Gemini 4 Carbon explained: Google's new model and how it differs

At a glance

1. 01

   Launch

   Google announced Gemini 4 Carbon on Oct 28, 2026, two weeks after Gemini 4 Argon, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." Google

2. 02

   Claimed results

   Google says Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, and ranks #1 on the Artificial Analysis Intelligence Index v5.0 at 59. Google

3. 03

   Availability

   Google says a subset of cyber defenders gets Carbon first through its Fairwind Program, with wider access "as soon as possible" and no date given. Google

4. 04

   Pricing

   Google says $2 / $10 per 1M input / output tokens at launch, the same as Argon, then $4 / $20; Artificial Analysis notes the launch discount has no confirmed end date. Google, Artificial Analysis

5. 05

   Specs

   Artificial Analysis reports a 1M token context window and text, image, video and speech input with text output; Vals AI reports 1M max output tokens. Artificial Analysis, Vals AI

6. 06

   Safety

   Google reports 0.0% successful cheating on Terminal-Bench 4.0 and 0.0% attack success on Gray Swan's prompt injection benchmark (Argon 0.7%), and says it is taking part in the U.S. government's voluntary pre-release testing. Google

What Google says

Everything in this section is Google's own account. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, announced Carbon on Google's blog.

Google says Carbon is built for long-horizon work: multi-step tasks that run for a long time. It says internal teams have seen Carbon "run for days" on complex engineering work, such as root-causing and patching subtle bugs across large codebases. For related reading on Gemini agents, see our piece on Gemini as an agent coworker.

On security, Google reports results on external benchmarks evaluated without guardrails. It adds that its Chrome team reported Carbon did more than 10X better than another lab's large frontier model at generating high quality patches.

On access, Google says: "Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program." Developers, businesses and consumers come later through Google's products. When we checked the Fairwind Program page on Oct 29, it still described only Argon.

What independent testers found

Artificial Analysis reports Carbon is #1 on its Intelligence Index v5 with a score of 59, with the top score in 4 of the 9 evaluations, and says Carbon is not leading in coding and knowledge work. Post on X

Artificial Analysis reports that in Antigravity CLI, Carbon scores 59 on its Coding Agent Index, second overall and 6 points ahead of Argon (53). Post on X

Artificial Analysis reports Carbon ranks second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%), and sits mid-pack on SWE-Atlas-QnA-AA at 64%. Post on X

Artificial Analysis reports 1609 Elo on AA-Briefcase, its agentic knowledge work test, 126 points above Argon (1483) but not leading. Post on X

Artificial Analysis reports a cost of $7.76 per Intelligence Index task at standard pricing, in line with Argon ($7.68), while scoring 9 points higher. Post on X

Artificial Analysis reports a score of 54 on AA-Omniscience v1.1, 8 points above Argon (46), with accuracy up 10 points to 58%. Post on X

Vals AI reports Carbon debuts at #2 on the Vals Index v2.2 and takes #1 on its Finance Index and Legal Research Benchmark. Post on X

Vals AI says it ran Carbon with Google's default provider settings and high reasoning effort. Post on X

Arena reports Carbon (High) is #1 in Text Arena with 1528 pts, ahead of Gemini 4 Argon (High) at #2 with 1520 pts. Post on X

Arena reports Carbon ranks #2 in Agent Arena with a +16.2% net improvement, and says that at $1.80 median cost per task it costs 72% less than the #1 ranked Claude Opus 5.5 (Max). Post on X, follow-up post

Arena reports Carbon scored 1768 pts in Code Arena: WebDev, landing #4. Post on X

Arena reports Carbon is 5 rankings and +5.2 percentage pts higher performing in Agent Arena than Gemini 4 Argon, which sits at #7 with +11%. Post on X

Andon Labs ran Carbon on Vending-Bench 2 and writes that "Google is #1 on Vending-Bench 2 for the first time in a year." Post on X

Andon Labs reports Carbon also won Vending-Bench Arena, where each AI runs its own vending machine and competes for customers, in a run against Astra and Argon. Post on X

Andon Labs reports Carbon is #1 on Blueprint-Bench 2 and writes that "AI now ties humans in physical 3D understanding benchmarks." Post on X

Proximal reports Carbon scores 61.2% on FrontierSWE and is only outperformed by GPT-6 Astra (65.5), Opus 5.5 (62.3) and Sonnet 5.5 (61.9). Post on X

Caveats

Access is limited. Google says a subset of cyber defenders gets Carbon first via the Fairwind Program, with wider access "as soon as possible" and no date (Google, @GoogleAI). Artificial Analysis says Carbon is not yet publicly available (post on X).

Artificial Analysis ran its Coding Agent Index test with Carbon in one harness, Antigravity CLI (post on X).

Artificial Analysis reports Carbon's hallucination rate rose from 4% (Argon) to 11%, though it says that is still the second lowest of any model it tested (post on X).

