus where the app was already faster.

Learn more: TechCrunch on Uber's agent booking numbers

Tencent's Hy4 can see images and operate a computer — but you can't download it yet

What happened: Tencent's Hunyuan team released the full version of Hy4 on November 4, the first model in the series with vision and the first that can operate a computer through both a command line and a graphical interface, trained on professional software including Blender and CAD tools. It is live now in Tencent's own apps — Yuanbao, WorkBuddy, CodeBuddy, ima and the WeCom assistant Dayuan — and through Tencent Cloud's TokenHub and third-party API platforms including OpenRouter, which lists a context window of about 1 million tokens with image input. The size figures come from the August preview's model card: a mixture-of-experts design with 770 billion total parameters and 49 billion active. Pricing holds at the preview's level — about $0.90 per million input tokens and $2.70 per million output at list, with TokenHub calls 50% off through December 31 and a free usage window in WorkBuddy and CodeBuddy through November 15. Two honest caveats come from Tencent itself and from careful launch coverage: the full release's weights are not out yet — Tencent says it "plans to" release them, with a system card, in the coming days, and only the August preview weights currently sit on Hugging Face under Apache 2.0 — and Tencent concedes a "significant gap" to the top closed models on its hardest internal tests, with its own Terminal Bench 4.0 score of 44.4 against 66.4 for Claude Opus 5.5.

Who it matters to: Developers and teams evaluating open or low-cost frontier alternatives, especially for computer-use and document-heavy work. Hy4's API price is a fraction of the U.S. frontier labs', and vision plus computer use in one release is exactly the combination agent builders have been waiting on from the open side — once the weights actually land.

What you can do: You can try Hy4 today through OpenRouter or Tencent's apps without committing to anything. Do not plan a self-hosted deployment around it yet: there is no download, no confirmed license for the full release, and no independent benchmark of the final model. Re-check when the weights and system card appear — that is the moment Hy4 becomes an open model rather than a hosted one.

Learn more: Launch analysis of Tencent's Hy4 release, its pricing and its missing weights

Moonshot published Kimi K3.1 Pro's weights — 2.8 trillion parameters, on schedule

What happened: Moonshot AI posted the full weights of Kimi K3.1 Pro to Hugging Face on November 4, the date it promised at the model's October 23 launch. The model card lists a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active per token, a context length of 1,048,576 tokens, and native text-and-image input; the ungated repository holds 96 weight files totaling about 1.56 TB. The license is Moonshot's own, not a standard open-source license: anyone may use, modify and sell with it, but a model-as-a-service business whose revenue with affiliates exceeds $20 million over any 12 months must sign a separate agreement with Moonshot first, and products above 100 million monthly users or $20 million in monthly revenue must display the model's name. In Moonshot's own benchmark table, K3.1 Pro scores ahead of Claude Opus 5.5 on 16 of 37 comparable non-internal rows — by small margins — while trailing by more than 20 points on three. Nothing changes for hosted users: the API stays at $2 per million input tokens and $10 per million output.

Who it matters to: The small set of teams who can actually run a 1.56 TB model — and everyone who rents from them. Self-hosting needs roughly eight data-center GPUs in the simplest published vLLM setup, so for most people K3.1 Pro's weights matter indirectly: more hosts can offer it, which is what pressures the $2/$10 price. The revenue-gated license is also the new normal to read carefully — "open weights" no longer automatically means "no strings."

What you can do: If you use Kimi through an API, nothing changes today; watch for new hosts and price movement now that the weights are public. If you are evaluating self-hosting, read the license's $20 million clause before you build on it, and budget for the hardware honestly — this is a data-center model, not a laptop download.

Learn more: Coverage of the Kimi K3.1 Pro weights release and its license terms

Muse AI: no new Meta launch — Conductor joins Meta, and a Muse commerce ecosystem forms

What happened: Our daily priority check found no new Muse product announcement from Meta today; the 10 million-plus setups, roughly 25% daily use and shopping-led usage from the Q3 call stand as the latest official numbers. Two developments landed around it on November 4. First, Meta acquired Conductor, the AI coding startup whose platform spins up multiple agents — running models from labs including Anthropic and OpenAI — into parallel workspaces, TechCrunch reports. Founder Charlie Holtz said he will work with Meta Superintelligence Labs leadership, teased "something new" by early next year, and confirmed Conductor's cloud service shuts down on November 20, when deletion of cloud user data begins; Meta confirmed the acquisition without further detail. Second, ReFiBuy, an agentic-commerce optimization platform, announced support for Meta Muse, pitching brands a way to enrich the products Muse surfaces — its research arm previously found Muse draws on Meta's product catalog, Shopify stores and web product pages, with Black Friday on November 27 set as Muse's first big retail test.

