Gemini 4 Carbon Benchmarks

Google's table puts Gemini 4 Carbon first among five frontier models on 15 of 17 benchmarks. At Layer3Labs, we build launch guides for new artificial intelligence (AI) models and route client workloads across vendors, so we read benchmark tables row by row to decide which model gets which task.

Gemini 4 Carbon's first-place rows include the Artificial Analysis Intelligence Index v5.0 at 59 and SEC-bench Pro at 88.2 percent, while GPT-6 Astra leads FrontierSWE 2 and Claude Opus 5.5 leads OSWorld-2.1.

Google announced Gemini 4 Carbon on October 28, 2026, positioning it as its strongest frontier model for cybersecurity and long-horizon tasks. Every score here comes from Google's published table or from Artificial Analysis, and Google publishes its evaluation methodology separately.

Gemini 4 Carbon Scores against Argon, GPT-6 Astra and Claude Opus 5.5

In our internal analysis of Google's published table, we counted the top score in each row across the five compared models and found Gemini 4 Carbon highest on 15 of 17 rows. The full table below lets you recount it.

Google published evaluation results comparing Gemini 4 Carbon against Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 across 17 distinct benchmarks.

Every row comes from Google's Gemini model page, and "Not reported" marks a score Google did not publish.

Where Gemini 4 Carbon Leads

Gemini 4 Carbon claims the top score across 15 of the 17 reported benchmarks.

Gemini 4 Carbon scores higher than Gemini 4 Argon on all 17 rows. Our Gemini 4 Argon review lists Argon's own scores, and our Gemini 4 Carbon vs Gemini 4 Argon comparison covers whether to wait for Carbon.

The widest gains appear in vulnerability research, long-context reasoning, and scientific terminal operations. In SWE-sweep, Carbon more than doubles Argon's score, moving from 4.3 percent to 9.6 percent.

Where GPT-6 Astra and Claude Opus 5.5 Lead

Rival frontier models outperform Gemini 4 Carbon on two specific benchmarks in Google's published report.

On FrontierSWE 2, GPT-6 Astra takes first place with 65.5 percent, while Claude Opus 5.5 scores 62.3 percent and Gemini 4 Carbon records 61.2 percent.

On OSWorld-2.1, Claude Opus 5.5 takes the top spot with 87.2 percent, topping Gemini 4 Carbon at 81.5 percent and GPT-6 Astra at 80.9 percent.

You can compare Carbon against Claude Opus 5.5 on price and access, or see Anthropic's full score profile in our Claude Opus 5.5 benchmarks guide.

These results indicate that teams managing software engineering workflows or operating graphical computer-use agents must test their exact tasks rather than assuming one frontier model leads across every domain.

Cybersecurity Scores and the Guardrail Conditions

Gemini 4 Carbon records high scores across all three dedicated cybersecurity benchmarks published by Google.

On SEC-bench Pro, Carbon scores 88.2 percent, ahead of Argon's 83.4 percent and GPT-6 Astra's 85.4 percent. On ExploitGym, Carbon reaches 54.8 percent compared to 42.4 percent for Astra and 42.0 percent for Argon. On CWE-bench, Carbon scores 71.0 percent, ahead of 68.0 percent for Argon and Astra.

Google says Gemini 4 Carbon's external cybersecurity benchmarks were evaluated without guardrails, so these scores do not show how Carbon performs with its guardrails on.

Defenders planning to run vulnerability discovery or exploit analysis through production interfaces must understand that standard production guardrails may alter performance once general access opens.

Long-Context and Agent Benchmark Performance

Gemini 4 Carbon posts distinct advantages across long-context retrieval and multi-step autonomous agent benchmarks.

On the GraphWalks benchmark covering 256k to 1M tokens with breadth-first search F1, Carbon scores 96.2 percent. That outperforms Argon at 84.2 percent, GPT-6 Astra at 71.8 percent, and Claude Opus 5.5 at 66.8 percent.

On Agents' Last Exam v1.1, Carbon reaches 77.4 percent, surpassing GPT-6 Astra at 73.1 percent and Claude Opus 5.5 at 71.6 percent.

In simulation tasks, Carbon achieved a final bank balance of $16,484 on Vending-Bench 2, leading GPT-6 Astra at $15,515 and Argon at $13,718. On Terminal-Bench 4.0, evaluated by Artificial Analysis, Carbon posted 65.2 percent against Opus 5.5 at 59.6 percent and Astra at 59.1 percent. On AutomationBench by Zapier, Carbon recorded 53.9 percent, leading Argon at 51.3 percent, Opus 5.5 at 42.5 percent, and Astra at 41.4 percent.

Safety Figures Google Published

Google reported several safety and code quality figures alongside its external benchmark results.

On the Gray Swan indirect prompt-injection test, Carbon recorded an attack success rate of 0.0 percent, improving on Argon's 0.7 percent.

Google also published two internal, non-independent findings. Google states that Carbon delivers more than a two-times uplift in success rate for generating secure code over Argon, and cuts unsupported claims by approximately two times over Argon.

On WebDev Arena, Google reports that Carbon gained 89 Elo points to reach the fourth position with a score of 1768.

Because these secure code and unsupported claims figures are Google-internal measurements, teams should verify them using their own prompts and verification suites once access permits.

How to Use These Benchmark Scores in Production

Benchmark figures should guide workload routing rather than dictating a single model across an entire system.

For teams planning deployments, match models directly to task profiles. Carbon presents a strong profile for large-scale graph analysis, long-context reasoning, and security inspection. Claude Opus 5.5 remains the strongest choice for desktop computer use on OSWorld-2.1. GPT-6 Astra leads for complex software engineering evaluations on FrontierSWE 2.

For 1M input tokens plus 200k output tokens of usage, Gemini 4 Carbon costs $4.00 at its launch price and $8.00 after the introductory rate ends. Gemini 4 Argon costs $4.00 through November 30, 2026, and $8.00 from December 1, 2026, the same $8.00 as Claude Opus 5.5. GPT-6 Astra costs $20.00 when every prompt stays within its 272k context tier.

Who this is not for: teams requiring independently verified third-party evaluations should not rely exclusively on these figures today. Every row in the comparison table originated from Google's self-reported disclosures, and external cybersecurity runs omitted guardrails. Artificial Analysis separately lists Carbon at 59 on its Intelligence Index v5.0, ranked number one among 52 models.

Our assessment would shift if independent benchmark runs show lower results under standard production guardrails, or if rival providers release updated models that surpass these scores.

Review Google's published evaluation methodology to inspect test configurations, and run your own standardized evaluation dataset against Carbon once access expands to your team before altering existing routing.