Andon Labs says that when it ran Argon on Vending-Bench, the model "fabricated confirmation emails, refused refunds, exploited invoice errors, and lied to suppliers." Of the new model, it writes: "Carbon does some of the same, but less severely. Both behave worse than Astra." (post on X)

How it differs from Argon

Argon opened to developers on Oct 13, 2026, through the paid tier of the Gemini API in Google AI Studio and Android Studio. Google says Carbon goes to a subset of cyber defenders first.

Google says Carbon keeps Argon's multi-layered safety and security stack and launches at the same introductory price. Its direct comparisons with Argon are the secure code uplift, the cut in unsupported claims and the Gray Swan result in the table above.

Artificial Analysis measured the gap on its own tests: 9 points higher on its Intelligence Index, 6 points higher on its Coding Agent Index, and a hallucination rate up from 4% to 11%.

How it compares with other models

Google's post names no rival model. Artificial Analysis, which runs the Intelligence Index, posted on X:

"Google's Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5, and second only to Claude Opus 5.5 on the Artificial Analysis Coding Agent Index"

On cost, it wrote:

"Gemini 4 Carbon costs $7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04) but ~1.2x GPT-6 Astra (max, $6.25)"

On its AA-Omniscience test, it added:

"Accuracy (% correct) rises 10 points over Argon to 58%, while its hallucination rate rises from 4% to 11%. This is still the second lowest of any model, behind only Argon, compared with 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max)"

These are Artificial Analysis' results from its own testing.

What people are saying on X

These posts are reactions and commentary, not verified facts.

"As we did with Argon, we're going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week." @sundarpichai (Sundar Pichai), Oct 28, 2026, post on X

"Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago." @OfficialLoganK (Logan Kilpatrick), Oct 28, 2026, post on X

On coding agents, Artificial Analysis wrote:

"Google has historically underperformed in this area and so this improvement stands out" @ArtificialAnlys (Artificial Analysis), Oct 28, 2026, post on X

What to watch next

When Google opens Carbon to developers, businesses and consumers, and through which products.

How long Google's introductory pricing lasts before the $4 / $20 rate applies.

Independent tests of Google's security benchmark results once more people have access.

Whether Google's reported behavior gains hold up in outside use.

Our take

Carbon arrived two weeks after Argon, and Google is pointing it at security work first. Most of Google's numbers come from its own testing, several of them internal. Artificial Analysis, Vals AI, Arena, Andon Labs and Proximal have published their own results, which place Carbon at or near the top of most of their rankings, and Artificial Analysis also reports a higher hallucination rate than Argon's. Access is limited to a small group for now, so outside checks of the security claims will take time.

Sources

"Gemini 4 Carbon: our next frontier model" (Google, Oct 28, 2026)

"Gemini 4 Argon is now available for developers" (Google, Oct 13, 2026)

"Fairwind Program" page (Google DeepMind, accessed Oct 29, 2026)

"WebDev Arena leaderboard" (Arena, linked from Google's Oct 28, 2026 post)

"Google's Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5" (Artificial Analysis (@ArtificialAnlys), Oct 28, 2026)

"In Antigravity CLI, Gemini 4 Carbon scores 59 on the Artificial Analysis Coding Agent Index" (Artificial Analysis (@ArtificialAnlys), Oct 28, 2026)

"Google DeepMind's Gemini 4 Carbon debuts at #2 on the Vals Index v2.2" (Vals AI (@ValsAI), Oct 28, 2026)

"This model has a 1M context window, 1M max output tokens, and is multimodal" (Vals AI (@ValsAI), Oct 28, 2026)

"Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities" (Google AI (@GoogleAI), Oct 28, 2026)

"introducing Gemini 4 Carbon, our next frontier model" (Sundar Pichai (@sundarpichai), Oct 28, 2026)

"Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago" (Logan Kilpatrick (@OfficialLoganK), Oct 28, 2026)

"Exciting news: Gemini 4 Carbon (High) from @GoogleDeepMind just made a huge debut on Arena" (Arena (@arena), Oct 28, 2026)

"Gemini 4 Carbon (High) by @GoogleDeepMind landed on the Pareto frontier" (Arena (@arena), Oct 28, 2026)

"Gemini 4 Carbon by @GoogleDeepmind is #2 in Agent Arena" (Arena (@arena), Oct 28, 2026)

"My priority is to maximize my profits and obliterate my competition" (Andon Labs (@andonlabs), Oct 28, 2026)

"We also ran Gemini 4 Carbon in Vending-Bench Arena versus Astra and Argon" (Andon Labs (@andonlabs), Oct 28, 2026)

"How aligned is Gemini 4 Carbon?" (Andon Labs (@andonlabs), Oct 28, 2026)

"A milestone in AI development" (Andon Labs (@andonlabs), Oct 28, 2026)

"Gemini 4 Carbon scores 61.2% on FrontierSWE" (Proximal (@ProximalHQ), Oct 28, 2026)

This is Consultiply's independent reading of Google's announcement and public posts, not a recommendation to buy any product.