Who it matters to: Developers on Conductor, who have a hard November 20 deadline, and brands watching agent commerce, who just got the clearest sign yet that an optimization industry is forming around how products appear inside Muse — even while Uber's numbers above show how little agent buying has actually happened so far. Meta buying a multi-agent coding tool also fills a real gap next to its own Muse Code, launched in August.

What you can do: If your team uses Conductor's cloud service, export your work before November 20 — that is when it shuts down and data deletion begins. If you sell products, ReFiBuy's pitch is worth reading as a map of what agent-readiness now means — clean catalog data and product pages an agent can quote — whether or not you buy the tool.

Learn more: TechCrunch on Meta's acquisition of Conductor

Apple is rationing its own engineers' AI tokens

What happened: Apple has cut the daily amount its engineering teams can spend on AI, Bloomberg reported via TechCrunch. In October, Apple lowered the daily AI spending cap for engineering teams from $300 to $200, more cuts are reportedly in the works with some allowances falling to $100 a day, and outside engineering the company is weighing a daily limit as low as $25. TechCrunch frames it as the end of this year's "tokenmaxxing" era — companies pushed staff to use AI as much as possible, the token bills came due, and rationing followed. Apple has not commented publicly; treat the specific figures as Bloomberg's reporting, not an Apple announcement.

Who it matters to: Anyone who manages a team using AI tools — which is to say, most managers now. Apple is the most cost-disciplined company in tech, and it is treating internal AI the way it treats every other input: a budgeted resource with a per-person cap, not an all-you-can-eat perk. Coming the same week Anthropic's filing showed compute eating 58% of revenue, it is the demand-side version of the same story.

What you can do: Copy the mechanism, not just the headline: set an explicit per-person or per-team AI budget, look at what it buys monthly, and cut the usage that produces nothing measurable before you cut the usage that ships work. If your provider can't show you per-user spend, that is the first thing to fix.

Learn more: TechCrunch on Apple's reported internal AI budget cuts

Gemini Spark: no new Spark announcement — and Google's school Search plan draws a backlash

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 countdown underneath it is now the story: personal-account Gems become read-only on November 17 — 12 days from today — and are permanently deleted on December 17, with Google auto-converting Gems to Skills in a portable SKILL.md format. The adjacent Gemini story did move: Google's plan to turn its AI tools off by default for K-12 students next year — with schools that restrict AI altogether also losing Google Search — drew swift political backlash on November 4, TechCrunch reports. Senator Bernie Sanders called the trade-off "extortion," former FTC chair Lina Khan called it monopoly coercion, and Google clarified that Gemini, Gemini Notebook and Search will be off by default and can be enabled individually, saying the changes give schools "more choice, not less." DuckDuckGo is meanwhile pitching its own search engine to schools as the no-bundle alternative.

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 fight matters more broadly because it is the consumer version of the same question: when AI and the underlying platform are fused, "opt out of AI" stops being a real option, for a school or for anyone else.

What you can do: Do the Gems backup this week, 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 per-Skill cap or in formats that don't convert. That file then works in Gemini's new Skills — and, because it is the shared format, in Claude as well.

