Gemini 4 Carbon vs Gemini 4 Argon: Benchmarks, Pricing, and Key Differences
Google announced Gemini 4 Carbon on October 28, 2026, just two weeks after Gemini 4 Argon became broadly available. Carbon now leads the Artificial Analysis Intelligence Index v5.0, positioning it as Google's newest frontier model for complex reasoning and long-running agentic tasks.
However, Carbon is not yet available to everyone. Its rollout begins with selected cybersecurity defenders before expanding to trusted partners, developers, businesses, and consumers.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is the latest model in Google's Gemini 4 family, positioned above Gemini 4 Argon. It is designed for long-horizon tasks that require multiple steps, tool calls, and recovery from errors.
Google DeepMind says its internal teams have used Carbon for complex engineering work, including debugging and patching large codebases and low-level systems.
Its reported strengths include:
Coding and full-stack development
Agentic workflows and computer use
Scientific reasoning
Multimodal understanding
Google has not yet published Carbon's technical report, architecture, parameter count, or API specification. Its context window and supported input formats have therefore not been confirmed.
Gemini 4 Carbon vs Gemini 4 Argon
Feature | Gemini 4 Argon | Gemini 4 Carbon |
Announced | September 30, 2026 | October 28, 2026 |
Broad availability | October 13, 2026 | Limited rollout through Fairwind Program |
Artificial Analysis Intelligence Index v5.0 | 50 | 59 |
Introductory input/output price per 1M tokens | $2 / $10 | $2 / $10 |
Carbon improves on Argon across all 17 rows of Google's published benchmark comparison, while retaining the same introductory pricing.
Gemini 4 Carbon Benchmarks
Google-reported results
Carbon leads in 15 of 17 benchmarks in Google's comparison. Its two exceptions are:
Software engineering: GPT-6 Astra and Claude Opus 5.5 outperform Carbon on FrontierSWE 2.
Computer use: Claude Opus 5.5 leads on the OSWorld-2.1 offline subset.
Carbon's strongest reported advantage is long-context reasoning, where it scores substantially higher than GPT-6 Astra and Claude Opus 5.5 on GraphWalks.
On Vending-Bench 2, a simulated business benchmark, Carbon achieves a balance of $16,484, ahead of Argon and GPT-6 Astra.
Google notes that its results do not include uncertainty estimates. Its security benchmarks were also conducted without guardrails, so real-world performance may differ.
Independent evaluations
Artificial Analysis's evaluations broadly support Carbon's strong performance, with some important qualifications:
Intelligence Index v5.0: Carbon ranks first, narrowly ahead of Claude Opus 5.5 and GPT-6 Astra.
Coding Agent Index: Carbon scores 59, improving on Argon's 53 but trailing Claude Code with Opus 5.5 at 62.
AA-Briefcase: Carbon scores 1,609 Elo, up 126 points from Argon, but remains behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro.
AA-Omniscience: Carbon scores 54 and answers 58% of questions correctly. Its measured hallucination rate rises to 11%, compared with Argon's 4%.
WebDev Arena: Carbon ranks fourth with 1,769 Elo, an increase of 89 points.
Overall, Carbon is a strong general-purpose model, but specialized coding and computer-use workloads remain competitive.
Gemini 4 Carbon Cybersecurity
Cybersecurity is a central part of Carbon's launch. Google reports results across three stages:
Vulnerability discovery: 88.2% on SEC-bench Pro
Exploit validation: 54.8% on ExploitGym
Patching: 71.0% on CWE-bench
Google also reports that Carbon is more than twice as successful as Argon at generating secure code and performed over ten times better than an unnamed frontier model at producing high-quality patches.
These capabilities could help defenders identify and fix vulnerabilities, but they also create potential misuse risks. This is one reason Google is limiting access during the initial rollout.
Safety and Reliability
Carbon uses the same layered safety approach as Argon, including safety training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring.
Google reports the following improvements over Argon:
Unsupported claims reduced by roughly half in internal testing
High-severity misalignment alerts in live coding environments reduced by nearly 80%
Overreach reduced by 26% and under-service by 29%
No successful benchmark cheating on Terminal-Bench 4.0
No successful attacks on Gray Swan's indirect prompt-injection benchmark
These internal results should be considered alongside the higher hallucination rate measured by Artificial Analysis on its separate AA-Omniscience evaluation.
Gemini 4 Carbon Pricing
Carbon launches at the same introductory price as Argon.
Pricing | Introductory period | After introductory period |
Input per 1M tokens | $2.00 | $4.00 |
Output per 1M tokens | $10.00 | $20.00 |
Cached input | 95% discount (~$0.10) | 95% discount |
Google has not announced when Carbon's introductory pricing will end. For comparison, Argon's introductory pricing ends November 30, 2026.
A workload using 10 million input tokens and 1 million output tokens would cost approximately $30 at introductory rates or $60 afterward.
How to Access Gemini 4 Carbon
Carbon is currently available only to selected cybersecurity defenders. Google plans to expand access gradually through its Fairwind Program before opening it to developers, businesses, and consumers.
Until then, Gemini 4 Argon is available and offers the same introductory pricing, making it an option for teams preparing workflows for Carbon.
Conclusion
Gemini 4 Carbon builds on Argon's capabilities with stronger benchmark performance, particularly in long-context reasoning, agentic work, and cybersecurity. Its pricing matches Argon's introductory rates, making it an attractive upgrade once access expands.
However, Carbon does not lead every evaluation, and its public availability remains limited. Developers and businesses should weigh its reported strengths against their specific workloads and wait for broader access and a published API specification before planning a full migration.
Comments