AI News Today — October 29, 2026: Google Gemini 4 Carbon Tops AI Benchmarks for Cybersecurity, Anthropic Cuts Claude Prices 40% for Government, and Muse AI Reaches Singapore

The big picture

October 28 was the day the AI race showed its receipts. Google launched a new frontier model, Gemini 4 Carbon, but gave it to cybersecurity defenders first instead of the public. Anthropic spent the day on the money side of AI: a 40% Claude discount for every level of U.S. government and new tools that show companies what their token spending actually produces. And earnings made the concentration plain — Microsoft disclosed Azure revenue in dollars for the first time, $33.2 billion in one quarter, and $10 billion of it came from OpenAI alone.

Underneath the headlines, the agent platforms kept opening up: Google let any developer build for Gemini and Search, Chrome learned to pull form details out of your Gmail, and Meta's Muse added Singapore as its fourth market. More capable, more connected, more embedded — and, for the first time, priced and measured in public.

What happened today

Google Gemini 4 Carbon tops the benchmark chart — and launches defenders-first for cybersecurity

What happened: Google unveiled Gemini 4 Carbon on October 28, just two weeks after Gemini 4 Argon, calling it its strongest frontier model with major gains in coding, long-horizon agent work, and cybersecurity. Artificial Analysis places Carbon first on its updated Intelligence Index v5.0 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. On the security benchmarks Google led with, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71% on CWE-bench for patching, and Google says it generates secure code at more than twice Argon's success rate on internal tests. Pricing matches Argon's introductory rates — $2 per million input tokens and $10 per million output, with cached input 95% off — stepping up to $4 / $20 after the introductory window. The catch is access: Carbon rolls out first to a subset of cybersecurity defenders through Google's Fairwind Program while safeguards are tested, with broader release promised "as soon as possible." Those benchmark figures are Google's and the evaluator's published results; independent reproduction is still pending, and Artificial Analysis testing also found a higher hallucination rate than Argon on one factual-knowledge test.

Who it matters to: Developers, security teams, and anyone choosing a frontier model this quarter. A top-ranked model you cannot yet use is a roadmap signal, not a buying option — and the defenders-first rollout tells you where Google thinks the capability risk sits.

What you can do: Do not rebuild a workflow around Carbon until your access actually opens. If you budget on the $2 / $10 introductory rate, plan for the published $4 / $20 step-up, and if your work is security-sensitive, look at whether the Fairwind route applies to your organization.

Learn more: VentureBeat on Google's Gemini 4 Carbon launch and SiliconANGLE on Gemini 4 Carbon's cybersecurity upgrades

Anthropic cuts Claude 40% for government — and launches tools to prove what AI spending buys

What happened: Two Anthropic moves landed within a day of each other, both about money. First, Anthropic expanded its partnership with the U.S. General Services Administration (GSA): a 40% discount on Claude Enterprise, Claude for Government, and the Claude API, extended for the first time beyond federal agencies to state, local, tribal, and territorial governments. There is no minimum spend — agencies pay only for what they use — the discount applies across every product surface, direct today and through cloud providers in the coming months, and Anthropic estimates it will save the government $1 billion. Anthropic says about 500 government organizations already use Claude. Second, Anthropic launched enterprise reporting tools that show managers what a team produced over a date range, what it cost, where workflows broke down, and how to improve the results, alongside pooled group budgets. Anthropic says the data is not used for model training by default.

Who it matters to: Public-sector buyers first — a 40%, no-minimum, consumption-based price changes the procurement math for smaller agencies that could never sign a large upfront deal. For private companies, the reporting tools matter just as much: "what did this token spend produce?" is now a product feature, not a spreadsheet exercise.

What you can do: If you work in or sell to government, check the OneGov ordering path before renewing anything at list price. If you run Claude at work, turn on the new reporting before your next budget review — you will be asked what the spend bought, and now there is an answer inside the product.

Learn more: Anthropic on expanding its GSA partnership, the GSA's OneGov announcement, and TechCrunch on Anthropic's AI spending tools

Muse AI launches in Singapore, its fourth market in seven weeks

What happened: Meta's Muse personal AI agent is now live in Singapore, with Meta's newsroom stamping the launch October 27. Singapore is Muse's fourth market since its September debut, after the United States, Canada, and Japan (October 23). Meta named eight local connectors — Shopee, Grab, Singapore Airlines, Agoda, ShopBack, Klook, Trip.com, and Lazada — spanning shopping, groceries, dining, and travel, and was explicit that the connectors switch on progressively, so day one is a partial shelf. The trust model travels with the product: credentials stay on a dedicated Muse Secure VM out of the agent's view, and Muse checks with the user before it sends an email or buys something. Muse's workplace tier remains announced but unshipped and unpriced.

