Google DeepMind releases Gemini 4 Carbon, focused on cybersecurity defense and coding
Google DeepMind has launched Gemini 4 Carbon, calling it a frontier model for long-horizon agentic coding and cybersecurity defense, ranking first with 59 points on the Artificial Analysis Intelligence Index v5.0. The model achieved 96.2% on the GraphWalks long-context test, and 88.2%, 54.8%, and 71.0% on SEC-bench Pro, ExploitGym, and CWE-bench, respectively, but on OSWorld-2.1 it trailed Claude Opus 5.5's 87.2% with 81.5%. Carbon is currently available only to vetted cybersecurity defenders, with introductory pricing of $2 per million input tokens and $10 per million output tokens, later $4 and $20; the timeline for opening it to developers and the public has not been announced.
Why read this
SelectedAI score 85
The original text lists Gemini 4 Carbon's specific scores on cybersecurity and coding benchmarks and comparisons with competitors, and also states that it is currently open only to vetted defenders.
Topics
Body
Add as a preferredsource on Google
Google DeepMind has introduced Gemini 4 Carbon, its strongest model to date. Gemini 4 Carbon is a frontier model from Google built for long-horizon agentic coding and cybersecurity defense. It ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. For developers, the catch is access. Carbon starts with a vetted group of cyber defenders, not the public API.
TL;DR
Size: Parameter count and context window not disclosed. Its predecessor, Gemini 4 Argon, has a 1,048,576-token context.
Runs on: Google-hosted only. No open weights, no self-hosting, no quantized variants.
Performance: Leads 15 of 17 published benchmarks versus Argon, GPT-6 Astra, Opus 5.5 and Fable 5.1.
Best: 96.2% BFS F1 on GraphWalks (256K to 1M tokens), versus 84.2% for Argon.
Worst: 81.5% on OSWorld-2.1, trailing Claude Opus 5.5 at 87.2%.
Bottom line: Best: state-of-the-art vulnerability finding, exploitation and patching. Worst: you probably cannot call it yet.
What is Gemini 4 Carbon?
Carbon is the second Gemini 4 model, following Gemini 4 Argon. Google team states that the model targets hard, long-running engineering work. Internal teams saw Carbon run for days on single tasks. Examples include root-causing and patching subtle bugs across large codebases and low-level systems.
TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored
Google also tuned Carbon for full-stack web development and multimodal understanding. On WebDev Arena, Carbon gained 89 Elo points to reach #4 at 1769.
How does Gemini 4 Carbon perform on benchmarks?
Google published a 17-row comparison in its launch post. Carbon was run at the highest thinking setting. Most rival numbers are self-reported by each provider, per Google’s evaluation methodology.
Agentic coding: Carbon scores 65.2% on Terminal-Bench 4.0 (AA), ahead of Opus 5.5 at 59.6%. It hits 9.6% on SWE-sweep, narrowly ahead of GPT-6 Astra at 9.5%. On FrontierSWE, GPT-6 Astra leads at 65.5% versus Carbon’s 61.2%.
Computer use and knowledge work: Carbon reaches 77.4% on Agents’ Last Exam v1.1. On Vending-Bench 2, it ends with a $16,484 bank balance, the highest listed.
Long context and science: GraphWalks is the widest gap: 96.2% versus 71.8% for GPT-6 Astra. Carbon also posts 80.0% on RiemannBench.
Why does Carbon relevant for cybersecurity?
Google frames Carbon around the full defense lifecycle: find, validate, then fix. Cyber benchmarks were evaluated without guardrails.
Finding: 88.2% on SEC-bench Pro, versus 85.4% for GPT-6 Astra.
Exploiting: 54.8% on ExploitGym, versus 42.4% for GPT-6 Astra and 34.5% for Opus 5.5.
Patching: 71.0% on Collinear’s CWE-bench, versus 68.0% for Argon and GPT-6 Astra.
Google computed SEC-bench Pro and ExploitGym scores itself, using its Antigravity harness. ExploitGym ran with mitigations off. Internally, Carbon showed more than 2x higher success at generating secure code than Argon. Google team states that the Chrome team found Carbon over 10x better than another lab’s model at high-quality patches. Wiz reports Carbon leading its web penetration testing benchmarks.
What safety changes ship with Carbon?
Carbon keeps Argon’s layered stack: training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Google reports several alignment gains:
Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheating attempts at 0.4%.
High-severity misalignment alerts in live internal coding dropped by almost 80%.
Unsupported claims fell roughly 2x versus Argon.
Overreach fell 26% and under-service fell 29%.
Attack success on Gray Swan’s Indirect Prompt Injection benchmark is 0.0%, down from Argon’s 0.7%.
These are Google’s internal or self-run figures. Independent replication is not yet available.
How can developers access Gemini 4 Carbon?
Not yet, for most. Google is participating in the U.S. government’s voluntary pre-release testing process. This week, a subset of cyber defenders gets access. More trusted partners follow through the Fairwind Program, which prioritizes governments, critical infrastructure operators and core tech platforms. Developers, businesses and consumers get Carbon later, with no date given.
Gemini 4 Carbon vs competitors
FeatureGemini 4 CarbonGemini 4 ArgonGPT-6 AstraClaude Opus 5.5Claude Fable 5.1
ParametersNot disclosedNot disclosedNot disclosedNot disclosedNot disclosed
Context windowNot disclosed1,048,5761,050,0001M1M
Max outputNot disclosed65,536 (1,048,576 long decode)128,000128K128K
Price in / out (per 1M)$2 / $10 intro, then $4 / $20$2 / $10 intro (Google)$10 / $50$4 / $20 (pricing)$10 / $50 (pricing)
AA Index v5.05950575849
SEC-bench Pro88.2%83.4%85.4%Not disclosedNot disclosed
ExploitGym54.8%42.0%42.4%34.5%Not disclosed
OSWorld-2.181.5%74.7%80.9%87.2%76.8%
FrontierSWE61.2%55.0%65.5%62.3%56.3%
Benchmark rows: Google launch post; rival scores mostly self-reported per methodology.
Key Takeaways
Gemini 4 Carbon ranks #1 on Artificial Analysis Index v5.0 at 59.
It leads SEC-bench Pro (88.2%), ExploitGym (54.8%) and CWE-bench (71.0%).
GraphWalks long context hits 96.2%, its widest margin.
Claude Opus 5.5 still wins OSWorld-2.1; GPT-6 Astra wins FrontierSWE.
Access starts with cyber defenders at $2/$10 per million tokens.
Asif RazzaqWebsite | + postsBio ⮌
Asif RazzaqBaseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens
Asif RazzaqPrime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass
Asif RazzaqNVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters
$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored
RELATED ARTICLESMORE FROM AUTHOR
Baseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens
Prime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass
NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters