Gemini 4 Carbon: Google's New Table Leader, Gated to Cyber Defenders

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon reached general availability. On Google's own comparison table, Carbon beats Argon on every row and posts the top score on 14 of 16. The widest leads sit in long context, exploitation, and knowledge work. Two clear losses remain: FrontierSWE 2 to GPT-6 Astra, and OSWorld-2.1 to Claude Opus 5.5.

Access is the gate. Carbon starts with a subset of cyber defenders, then phased trusted partners through Google's Fairwind Program. Developers, businesses, and consumers come "as soon as possible," with no date. There is no public API and no model code. Scores below are self-reported by Google.

At a Glance

• Maker: Google DeepMind
• Announced: October 28, 2026 (Google blog, Koray Kavukcuoglu), two weeks after Argon reached general availability
• Google's framing:"our strongest model with state-of-the-art cybersecurity and long horizon capabilities"
• Access: subset of cyber defenders now; phased trusted partners via Fairwind; developers, businesses, and consumers later with no date
• Public API: none. No model code. Not in AI Studio, Vertex AI, the Gemini app, OpenRouter, or Vercel AI Gateway as of October 28, 2026
• Context window / max output / cutoff: not published
• Pricing (announced, not purchasable): introductory $2 / $10 per 1M (cached input $0.10), then $4 / $20 (cached $0.20). Intro end date not given
• Google's table: ahead of Argon on 16 of 16 rows; top score on 14 of 16
• Verification: scores self-reported by Google, not independently verified by third parties or by LLM Stats

Can I Use It? What to Do Now

Almost nobody can. Starting the week of October 28, 2026, Google is making Carbon available to a subset of cyber defenders, then gradually phasing access to additional trusted partners through the Fairwind Program. Google is engaged in the U.S. government's voluntary pre-release testing process. Developers, businesses, and consumers get Carbon "as soon as possible." No date.

Who qualifies for Fairwind

Priority partners: governments and national cyber authorities, critical infrastructure operators, and core technology platforms. Academic labs focused on defensive benchmarking may apply. Applicants are vetted. Access is only for internal cybersecurity, incident response, or penetration-testing teams. No sharing, redistribution, or resale. The program works with over 650 partners globally. As of October 28, 2026 the Fairwind page still describes Argon access, not Carbon. Non-eligible organizations are pointed to CodeMender with publicly available models.

Everyone else: ship on Argon

Build on Gemini 4 Argon now on the Gemini API paid tier (not the free tier), AI Studio, Gemini Enterprise, Antigravity, or the Gemini app on Google AI Ultra or Pro. Argon specs that matter today: 1,048,576-token context; thinking low, medium, high; implicit caching. Same announced introductory price as Carbon: $2 / $10 per 1M. Argon moves to $4 / $20 on December 1, 2026 (intro ends November 30, 2026). Keep prompts and evals model-agnostic so a Carbon swap is a config change when it ships.

Precedent, not a promise: Argon was announced September 30, 2026 with the same Fairwind-first, "as soon as possible" wording, and reached general availability October 13, 2026 (13 days later). Google has not said Carbon will follow that path.

Benchmarks, by Job

Prefer Google's comparison table for the hero read. Columns: Gemini 4 Carbon, Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1. Every score below is self-reported by Google. None of these scores has been independently verified by third parties or by LLM Stats. Methodology notes live on Google's evals methodology page.

Agentic coding

The biggest jump is Terminal-Bench 4.0 at 65.2 (a third-party run reported by Google), versus Argon 57.1 and Opus 59.6. SWE-sweep at 9.6more than doubles Argon (4.3) but is a dead heat with Astra (9.5): call it roughly tied. Gemini scores use Google's own mini-swe-agent harness; peers come from the public leaderboard. FrontierSWE 2 is the coding loss: Astra 65.5, Opus 62.3, Carbon 61.2. Carbon leads two of three coding rows and loses the third. Do not treat it as the top coding model on this evidence.

Computer use

Agents' Last Exam v1.1 at 77.4 (averaged rubric score) is +4.3 over Astra. OSWorld-2.1 at 81.5 (82-task offline subset, partial score) is +6.8 over Argon but 5.7 behind Opus 5.5 (87.2), Carbon's largest deficit on the table. Versus Astra (80.9) the OSWorld gap is +0.6: roughly tied.

Knowledge work

AutomationBench at 53.9 leads Opus 42.5 and Astra 41.4 by double digits. Vending-Bench 2 at $16,484 final bank balance leads Astra ($15,515) and Argon ($13,718).

Long context

GraphWalks 256k to 1M, BFS F1 at 96.2versus Argon 84.2, with no rival above 71.8. That is the largest lead on the table (+24.4 over Astra). Carbon's own context window is still not published; the 256k to 1M figure is the benchmark range.

Multimodal

Chartography at 72.5 versus Argon 71.6 and Astra 71.0 is a near tie: call it roughly tied. Blueprint-Bench 2 at 58.6 leads Opus 51.2 by 7.4.

Science and math

RiemannBench at 80.0 is +8.0 over Astra. Terminal-Bench-Science 0.1 at 64.8(a third-party run reported by Google) is +11.9 over Argon's 52.9from that same third-party setup in Google's table. Versus Astra at 63.3 the gap is +1.5: roughly tied.

General

Vals Index v2.2 at 70.4versus Opus 70.2 (+0.2): roughly tied. Footnote: Argon's column here is 69.2; Argon's own launch table printed 68.9.

Cybersecurity

Google's headline domain, evaluated without guardrails. ExploitGym at 54.8 (labeled as the 10-hour run; the default 6-hour score is 53.5) versus Astra 42.4, Argon 42.0, Opus 34.5: the biggest rival gap among cyber rows. SEC-bench Pro at 88.2 versus Astra 85.4. CWE-bench at 71.0versus a three-way 68.0 (Argon, Astra, and Grok 4.7 on Google's chart).

Where it leads clearly

• GraphWalks: 96.2 vs best rival 71.8 (+24.4)
• ExploitGym: 54.8 (10h; 53.5 at 6h) vs Astra 42.4 (+12.4)
• AutomationBench: 53.9 vs Opus 42.5 (+11.4)
• RiemannBench: 80.0 vs Astra 72.0 (+8.0)
• Blueprint-Bench 2: 58.6 vs Opus 51.2 (+7.4)
• Terminal-Bench 4.0: 65.2 vs Opus 59.6 (+5.6)
• Agents' Last Exam v1.1: 77.4 vs Astra 73.1 (+4.3)

Where it trails or ties

• FrontierSWE 2: trails Astra 65.5 (Carbon 61.2) and Opus 62.3
• OSWorld-2.1: trails Opus 87.2 (Carbon 81.5); roughly tied with Astra 80.9
• Vals Index v2.2: 70.4 vs Opus 70.2, roughly tied
• SWE-sweep: 9.6 vs Astra 9.5, roughly tied
• Chartography: 72.5 vs Argon 71.6 / Astra 71.0, roughly tied
• Terminal-Bench-Science 0.1: 64.8 vs Argon 52.9 on the same third-party setup; vs Astra 63.3 roughly tied

On the WebDev Arena leaderboard (dated October 27, 2026), Carbon (High) sits at 1768(±15), with 2,230 votes. The leaderboard marks that score preliminary and ranks it #4; the confidence interval overlaps ranks 2 to 6. Those preliminary / ±15 / #4 details come from WebDev Arena, not from Google. The +89 Elo is over Gemini 4 Argon (High) at 1679 on that board. Above Carbon sit Opus 5.5 (Max) 1810, Astra (Max) 1773, and Claude Sonnet 5.5 (xHigh) 1770. Crowd-vote leaderboard, not a Google self-report; it stays out of the chart above.

Google also makes qualitative claims we do not chart: models that "run for days" on complex engineering tasks, a claimed ">2x uplift in secure code generation," Chrome calling patches ">10X better than another lab's large frontier model," and Wiz calling Carbon leading on web pen-testing benchmarks. Treat those as Google says, not as measured facts.

Pricing and Access

Carbon "will launch" at the introductory rates below. None of it is purchasable today. Google gives no end date for Carbon's introductory period. Argon's dates are explicit and live on the Gemini API.

Carbon's standard price equals Opus 5.5's list price. Its intro price is half that and one-fifth of Astra's. For Argon, output price includes thinking tokens. Per-token price is not per-task cost; Google publishes no cost-per-task data for Carbon. Do not claim Carbon's caching is implicit-only (that is documented for Argon only).

Safety and Alignment

Google says Carbon uses the same multi-layered safety stack as Argon (see Argon's Safety and Security report). No Carbon-specific safety report or model card is published yet.

• Terminal-Bench 4.0:successful cheating 0.0%, attempts 0.4% (Argon's report: withheld information obtained in 1.4% of audited runs)
• Gray Swan indirect prompt injection attack success at k=15: Carbon 0.0% vs Argon 0.7%, Opus 5.5 1.0%, Fable 5.1 1.0%, GPT-6 Astra 8.5%
• High-severity misalignment alerts in live coding down almost 80% (Google says)
• Unsupported claims about 2x lower; overreach -26%, under-service -29% (Google says)

Who Should Care

• Security teams at governments, critical infrastructure, or core platforms:apply to Fairwind now. Carbon's cyber rows are the clearest lead on the table.
• Teams on Argon:nothing to do today. Keep building. Carbon is ahead on every row of Google's table, so plan an eval-driven swap when it ships. Watch December 1 pricing on Argon.
• Long-context and document-heavy agents: the GraphWalks gap is the strongest signal to wait for Carbon.
• Desktop automation: Opus 5.5 still leads OSWorld-2.1 by 5.7. Do not move for computer use alone.
• Hard software engineering:Astra leads FrontierSWE 2. Carbon's coding lead is Terminal-Bench, not repo-scale SWE.
• Price-sensitive buyers: at $2 / $10 Carbon would undercut Opus 5.5 and Astra, but only during an intro window Google has not dated.

FAQ

Outlook

On Google's table, Carbon is the breadth-and-cyber upgrade over Argon: 16 of 16 against its predecessor, 14 of 16 as the top score, with GraphWalks, ExploitGym, and AutomationBench as the clearest signals. It is not unbeatable on coding or computer use, and every number is self-reported by Google.

Open questions: an API date and model code; context and output limits; Carbon's intro price window; a Carbon model card or safety report; independent reruns of the cyber and long-context rows; whether the OSWorld and FrontierSWE gaps close. Until then, ship on Argon, watch Fairwind, and keep evals ready for a swap.

Rival framing for the same period: GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1.