ModelsArchitectures & capability
Google launches Gemini 4 Carbon to cyber defenders first as it tops Artificial Analysis Intelligence Index
Google's Gemini 4 Carbon, released two weeks after Gemini 4 Argon, scores 59 to lead Artificial Analysis's Intelligence Index v5 by one point. Its strength at finding and exploiting software flaws means access starts with Fairwind Program cyber defenders, with no wider date.
From the The Verge · AI article of October 28, 2026 and cited sources: Carbon scored 59 on Artificial Analysis's Intelligence Index v5, one point above Claude Opus 5.5. [8] [13]
SOURCES[8] Artificial Analysis: Carbon #1 on Index v5 artificialanalysis.ai[13] Cornford and Cross: One-point lead explained cornfordandcross.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Carbon scored 59 on Artificial Analysis's Intelligence Index v5, one point above Claude Opus 5.5. [8] [13]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 28, 2026 article published
- Covered
- Oct 29 edition
- Evidence
- 14 sources across 13 sites
- 1Google launches Gemini 4 Carbon with a ‘significant leap’…theverge.com
- 2Gemini 4 Carbon: Our next frontier modeldeepmind.google
- 3Gemini 4 Carbon is INSANE! 🤯reddit.com
- 4Google: Fairwind Programblog.google
- 5noze: Gemini 4 Argon and Fairwindnoze.it
- 6SmartScope: Argon access and pricingsmartscope.blog
Reach: not measured yet
Carbon scored 59 on Artificial Analysis's Intelligence Index v5, one point above Claude Opus 5.5. The lead is narrow. Carbon won 4 of the 9 evaluations and lost the rest, including coding and office-style work. Opus 5.5 still leads the Coding Agent Index, 62 to Carbon's 59. [8] [13]
Google says Carbon can exploit vulnerabilities, and gives that as the reason it is going first to selected defenders and trusted Fairwind partners. Developers, businesses and consumers have no release date. The introductory price matches Argon's at $2/$10 per million input/output tokens and later rises to $4/$20. [7] [14] [4]
Google reports headline security scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These are vendor figures that no third party has replicated, as are claims of twice as much secure code as Argon and multi-day autonomous runs. [7] [10]
Carbon uses fewer tokens, which lowers its cost per task. It averaged about 123k output tokens per index task against 173k for Opus 5.5, for a cost of $7.76 per task versus $10.04. Its win on the AA-Omniscience knowledge test came from caution rather than more knowledge: lower raw accuracy, but an 11% hallucination rate. One outlet says that rate is up from Argon's 4%. [8] [13] [11]
Artificial Analysis ranks Carbon first by a narrow margin and finds it cheaper per task than Opus 5.5.
Read the full assessment
Implication: security teams outside Fairwind can't use it yet, and coding-agent buyers should test on their own workloads, since Opus 5.5 still leads that sub-index.
Sources are gathered; writing the dossier now.
Executive brief
Google's newest frontier model, Gemini 4 Carbon, took first place on the Artificial Analysis Intelligence Index v5 with 59 points, one point ahead of Claude Opus 5.5. Most people can't use it yet. Google is releasing it first to cyber defenders in its Fairwind Program because the model is strong at finding and exploiting software vulnerabilities. Google reports 88.2% on SEC-bench Pro. Carbon came out only two weeks after Gemini 4 Argon and keeps Argon's introductory price. An independent evaluator confirms the overall ranking. Most of the cybersecurity and safety figures come from Google alone.
What changed and event timeline
Fairwind Program launches
Google opened a restricted cyber-defense program built on Gemini 3.8 Flash Cyber and the CodeMender patching harness. Google says it has more than 650 partners, including CrowdStrike, Palo Alto Networks and Wiz ().
Gemini 4 Argon announced
Some selected partners reportedly received it "without cyber guardrails," with phishing-resistant MFA required ().
Argon generally available
The Gemini API listed Argon as stable at $2/$10 per million tokens until Nov 30, then $4/$20 ().
Carbon announced
Google called Carbon its strongest model for cybersecurity and long tasks. Access starts with Fairwind cyber defenders (;).
Independent index ranks Carbon #1
Artificial Analysis scored Carbon 59 and Claude Opus 5.5 second. Carbon leads 4 of the 9 index evaluations ().
Capabilities and access
- Model: Gemini 4 Carbon, the successor to Gemini 4 Argon (Engadget).
- Access: Selected cyber defenders first, then trusted Fairwind partners. Developers, businesses and consumers get it "as soon as possible," with no date.
- Price: $2 per million input tokens and $10 per million output tokens to start, with a 95% discount on cached input.
Read the full section
- Model: Gemini 4 Carbon, the successor to Gemini 4 Argon (Engadget).
