Google DeepMind Releases Gemini 4 Carbon, Claims Top Benchmark Spot With Restricted Access

Google DeepMind announced Gemini 4 Carbon on October 28, positioning it as the company's strongest model yet and claiming the top score on the Artificial Analysis Intelligence Index v5.0. The release comes two weeks after the broader rollout of Gemini 4 Argon on October 13. Carbon is initially available only to selected cybersecurity defenders through Google's Fairwind Program, with no date set for general access [1][2].

Benchmark Claims and Independent Verification

Artificial Analysis reports Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57), and Argon (50) [7]. The evaluator attributes the lead to top scores in complex reasoning, knowledge retrieval, and scientific tasks [7]. However, the index was recently revised: 60% of v5 uses private tests or withheld answers, and several components were replaced or updated [3]. Carbon's lead and its comparison with Argon are on this revised version and should not be mixed with older index scores [3].

On the Coding Agent Index, which evaluates models inside a coding tool (Antigravity CLI), Carbon scores 59 — three points behind Claude Code with Opus 5.5 at 62 [2][7]. Artificial Analysis places Carbon second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%) [3][7], but the individual component leadership is unclear: Source [7] states Opus 5.5 (max) scores 56% without specifying which benchmark that refers to, and if it applies to Terminal-Bench 4.0, Carbon's 64% would place it first. The composite index shows Opus 5.5 ahead overall. In agentic knowledge work (AA-Briefcase), Carbon reaches 1,609 Elo, a 126-point gain over Argon, but trails Grok 4.7 (1,638), Muse Wonder (1,659), and Kimi K3.1 Pro (1,678) [2].

Vals AI places Carbon second on its Index v2.2 at 70.36% accuracy, behind Claude Sonnet 5.5 (70.85%) and ahead of Opus 5.5 (70.18%) and Argon (69.24%) [2]. Carbon also leads Vals' Finance and Legal Research benchmarks. However, tasks take more than twice as long as with Argon, and Carbon generates 2.3x more reasoning tokens and 64% more output tokens [2]. Arena.ai puts Carbon first in Text Arena (1,528) and second in Agent Arena (+16.2% net improvement) behind Opus 5.5 Max (+21.1%) [2].

The Decoder reports Google's benchmark selection shows Carbon leading in 15 of 16 tests [2]. VentureBeat counts 17 reported rows in Google's comparison and says Carbon leads in 15 of them [3]. In that comparison, OpenAI's GPT-6 Astra remains ahead on FrontierSWE 2 (65.5% vs 61.2%) and Anthropic's Opus 5.5 leads on the OSWorld-2.1 offline subset (87.2% vs 81.5%) [3][5]. Google has not released a technical report, parameter count, or model architecture details [3]; Ars Technica notes Carbon could be a post-training upgrade of Argon rather than a new architecture [5].

Cybersecurity Focus and Guardrail Caveats

Google organizes Carbon's security results around three stages: vulnerability discovery (SEC-bench Pro: 88.2%), exploitation validation (ExploitGym: 54.8%), and patching (CWE-bench: 71.0%) [1]. The company states these external cybersecurity evaluations were conducted "without guardrails" [1]. That means the scores reflect the unmitigated model; customers will receive a version with safety controls that may yield different results [3][4].

Jake Handy's review notes the ExploitGym headline used a 10-hour cap, while Opus 5.5 was evaluated at 6 hours; at 6 hours Carbon scores 53.5% [4]. Google's internal claims — more than 2x Argon's secure-code success rate, Chrome team observing >10x better patches than an unnamed competitor's model, Wiz reporting leading web penetration testing benchmarks — are not independently verifiable [1][3][4].

Restricted Rollout and Pricing

Access begins this week with a subset of cyber defenders, expanding to trusted partners via the Fairwind Program. Google says it is participating in the U.S. government's voluntary pre-release testing process and will gather feedback before opening Carbon to developers, businesses, and consumers "as soon as possible" [1][2]. The Fairwind Program prioritizes governments, critical infrastructure operators, and core technology platforms, requires background checks, phishing-resistant MFA, tracked usage, and bans malware creation [4]. As of October 29, the Fairwind page describes Argon access and does not mention Carbon, so terms may differ [4].

Pricing matches Argon's introductory rates: $2 per million input tokens, $10 per million output tokens, with cached input at 95% off. Standard prices are $4 and $20. Google's Argon developer guide puts higher rates into effect December 1, 2026; Carbon's end date is unspecified [3]. Artificial Analysis calculates Carbon's cost per Intelligence Index task at $7.76 (standard pricing), about 23% less than Opus 5.5 Max ($10.04) but ~1.2x GPT-6 Astra ($6.25) [2][7]. Carbon uses 123k output tokens per task, 29% fewer than Opus 5.5 Max (173k) [7].

Safety Claims and Missing Evidence

Google reports Carbon halves unsupported claims versus Argon, reduces high-severity misalignment alerts by almost 80% in internal coding, cuts overreach by 26% and under-service by 29%, and achieves 0.0% attack success on Gray Swan's indirect prompt injection benchmark (vs Argon's 0.7%) [1]. On Terminal-Bench 4.0, successful cheating drops to 0.0% (attempts 0.4%) [1]. These are company-reported measurements from internal environments; external reproduction is pending [3].

The Argon Safety and Security Report, linked from Carbon's announcement, states Google "cannot rule out" that unmitigated Argon reaches the cybersecurity Critical Capability Level (CCL) and biological CCL [8][9]. Carbon's announcement does not include a separate system card or equivalent risk assessment [1][4]. Google says Carbon uses the same layered safety stack as Argon — input screening, activation probes, continuous monitoring, chain-of-thought monitoring — with additional alignment work for high-pressure tasks [1][8].

Artificial Analysis finds Carbon's hallucination rate on AA-Omniscience rises from 4% (Argon) to 11%, though accuracy improves from 48% to 58% [2][7]. The benchmark defines hallucination rate as incorrect responses divided by incorrect + partially correct + unanswered; 11% does not mean 11% of all answers are fabricated [3]. Carbon still has the second-lowest hallucination rate among tested models, behind Argon; Opus 5.5 Max and GPT-6 Astra Max each sit at 50% [2][7].

What Remains Unknown

• No Carbon model card, API specification, or context window confirmation from Google (third parties report 1M tokens) [4].
• No end date for introductory pricing.
• No general availability timeline.
• Cybersecurity benchmarks run without guardrails; deployed performance unknown.
• Internal claims (Chrome team, Wiz, secure-code uplift) lack independent verification.
• Whether Carbon is a new architecture or a post-training upgrade of Argon.

Google's two frontier releases in a month return it to the top of the main independent leaderboard, with margins of only a few points over Anthropic and OpenAI [5]. For most enterprises, the practical distinction remains: Argon has documented public access, API behavior, and a dated pricing transition; Carbon has stronger reported results, matching advertised token rates, and a narrower initial audience [3].