Gemini 4 Carbon Tops Intelligence Index, Cyber Benchmarks

Wide dusk view of a security lab with server racks, a cable-covered bench, and a distant back-turned figure.

Gemini 4 Carbon, unveiled by Google DeepMind on October 28, 2026, is the company’s strongest frontier model yet, ranked first on the Artificial Analysis Intelligence Index v5.0 at 59 and ahead on benchmarks for finding, exploiting, and patching vulnerabilities. Starting this week, Google will make it available to a subset of cyber defenders, then gradually phase access to additional trusted partners through the Fairwind Program. A wider release waits on further guardrail iteration.

Key Takeaways

  • Gemini 4 Carbon scores 59, first on Artificial Analysis Intelligence Index v5.0.
  • It leads SEC-bench Pro at 88.2%, ExploitGym at 54.8%, and CWE-bench at 71.0%.
  • This week a subset of cyber defenders gets access before Fairwind partners.
  • Introductory pricing matches Argon: $2 per million input, $10 per million output.

What Gemini 4 Carbon Is Built to Do

In Google’s announcement, Carbon builds on Gemini 4 Argon, which launched two weeks before the October 28 post. It is aimed at harder, long-horizon engineering problems and sustained agentic workflows rather than one-shot prompts. Internal teams have seen it run for days, persisting through dead ends on multi-step jobs such as root-causing and patching subtle bugs across large codebases and low-level systems.

Google says the same push covered coding, agentic and scientific tasks, full-stack web development, and multimodal understanding. With better web-development artifacts, Carbon jumps +89 Elo to #4 (1769) on WebDev Arena.

Cybersecurity Scores Across the Defense Lifecycle

Google says Carbon achieves state-of-the-art performance across the cybersecurity defense lifecycle — vulnerability detection, validation, and remediation. That lifecycle claim rests on external scores and on internal results, not on one evaluation. The no-guardrails note applies to the named external cyber benches below.

Benchmark Task Carbon result
Artificial Analysis Intelligence Index v5.0 Aggregate intelligence 59, ranked #1
SEC-bench Pro Finding vulnerabilities 88.2%
ExploitGym Exploitation 54.8%
CWE-bench Patching 71.0%
WebDev Arena Web development 1769, ranked #4 (+89 Elo)

On those external benches, evaluated without guardrails, Carbon leads SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation, and Collinear’s CWE-bench at 71.0% for patching. Google describes SEC-bench Pro as autonomously uncovering complex flaws in real-world software, and ExploitGym as a check on which vulnerabilities are exploitable.

Separate internal numbers are Google’s own. Secure-code success showed more than a 2x uplift versus Argon. The Chrome team reported that Carbon performed more than 10 times better than another lab’s large frontier model when generating high-quality patches. Wiz, which Google calls its leading cloud security platform, observes Carbon leading on web penetration-testing benchmarks, a core capability for its Scan for Good initiative.

The spread in those scores is the practical limit. An 88.2% find rate is not an 88.2% fix rate. ExploitGym success at 54.8% is only a slim majority, so many exploit-validation checks still fail. Treat the index lead as evidence of breadth, and these cyber scores as a triage aid, not a closed-loop defense.

Access, Pricing, and Release Pace

Google is in the U.S. government’s voluntary pre-release testing process. Starting this week, Carbon goes to a subset of cyber defenders. Only the later, gradual phase extends access to additional trusted partners through the Fairwind Program. Google does not date that later phase.

The company will gather feedback and iterate on guardrails before developers, businesses, and consumers see it on product surfaces, timed only as soon as possible. Introductory pricing matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. TestingCatalog’s pre-launch report cited Logan Kilpatrick’s post that Google processed around 25 trillion tokens in Argon’s first 72 hours.

Alignment Results and Pre-Launch Tester Notes

Carbon keeps Argon’s stack: model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. Google says it further mitigated shortcut attempts in training while keeping coding performance high. The results it cites:

  • Successful cheating on Terminal-Bench 4.0 fell to 0.0%, and cheat attempts to 0.4%, while accuracy improved.
  • High-severity misalignment alerts in internal live coding dropped by almost 80%.
  • In internal testing, unsupported claims fell by about 2x compared with Argon.
  • Overreach fell 26% and under-service fell 29%.
  • Gray Swan’s Indirect Prompt Injection benchmark attack success fell from Argon’s 0.7% to 0.0%.

Google also says Carbon retrieves relevant sources, surfaces discrepancies before acting, and lists what is still required instead of fabricating evidence. Google labels the misalignment-alert drop and the unsupported-claims cut as internal. Terminal-Bench 4.0 and Gray Swan’s IPI benchmark are named evaluations outsiders can check; they cannot audit the internal alerts.

Crypto Briefing reported that Jetski testing reportedly began in early October 2026 and compared coding with Anthropic’s Claude Opus 5.5. TestingCatalog described invite-only early access and LM Arena trials, even though the Trusted Tester Agreement says unreleased models, codenames, benchmarks, and outputs should remain private until launch and shouldn’t be used in production.

Those arena notes are not the table above. Via TestingCatalog, Lumina called Carbon “incredibly slow,” roughly two to four times slower depending on the task, and said a voxel comparison had “noticeably higher quality than Argon does.” Lentils described stronger 3D work, weak frontend and UI, and a model that was “less lazy.” An unknown checkpoint means those impressions should not be merged with Google’s scored claims.

The release shape matters as much as the score. A model measured on finding and exploiting flaws reaches a subset of defenders first, with no public date for a general launch — closer to constrained frontier work such as OpenAI training frontier models without live internet than to an open product drop. Google has not said what Fairwind feedback must show before that wider release.

Notice an error?

Help us improve our content by reporting any issues you find.