The short version
- Gemini 4 Carbon launched first for vetted cyber defenders. It costs $2 per million input tokens and $10 per million output tokens at an introductory rate, and Artificial Analysis reportedly ranks it first on its Intelligence Index.
- Apple's iPrefBench finds phone agents complete actions well but often fail to infer or ask for your preferences. The best model completed only 47.7% of tasks fully.
- Alphabet's quarterly revenue rose 25% to $127.6 billion, Google Cloud roughly doubled, and capital spending reached $51 billion, turning free cash flow negative.
- A good assistant should ask before acting on anything it cannot infer. Tell yours your preferences up front.
October 29 brought a new frontier model that its maker considers too capable in cybersecurity to release to everyone at once, a benchmark that measures whether phone agents understand what you actually want, and earnings that show what building AI is costing Google. Each section explains what happened and how to apply it.
01
Gemini 4 Carbon: frontier model, defenders first
What Google released
Google DeepMind launched Gemini 4 Carbon on October 28. Google describes it as its most capable model at finding and fixing security vulnerabilities, and because the same skills can be used to attack systems, it is rolling Carbon out first to selected cyber defenders through its vetted Fairwind programme. Eligible groups include government cyber authorities, critical infrastructure operators and core technology platforms.
Independent evaluator Artificial Analysis reportedly ranks Carbon first on its Intelligence Index v5.0 with a score of 59. Google reports strong results on long-document tasks and says the model is less prone to hallucination and to gaming tests. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with higher rates planned after the introductory period. Wider availability was to follow in stages.
Using AI to fix vulnerabilities in your own code
You do not need Carbon to benefit from AI-assisted security. Most teams lose more to known, unpatched issues than to novel attacks. Run a dependency and code scanner such as Snyk on your main repositories, and sort the findings by severity and by whether the vulnerable code is actually reachable. Then ask an assistant such as Gemini or Claude to explain each critical finding and propose a fix, and make it write a test that fails before the fix and passes after.
Keep two rules. Never paste secrets, keys or customer data into a chat. And review every AI-generated patch as you would a colleague's, because a fix that silences the scanner is not always a fix. Our Claude vs Gemini comparison covers the general assistants, and the AI Security category lists more security tools.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Apple's iPrefBench: agents that act well but do not know you
What Apple's researchers found
Apple researchers published iPrefBench, a benchmark that tests whether AI agents on a phone can handle vague requests by either inferring your preference from your history or asking you. It runs in a working iOS environment with a simulated user, covering 26 apps, 130 deliberately underspecified tasks and about 3,000 records of the same user's past activity.
The results show a clear gap. Agents are good at operating apps but weak at personalisation. The best model reached 91.6% on completing the required actions but completed only 47.7% of tasks fully, and agents asked very few questions, often guessing when they should have checked. Adding a clarification step before acting reduced missed questions but increased unnecessary ones. The researchers conclude that a truly personal agent must judge well when to act on its own and when to confirm.
Getting better results from any assistant
You can close part of this gap yourself. Give your assistant the preferences it cannot guess: in the custom instructions or memory settings, state your usual choices, such as preferred airlines, dietary needs, meeting times, tone for emails and which contacts are work or personal. For tasks with consequences, add a line to your request: "Ask me before you choose anything I have not specified."
When an agent books, buys or sends something, review the summary before confirming. Treat each wrong guess as information: add the missing preference to your instructions so it does not happen again. Assistants such as ChatGPT, Claude and Gemini all support saved instructions.
03
Alphabet's quarter: cloud doubles, spending surges
What Alphabet reported
Alphabet reported third-quarter revenue of about $127.6 billion on October 28, up 25% from a year earlier but slightly below some forecasts. Google Cloud revenue roughly doubled to $30.7 billion, with operating income of about $11.1 billion. Capital spending reached $51.4 billion in the quarter, mostly for AI computing and TPU systems, which turned free cash flow negative at about –$9.1 billion. Net income fell to about $26.2 billion, partly because of a loss on equity investments. Alphabet kept its 2026 capital spending guidance at $195 billion to $205 billion and said spending will rise significantly in 2027. The shares fell after hours.
Reading a tech earnings release for what matters to you
For people who buy AI services, three lines in a release like this are worth reading: cloud revenue growth, which shows demand; capital spending, which shows how much capacity is coming; and any comment on pricing or capacity constraints. Strong cloud demand with rising spending usually means more capacity later, while short-term limits on the newest models may persist.
Try this with an assistant and check it: paste the earnings release and ask for a table of revenue by segment and capital spending for this quarter and the same quarter last year, with the source sentence for each number. Then verify each figure against the release. Excelly-AI can turn the table into a chart in your own spreadsheet. This is a reading exercise, not investment advice.
What to try next
Tell your tools what you want
Scan your own code for known vulnerabilities and use AI to explain and fix them, with tests. Write your preferences into your assistant's instructions so it guesses less. When you read AI earnings, look at cloud growth and capital spending rather than the share price.
Explore Chatbots, AI Security and Finance AI to build your shortlist. Our Gemini review covers Google's assistant in detail.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.