Gemini 4 Carbon: #1 on the AI index at 59, cyber defenders first
Google announced a new frontier model today and immediately locked it behind a pilot program for security teams. The capability is the news; the distribution is the story.
Gemini 4 Carbon is now Google's strongest model, and it tops the freshly revised Artificial Analysis Intelligence Index v5 at 59 — one point above the Claude Opus 5.5 build that led the same index this morning. The launch came two weeks after Gemini 4 Argon and was framed almost entirely around cybersecurity: Google says Carbon shows a "significant leap" in finding, validating and fixing vulnerabilities, and it is releasing the model first to a hand-picked group of cyber defenders rather than to developers or consumers. According to Google's announcement, access starts this week for a subset of defenders and widens gradually through the company's Fairwind Program, after the model clears the US government's voluntary pre-release testing process.
The numbers Google published are unusually specific for a cybersecurity claim. Carbon scores 88.2% on SEC-bench Pro, a benchmark for autonomously finding real vulnerabilities in shipping software; 54.8% on ExploitGym, which tests whether a model can actually exploit what it finds; and 71.0% on Collinear's CWE-bench, which measures whether it can produce a working root-cause patch. Google also says internal tests show more than twice the success rate at generating secure code compared with Argon, and that its Chrome team found Carbon more than ten times better at writing high-quality patches than "another lab's large frontier model" — a comparison it declines to name.
Coding and long-horizon work are the second half of the pitch. Carbon gained 89 Elo points on WebDev Arena to reach 1,768 and fourth place, and Google says internal teams have had it run for days on multi-step engineering tasks, working through dead ends on bugs spread across large codebases. The model also debuted on LMArena today at first place in the text arena, second in the agent arena and fourth in the web-development code arena, per Arena's own announcement.
Set that against this morning's picture. We wrote up the v5 index five hours before Carbon existed — Artificial Analysis Intelligence Index v5: Opus 5.5 leads at 58 — and the takeaway then was that the frontier had compressed into a two-point band. Carbon moves the leader by exactly one point. That is not a rout; it is a photo finish in which Google bought the lead with a model tuned for one domain.
The safety numbers are the part worth reading closely, because they are about behaviour rather than capability. Carbon scored 0.0% on Gray Swan's indirect prompt-injection benchmark, meaning it resisted every attack attempt in the evaluation — against 0.7% for Argon, roughly 1% for Opus 5.5 and, by Ars Technica's reading of Google's eval table, 8.5% for GPT-6 Astra. Google says successful cheating on Terminal-Bench 4.0 fell to 0.0% with attempts down to 0.4%, high-severity misalignment alerts in its internal coding environments dropped by almost 80%, unsupported claims roughly halved versus Argon, overreach fell 26% and under-service 29%. A model that fabricates less and escalates ambiguity instead of guessing is a quieter selling point than a benchmark crown, and a more useful one for anyone pointing an agent at a production codebase.
Pricing is where the launch gets conventional. Carbon arrives at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. Ars Technica reports that the rate doubles once the introductory period ends, and that Google has not said when that is; Argon's discounted window runs only through the end of November. Anyone building a budget on the $2 figure should assume it is a launch promotion, not a list price.
There is a reasonable skeptical case. Carbon looks less like a new architecture than a post-training upgrade of Argon — it beats its predecessor by slim margins across Google's own table, and Google's framing ("builds on Gemini 4 Argon") does not contradict that. Every published benchmark here is either self-run or was evaluated without guardrails, and the WebDev and Chrome comparisons are against unnamed competitors. The gated rollout also means independent evaluation of the released model is impossible for now: researchers cannot test what they cannot call.
The rollout deserves more attention than the scoreboard. Gating a cyber-capable model behind a defender program is now a pattern at Google — it did the same with its cyber model behind Fairwind earlier this year — and the mechanism is the US voluntary pre-release testing regime rather than a regulator with a statute. That is a real shift in how frontier capability reaches the public: capability first, wide access later, with the delay set by the lab and the government jointly. It also hands Google a defensible answer to the obvious dual-use question, because a model that finds vulnerabilities at 88.2% on SEC-bench Pro is a model that can be pointed at software its owner does not own.
We flagged Carbon's coding reputation when it was only an internal build — Google's Gemini 4 Carbon reportedly matches Opus 5.5 on coding — and the launch confirms the direction while adding the thing the leak could not: numbers, a price and a distribution plan.
Apple has put 512GB of unified memory in the Mac Studio, and the upgrade alone costs $15,500. The top M5 Ultra configuration — 36-core CPU, 80-core GPU — now takes 512GB of RAM, which brings the entry price for that spec to $22,299 with 1TB of storage; maxing storage to 16TB takes it to $29,799, according to 9to5Mac's read of Apple's configurator. Orders ship in late November to early December.
For local inference this is the number that matters: 512GB of unified memory is enough to hold open-weight models that no single consumer GPU can address, and Apple is currently the only vendor shipping that much memory in one box you can plug into a wall. It is also, at roughly $37 per gigabyte of added RAM, a reminder that the local-AI ceiling we described in M5 Ultra reviews: local AI's ceiling is memory, not compute has not moved so much as been priced. The machine is not for hobbyists; it is for teams that would otherwise rent eight accelerators by the hour.
Microsoft is ending Copilot support on smart TVs, less than two years after LG and Samsung put it in their sets. LG TVs began showing a notification this week saying the Copilot feature is "scheduled for discontinuation during 2026" and that the date is subject to Microsoft's timeline, as The Verge reported. There was no formal announcement from either TV maker or from Microsoft.
The retreat is easy to read as a strategy shift rather than a failure of the technology. Copilot on a TV was a chat companion for what was on screen; Microsoft has since rebuilt Copilot as a productivity and agent surface, and a television is a poor host for that. The lesson for consumer AI more broadly is blunter: assistants bundled into hardware without a job to do get removed when the strategy changes, and users who never asked for the feature are the ones who notice.
What to watch: whether Carbon's intro pricing gets a published end date, and whether the Fairwind pilot produces any independent benchmark run before the model reaches the API.
If a model can find 88.2% of real vulnerabilities, should labs ship it to defenders first or to everyone at once? Write to us via our contact page.
