Technology

Home  •  Blog  •  Google's AI Double Play: Gemini...

Google's AI Double Play: Gemini 4 Neon Lands, and Gemini Can Now Drive Your Mac

A faster, cheaper frontier-class model and a desktop agent that clicks and types for you show Google pushing Gemini from chatbot to doer

A glowing glass sphere of swirling neon plasma floats above a laptop, streaming light into on-screen app windows as a cursor moves on its own

In barely 24 hours, Google shipped a faster, cheaper frontier model and handed its Gemini assistant the keys to your Mac.

Google has made back-to-back moves in the AI race this week. On Wednesday, November 18, it launched Gemini 4 Neon, a mid-tier model it says delivers near-flagship reasoning at higher speed and lower cost. A day later, on Thursday, November 19, it rolled out "computer use" in the Gemini app for Mac, letting the assistant click, type and move through desktop apps and websites to finish tasks for users. Together, the two launches show Google pushing Gemini from a chatbot that answers questions toward an agent that does the work.

What Gemini 4 Neon is, and where it fits

Neon slots into a Gemini 4 family that already includes Hydrogen for fast everyday help and Carbon for the hardest, long-horizon work. In its launch post, Google DeepMind's Koray Kavukcuoglu described Neon as bringing "frontier-level intelligence with fast end-to-end task completion" across software engineering, knowledge work and science workflows. It accepts text, images, audio and video in a context window of up to 1 million tokens.

In the Gemini app, the reshuffle is significant. According to 9to5Google, Neon replaces Argon and the older 3.1 Pro option for AI Pro and Ultra subscribers, leaving a model picker of Auto, Hydrogen for everyday help, Neon for deeper analysis and Carbon for code and complex work. Neon is also live in the Gemini API through Google AI Studio, in Gemini for business, and in Google's Antigravity coding platform for paid subscribers.

The numbers Google is pointing to

Google's case for Neon rests on a long list of benchmark results, all reported by the company. It says Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, ranking third overall behind larger, slower models, and ranks first on the Artificial Analysis Coding Agent Index with 58.6. Google also cites 65.2% on Terminal-Bench 4.1, 70.9% on the Vals Index (second place), first place on Zapier's AutomationBench at 56.3%, and a debut at No. 2 on Arena's new Agent Arena Index with a score of 1,336.

Pricing may matter more to developers than any leaderboard. Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. VentureBeat reports that those rates are 50% higher than Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.2 Sol, but 25% below the standard rates of Anthropic's flagship Opus 5.5, which Neon comes within about a percentage point of on two engineering benchmarks. Google says that at its default "high" setting, Neon costs 32% less per task than Carbon.

Some of the most striking claims come from customers. Glean's AI product lead Thai Tran said in Google's announcement that, in early evaluations, Neon "matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost." VentureBeat notes, however, that Google's post gives no absolute completion times for matched workloads, so that speed comparison rests on a customer's account rather than published measurements.

Security and safety claims

Google is also positioning Neon as a tool for cyber defenders. The company says it reaches 90.5% on the SEC-bench Pro vulnerability-finding test and 70.7% on a patching benchmark, and that a specialized "Neon for cyber" version will reach eligible members of its Fairwind Program over the coming days. On safety, Google says Neon keeps Carbon-level protections against severe misuse and resisted every attack in Gray Swan's indirect prompt injection benchmark when its real-time safeguards were on at the highest thinking level. As with the performance figures, these are Google's own evaluations, not independent audits.

ADVERTISEMENT

Gemini can now operate your Mac

The second launch puts that agentic pitch on the desktop. Google says users can describe a task and Gemini will navigate menus, type information and click through apps, working in an isolated background window with a live preview in the chat so the user's own screen and cursor stay free. Users choose which apps Gemini can access, can pause at any time, and approve key actions before they happen. Google's examples include batch color-grading photos, pulling weights and addresses from PDFs into shipping forms, building Google Slides decks from live dashboards, and staging resale listings from a folder of clothing photos.

The feature is rolling out in English in the U.S. for personal Google accounts of users 18 and older, on Apple Silicon Macs running macOS Sequoia 15 or later. TechCrunch noted that Google is "a little late to the game," since OpenAI, Anthropic and Meta already offer agentic features in their desktop apps, and that Google did not say when the feature would reach other markets or platforms.

The risks Google itself flags

Handing an AI agent control of a computer carries real risk, and Google's own documentation spells some of it out. According to AppleInsider, the support pages warn that Gemini can access active sessions in signed-in Mac apps, can read multifactor authentication codes visible on screen, and could be redirected by malicious instructions hidden in content in one app into acting in another. Granting it access to code editors or developer tools could enable shell commands or broader file-system changes, and some changes cannot be undone. AppleInsider stresses that these are documented risks, not evidence of actual harm, but still advises against giving any AI agent broad control of a primary Mac. The report also points out that Apple added tighter automation permission controls in a macOS 27.3 developer beta on November 16.

Why it matters and what to watch

Taken together, the launches sharpen Google's strategy: a capable, cheaper middle model to power everyday agent work, and a desktop surface where that work can happen. For developers, the immediate question is cost per finished task. VentureBeat calculates that Neon would need to use roughly a third fewer tokens than Sol or Sonnet to offset its higher list price. For consumers, the test is trust, meaning whether approvals and app-by-app permissions prove strong enough once agents meet messy real-world inboxes and websites.

Near-term milestones are already set. Google's promotional pricing for Carbon ends on November 30, which 9to5Google says will widen Neon's price advantage over it, and the cyber-focused version is due to reach Fairwind members within days. Independent benchmarking and early user reports over the coming weeks will show whether Neon's numbers hold up outside Google's own tests.

Verified Sources & References
TAGS: #Google #Gemini #AI Models #AI Agents #Mac #Artificial Intelligence
SHARE THIS STORY
Entro

Entro

Staff Writer & Senior Contributor

Curating in-depth stories, analysis, and ideas for Enterteno readers. Committed to editorial accuracy, independent reporting, and original perspective.