- Access: Selected cyber defenders first, then trusted Fairwind partners. Developers, businesses and consumers get it "as soon as possible," with no date. Google says it takes part in the US government's voluntary pre-release testing (Google).
- Price: $2 per million input tokens and $10 per million output tokens to start, with a 95% discount on cached input. The standard rate later is $4/$20 (Google).
Technical analysis for researchers and developers
- Architecture: Not published. The announcement includes no model card link (Google).
- Index methodology: Index v5 adds 2 new benchmarks and upgrades 3.
- Efficiency: Carbon used about 123k output tokens per index task, 29% fewer than Opus 5.5's 173k.
Read the full section
- Architecture: Not published. The announcement includes no model card link (Google).
- Index methodology: Index v5 adds 2 new benchmarks and upgrades 3. For the first time, most of its datasets are private, which limits outside replication (Artificial Analysis).
- Efficiency: Carbon used about 123k output tokens per index task, 29% fewer than Opus 5.5's 173k. A task costs $7.76, compared with $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra (Artificial Analysis).
- Safety stack: Google names training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring, but gives no implementation details.
Claims and evidence
- Index rank (independent): 59, #1 overall. Opus 5.5 still leads the Coding Agent Index at 62, 3 points ahead of Carbon (Artificial Analysis).
- Security scores (Google-reported): SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%. These are the headline numbers, and no third party has replicated them (Google).
- Internal-only claims: Twice as much secure code as Argon, Chrome patches "10x better" than an unnamed rival, and multi-day autonomous runs.
Read the full section
- Index rank (independent): 59, #1 overall. Opus 5.5 still leads the Coding Agent Index at 62, 3 points ahead of Carbon (Artificial Analysis).
- Security scores (Google-reported): SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%. These are the headline numbers, and no third party has replicated them (Google).
- Internal-only claims: Twice as much secure code as Argon, Chrome patches "10x better" than an unnamed rival, and multi-day autonomous runs. None of these has been independently verified (Interesting Engineering).
- Prompt injection: 0.0% attack success on Gray Swan's indirect prompt-injection benchmark, compared with 0.7% for Argon (The Decoder).
Context and prior work
Fairwind started with a specialized model, Gemini 3.8 Flash Cyber, which Google says succeeded more than 70% of the time on an internal vulnerability benchmark covering 20 languages (AI Weekly). Argon then used the same defenders-first rollout. Carbon follows that pattern, and Artificial Analysis notes the parallel directly. On the index, Carbon rose from Argon's 50 to 59 in two weeks (The Decoder).
Limitations, safety and contested findings
- Thin lead: The one-point lead comes from winning 4 of 9 tests and losing the other 5, including coding and office-style work.
- Knowledge-test win: Carbon's AA-Omniscience win comes from caution rather than more knowledge.
- Hallucination conflict: The Decoder reports that hallucination rose from Argon's 4% to 11% (The Decoder).
Read the full section
- Thin lead: The one-point lead comes from winning 4 of 9 tests and losing the other 5, including coding and office-style work. It is unclear whether the ranking holds across repeated runs (Cornford and Cross).
- Knowledge-test win: Carbon's AA-Omniscience win comes from caution rather than more knowledge. Its plain accuracy was lower (58%), but its hallucination rate was 11%, compared with about 50% for rivals (Cornford and Cross).
- Hallucination conflict: The Decoder reports that hallucination rose from Argon's 4% to 11% (The Decoder). That sits uneasily with Google's claim of half as many unsupported claims.
- Exploit risk: Google cites the model's ability to exploit vulnerabilities as the reason for the limited rollout (Crypto Briefing).
Business and practitioner implications
- Security teams: Fairwind members get first access. Everyone else has no timeline.
- Coding agents: On independent numbers, Opus 5.5 still leads the Coding Agent Index, so test Carbon on your own workloads before switching.
- Plan around the later $4/$20 rate, and note that per-task cost depends on how many tokens a model uses, not just its list price.
Read the full section
- Security teams: Fairwind members get first access. Everyone else has no timeline.
- Coding agents: On independent numbers, Opus 5.5 still leads the Coding Agent Index, so test Carbon on your own workloads before switching.
- Budgets: The introductory price matches Argon's. Plan around the later $4/$20 rate, and note that per-task cost depends on how many tokens a model uses, not just its list price.
- Procurement: Treat the security benchmark scores as Google's claims until a third party evaluates them.
The source trail.
Sources (14)
Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.comGemini 4 Carbon: Our next frontier model
Related coverage; assess separately
deepmind.googleGemini 4 Carbon is INSANE! 🤯
Related coverage; assess separately
reddit.com