Who it matters to: Anyone tracking how fast personal agents go global. The pattern is now clear: each new country arrives with local shopping and travel partners attached, because an agent is only useful where it can actually transact.

What you can do: If you are outside the four live markets, there is still nothing to sign up for — Meta has published no expansion timeline. If you build services, watch the connector list: being connectable to Muse, Gemini, and ChatGPT is becoming a distribution channel in its own right.

Learn more: Forkast on Muse's Singapore launch

Gemini Spark: no new announcement — Task mode stands and the November 17 Gems deadline is 19 days out

What happened: There was no new Gemini Spark announcement on October 28–29. Spark still lives inside the main Gemini app as Task mode, with the Chat/Task switch above the prompt box and a Scheduled page for active, paused, completed, and suggested tasks, and it still requires a Google AI Pro or AI Ultra subscription with a personal Google account. Note the distinction from today's headline: Gemini 4 Carbon is a developer frontier model, not a Spark upgrade — it does not change what Task mode does. The next hard date for Spark users remains the migration of Gems to Skills on November 17, 2026, now close enough that waiting for the automatic move is a gamble.

Who it matters to: Spark and Gemini power users with custom Gems built into their routines. The product is quiet; the deadline is not.

What you can do: Move your most important Gem by hand this week: copy its name, description, and instructions into a Skill from Gemini's settings, re-upload its knowledge files, and test it — while the original Gem still works as a fallback.

Learn more: FindSkill on where Gemini Spark went — it's Task mode now

Google opens Gemini and Search to any developer — and Chrome Autofill can now read your Gmail

What happened: Two Google openings landed the same day. First, Google launched its Connected Apps Platform in U.S. beta preview: any business or developer can build an app for Gemini and Search using a cloud-hosted MCP server or Android MCP, submit it through the Google Connected Console, and track traffic, latency, and errors. More than 50 connected apps already work in Gemini, and those integrations expand into Search over the next few weeks. Second, Chrome's AI-powered Autofill gained Gmail access: it can suggest form details pulled from your email — a flight confirmation number, membership ID, tracking or order number, address, or phone number. It is an expansion of Google's Personal Intelligence, it is opt-in in Chrome settings, Google says sensitive data is encrypted, and it is rolling out to U.S. desktop users and to Gemini in Chrome on iOS and Android.

Who it matters to: Businesses get a new front door: if customers start tasks inside Gemini or Search instead of opening your app, a connected app is how they still reach you. Users get a convenience with a real trade — autofill that reads email is helpful exactly because it can reach private information.

What you can do: If you run a business with an app or booking flow, join the Connected Apps waitlist and decide whether you want to appear in Gemini, Search, or both. If you use Chrome, turn enhanced autofill on only if the Gmail connection is worth it to you — and know where the off switch is in settings.

Learn more: TechCrunch on Google's Connected Apps Platform and TechCrunch on Chrome's Gmail-powered Autofill

Microsoft discloses Azure revenue for the first time: $33.2 billion — and $10 billion of Microsoft's quarter came from OpenAI

What happened: Microsoft reported Azure revenue in dollars for the first time: $33.2 billion for the July–September quarter, up 48% year over year and slightly ahead of Google Cloud's $30.7 billion for the same period. It was Microsoft's tenth straight record quarter: $92.3 billion in total revenue, up 19%, with net income of $36.6 billion and record capital spending of $45.2 billion. Inside that total, Microsoft recorded $10 billion in revenue from OpenAI in the quarter — nearly 11% of everything Microsoft made — from revenue-sharing and commercial arrangements disclosed in a regulatory filing, with $7.2 billion in outstanding receivables from OpenAI at quarter's end.

Who it matters to: Anyone trying to judge whether AI spending is real revenue or circular accounting. Both readings got evidence today: cloud demand is enormous and growing, and a striking share of the largest player's growth is one AI lab paying its investor-cloud provider.

What you can do: If you buy cloud or AI services, use the newly comparable dollar figures — Azure $33.2 billion versus Google Cloud $30.7 billion this quarter — when vendors negotiate as if scale were a mystery. Amazon reports next, which completes the comparison.