Learn more: TechCrunch on the backlash to Google's school Search and AI plan

Also worth knowing

• OpenAI is already shaping its 2028 election agenda. Global affairs chief Chris Lehane published an essay framing 2028 as a defining election for AI and urging a New Deal-style program across worker protections, education, enforceable safety standards and data center development; critics told TechCrunch the plan presumes mass adoption while offering few binding ways to share AI's gains. It lands the same week TechCrunch reports AI PACs spent millions on the midterms and a bigger post-midterm clash over data centers could be coming.
• OpenAI had a four-part November 4, per roundups summarizing its announcements: Images 2.6 with faster generation, GIF creation and finer editing controls; an agreement to acquire Pierre Computer Company to expand Codex's developer tools; a Linux-kernel research campaign reporting 948 unique local privilege-escalation vulnerabilities dynamically reproduced from 2,308 candidates — OpenAI's own figures, reported via its research write-up; and Daybreak Red, its GPT-5.6 Cyber access program for approved defensive-security teams, made available through Amazon Bedrock in AWS GovCloud (US-East).
• Claude helped resolve a protein structure. Anthropic reported on November 4 that Claude assisted sequence searches, purification planning and microscopy analysis behind a cryo-EM structure of a bacteriophage-derived vault particle built from 96 protein copies — a research result with substantial human direction, not an autonomous discovery.
• Snowflake put its agents on schedules and phones. At Expedition 2026, Snowflake announced CoCo and CoWork Bots — always-on versions of its coding agent for data teams (9,100+ accounts) and personal work agent for knowledge workers (5,800+ accounts) — plus Automations for recurring work, a unified mobile app and skills sharing; the Bots are marked private preview soon, while Automations are generally available, per Snowflake's announcement.
• Salesforce agreed to buy voice-AI platform Cresta. The definitive agreement, announced November 2, brings Cresta's end-to-end voice agents, live coaching for human representatives and conversation analytics into Service Cloud and Agentforce; Salesforce did not disclose a price — The Information reported $2.7 billion — and the deal is expected to close in the first quarter of fiscal 2028 (February–April 2027), with Cresta continuing to support rival contact-center platforms after close.
• Nothing's first AI-native device ships in December. CEO Carl Pei says it is a screenless productivity tool that "couldn't have been an app," pointedly separate from the industry's hunt for an AI pendant, pin or glasses; no date or product details yet, TechCrunch reports.
• Agility Robotics is going public via SPAC. The deal raises $465 million and makes Agility the first pure humanoid-robot company on the stock market; its Digit robots have moved more than 100,000 totes at one GXO facility and it cites $300 million in planned Digit 5 contracts — against $174,000 of revenue and a $362 million loss in the first half of 2026, TechCrunch reports.
• Siri Recaps are useful — and a little unsettling. TechCrunch's hands-on with the Audio Intelligence features in the iOS 27.3 developer beta found the Apple Watch's automatic conversation summaries genuinely good at capturing prices, deadlines and repair details, processed on-device where possible — and hard to forget about at the family dinner table, TechCrunch reports.
• CrowdStrike opened early access to AI red teaming on its Falcon platform at Fal.Con Europe on November 4 — an agentic emulation workflow that chains exposed weaknesses, recommends fixes and retests; broad availability and independent results are not yet established, so treat it as an announcement, not a proven capability.
• Google bundled more of the cloud into agent workflows. Its November 4 Agentic Data Cloud updates made cross-cloud interconnect and caching for AWS and Azure generally available and described direct Iceberg access in AlloyDB, letting a single query span operational and lakehouse data — plumbing, but the kind agents starve without.

One practical takeaway

Count transactions and tokens, not setups and seats. Today gave you the three numbers that actually describe where AI stands: four agent-booked rides in Uber's busiest week, a $300-to-$200 daily token cap inside Apple, and 1.56 TB of weights anyone can now download. Run the same audit on your own stack this week: how many tasks did your agents complete end-to-end, what did your AI usage cost per person, and which of your setups — Gems included, 12 days before they freeze — would survive your provider changing the terms tomorrow?

Sources

TechCrunch (Uber's Q3 call agent numbers — four rides, ~500 Eats orders, ~320M trips and orders in Halloween week, November 4; Meta's acquisition of Conductor and the November 20 cloud shutdown, November 4; Apple's reported internal AI budget cuts, via Bloomberg, November 4; backlash to Google's K-12 Search and AI plan, with Google's clarification, November 4; OpenAI's 2028 election essay and critics' responses, November 4; Nothing's December AI-native device, November 4; Agility Robotics' SPAC debut, $465M raise and Digit 5 contracts, November 5; hands-on with Siri Recaps in the iOS 27.3 beta, November 4); Madrobot and APIyi launch analyses checked against Tencent's Hy4 announcement and OpenRouter's listing (Hy4's vision and computer use, preview model-card size of 770B total / 49B active, ~1M context, $0.90/$2.70 list pricing, 50% TokenHub discount through December 31, weights planned but not yet released, November 4); FourWeekMBA and AI Notis checked against Moonshot's Hugging Face model card (Kimi K3.1 Pro weights published November 4 — 2.8T total, 104B active, 1,048,576 context, 96 files, ~1.56 TB, $20M model-as-a-service license clause, $2/$10 API pricing); BusinessWire (Snowflake's Expedition 2026 CoCo and CoWork announcement — 9,100+ CoCo and 5,800+ CoWork accounts, Bots in private preview soon, Automations generally available, November 3); Futurum Group and Silicon Report (Salesforce's definitive agreement to acquire Cresta, November 2 — close expected February–April 2027, $2.7B price as reported by The Information); The Daily Haze carrying ACCESS Newswire (ReFiBuy's Meta Muse support launch for brands, November 4); Dosa AI Tools and the SXF Brief (OpenAI's Pierre Computer Company acquisition and Images 2.6, Anthropic's Claude-assisted vault structure, CrowdStrike's early-access Agentic Adversary Emulation, and OpenAI's Daybreak Red on AWS GovCloud, November 4–5); AI Developer Code's daily summary (OpenAI's Linux-kernel vulnerability campaign — 2,308 candidates, 948 unique privilege-escalation findings as reported by OpenAI — and Google's Agentic Data Cloud updates, November 4); 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.