Learn more: TechCrunch on Azure's $33.2 billion disclosure and TechCrunch on Microsoft's $10 billion from OpenAI

Liquid AI's LFM3 puts six open-weight models on phones and Raspberry Pis

What happened: Liquid AI released LFM3, a family of six open-weight models in the roughly 2-billion-parameter class, all branching from one shared backbone: text (with a 256,000-token context window), vision-language, audio, an omni model handling text, images, video, and audio, and two decision models that return calibrated answers instead of generated text. Liquid says the vision model performs within about a point of a model more than six times its size on its reported average, the text model fits in about 1 GB at 4-bit precision, and the family is tuned to run across phones, laptops, Raspberry Pi-class boards, and NPUs with support for llama.cpp, MLX, vLLM, and Ollama. The weights are on Hugging Face under the LFM Open License — free for commercial use below a $10 million annual revenue threshold, per the license terms. The benchmark figures are Liquid AI's own and have not been independently verified.

Who it matters to: Builders who need AI that runs on the device — for privacy, latency, offline use, or cost. A fine-tunable 2B model that runs on one GPU or a phone changes what small teams can ship without an API bill.

What you can do: Download the text or vision model from Hugging Face and test it on your own task before believing the comparisons. Liquid itself notes the small text model is not recommended for agentic coding or complex multi-step reasoning — use it for local agents, tool calling, and structured outputs.

Learn more: Liquid AI's LFM3 announcement

Also worth knowing

• Google is taking AI Search ads global as core growth slows. Alphabet reported $127.6 billion in third-quarter revenue, up 25%, with Google Cloud up 103% to $30.7 billion — but Search advertising growth slowed to 12%, and Google is expanding ads inside AI Overviews and AI Mode internationally for the first time. Alphabet shares fell 8% after the results.
• OpenAI reportedly plans to buy Google's AI chips — reported, not confirmed. CoinCentral, summarizing reporting first published by The Information and also carried by Reuters and Bloomberg, says OpenAI plans to purchase Google TPUs; no size, price, or timing was disclosed, and both companies declined to comment. Treat it as a reported plan, not a signed deal. OpenAI is also developing its own chip with Broadcom.
• Nvidia released a free open transcription model. Nemotron 3.5 Transcribe is a 30-billion-parameter open-weight model that transcribes up to 40 languages, translates 31 to and from English, handles recordings up to an hour, and separates up to eight speakers — for recorded audio, not livestreams — alongside version 2 of Nvidia's open Granary speech dataset.
• South Korea's financial sector was hit by another wave of cyberattacks. TechCrunch reports Daishin Securities said personal data for 38,490 customers was exposed, while Mirae Asset Global Investments and crypto exchange Bithumb reported separate incidents without confirming data loss; regulators called for stronger measures.
• The "vulnpocalypse" warning is getting louder. Wiz's CTO argues that AI tools finding ancient, exploitable flaws in aging internet code — the pattern behind this month's PHP findings — is about to become a sustained wave for defenders, not a one-off scare.
• AI executives may be pitching a future the public does not want. A TechCrunch commentary on reaction to AI assistants for parenting and friendship cites Pew research: 50% of U.S. adults think AI will reduce people's ability to form meaningful relationships, and only 5% think it will make connections easier.

One practical takeaway

Measure your AI spend before the December price resets do it for you. This week handed you both sides of the ledger: Anthropic now sells the measurement itself, Carbon's introductory $2 / $10 pricing already has a published step-up, and Microsoft's filing shows how large these flows get once usage compounds. Pick one workflow this week, record what it costs and what it produced, and test a small open model like LFM3 or Nemotron Transcribe on the parts that never needed a frontier API in the first place.

---

Sources: VentureBeat, SiliconANGLE, TechCrunch AI, Anthropic, U.S. General Services Administration, Forkast, Liquid AI, FindSkill, CoinCentral (summarizing The Information / Reuters / Bloomberg reporting on the unconfirmed TPU plan). Summaries are original; follow the links for full reporting. Gemini 4 Carbon benchmark figures are Google's and Artificial Analysis's published results, not yet independently reproduced. Anthropic's $1 billion government savings figure is Anthropic's own estimate. LFM3 performance figures are Liquid AI's own. There was no new Gemini Spark announcement on October 28–29.