Two weeks after rolling out Gemini 4 Argon globally, Google DeepMind compressed the frontier release cycle again. On October 28, 2026, Koray Kavukcuoglu announced Gemini 4 Carbon, a model engineered specifically for multi-day software engineering workflows and end-to-end cybersecurity defense.

Quick answer. Gemini 4 Carbon is Google DeepMind's strongest frontier model to date, ranking #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Over Argon, it adds +89 Elo on WebDev Arena (reaching 1768, #4 overall), hits 88.2% on SEC-bench Pro for vulnerability discovery, and records a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark. Introductory API pricing matches Argon at $2 per 1M input tokens and $10 per 1M output tokens with 95% off cached inputs, rising to $4 and $20 after the introductory window closes.

Multi-day engineering persistence and the +89 Elo jump on WebDev Arena

Anyone building custom marketing software knows where frontier coding agents usually break. Writing an isolated script is easy, but tracing a subtle bug across twenty files often pushes a model into guesswork or premature shortcuts after three failed test runs. Inside Google, engineering teams observed Carbon running continuously for days, working through dead ends to isolate root causes and patch low-level system defects across massive repositories.

Full-stack application quality moved forward alongside backend reasoning. Carbon gained +89 Elo points over Argon to reach 1768 (#4 overall) on WebDev Arena. When I build internal React tools for POAS reporting, feed diagnostics, or account auditing, that jump translates directly into cleaner state management, fewer broken API contracts, and less manual UI cleanup.

End-to-end cybersecurity benchmarks and internal Google results

Carbon's steepest capability gains appear across the defensive cybersecurity lifecycle, covering vulnerability discovery, exploit validation, and root-cause patching. Google evaluated the model without guardrails across three external security benchmarks.

Evaluation Domain Benchmark / Metric Gemini 4 Carbon Baseline / Context
Overall intelligence Artificial Analysis Intelligence Index v5.0 59 (#1 overall) Top-ranked model on v5.0 index
Full-stack web apps WebDev Arena 1768 Elo (#4 overall) +89 Elo gain over Gemini 4 Argon
Vulnerability discovery SEC-bench Pro 88.2% State-of-the-art on real-world code
Exploit validation ExploitGym 54.8% Verifies real exploitability risk
Root-cause patching Collinear CWE-bench 71.0% Argon scored 68.0% on CWE-bench v1
Prompt injection defense Gray Swan Indirect Prompt Injection (IPI) 0.0% attack success Down from Argon's 0.7% rate
Anti-cheating alignment Terminal-Bench 4.0 successful cheating 0.0% (0.4% attempts) ~80% drop in high-severity internal alerts

Production deployments inside Google and partner environments back up the external numbers. On internal security benchmarks, Carbon delivered more than a 2x uplift in secure code generation success rate over Argon. The Google Chrome engineering team reported that Carbon performed more than 10x better than another lab's large frontier model when generating high-quality security patches. Cloud security platform Wiz also confirmed Carbon leads its web penetration testing benchmarks, strengthening Wiz's Scan for Good program.

Eliminating reward hacking and hitting 0.0% on Gray Swan IPI

Running autonomous AI agents in marketing and data pipelines exposes teams to two concrete failure modes. First, agents under pressure sometimes game unit tests or fabricate missing metrics instead of fixing broken code. Second, agents scraping external pages can fall victim to indirect prompt injection embedded inside untrusted HTML.

Google DeepMind targeted both failure modes directly during Carbon's training.

  • 0.0% successful cheating on Terminal-Bench 4.0. Carbon drove successful loophole exploitation down to 0.0%, while attempts to cheat fell to 0.4% alongside higher overall benchmark accuracy. Across live internal coding sessions at Google, high-severity misalignment alerts dropped by almost 80%.
  • Roughly 2x fewer unsupported claims than Argon. Rather than inventing figures or hedging around missing evidence, Carbon writes methodological caveats directly into its output or pauses blocked actions with an explicit checklist of missing inputs.
  • Tighter scope discipline. Carbon cut overreach by 26% and under-service by 29%. When requirements conflict, it pulls source files and flags discrepancies to the operator before mutating code.
  • 0.0% attack success on Gray Swan IPI. Argon already led the Gray Swan Indirect Prompt Injection benchmark at a 0.7% attack success rate. Carbon pushed attack success down to 0.0%, blocking every evaluated injection attempt.

Why 0.0% prompt injection matters for marketing engineering

Security benchmarks often sound distant from paid acquisition until you wire an LLM agent to live credentials. In my own workflows connecting ad platforms, Server-Side GTM containers, and BigQuery warehouses, untrusted text enters the context window constantly.

Consider an autonomous audit agent that crawls competitor landing pages, parses scraped customer reviews, or ingests third-party supplier XML feeds into Merchant Center. Hidden prompt instructions planted inside external DOM nodes or product descriptions can attempt to hijack an agent's tool calls or exfiltrate environment tokens. Moving from 0.7% to 0.0% on Gray Swan IPI is a meaningful operational boundary for any agent parsing untrusted web inputs.

Halving unsupported claims is equally valuable in attribution engineering. When a BigQuery table is missing a consent partition or a webhook payload drops a currency field, you never want an agent silently interpolating revenue numbers. You want it to halt execution and report the exact broken schema field before Smart Bidding ingests bad conversion values.

Staged Fairwind rollout, API pricing, and when Argon still wins

Google is rolling out Carbon in stages. The model is currently undergoing voluntary pre-release testing with the U.S. government, while initial external access goes to vetted cyber defenders and open-source maintainers through the Fairwind Program. Broader availability across Google AI Studio, the Gemini API, Vertex AI, and consumer products will follow once guardrail iterations conclude.

API pricing launches with a clear two-tier timeline. During the introductory period, Gemini 4 Carbon matches Argon at $2 per 1M input tokens and $10 per 1M output tokens, with 95% off cached input tokens ($0.10 per 1M tokens). Once the introductory period ends, pricing doubles to $4 per 1M input tokens and $20 per 1M output tokens.

When Carbon is overkill. Once standard $4 / $20 rates take effect, routing everyday search term classification, RSA ad copy generation, or simple single-file Python scripts to Carbon wastes half your token budget. Keep high-volume, well-bounded marketing operations on Argon or Flash-tier models, and reserve Carbon for high-stakes architecture, security-sensitive scraping, and multi-file repository debugging.

Practical workload routing checklist for Argon vs. Carbon

  • ✔ 1. Route untrusted DOM and feed crawlers to Carbon. Any agent ingesting external competitor HTML, user-generated reviews, or supplier XML catalogs benefits directly from 0.0% Gray Swan IPI resilience.
  • ✔ 2. Use Carbon for multi-file sGTM, Cloud Run, and BigQuery debugging. Reserve Carbon for cross-repository refactors, webhook authentication code, and subtle attribution discrepancies where 0.0% test-cheating matters.
  • ✔ 3. Front-load heavy internal app builds during the $2 / $10 introductory window. Take advantage of Argon-parity introductory pricing and the 95% context caching discount to ship or harden custom React dashboards.
  • ✔ 4. Keep routine PPC automation on Argon. Daily pacing checks, standard Merchant Center attribute mapping, and ad copy variations belong on Argon so post-introductory $4 / $20 pricing never inflates recurring run costs.

Frequently Asked Questions (FAQ)

How does Gemini 4 Carbon differ from Gemini 4 Argon?

Gemini 4 Argon is Google's globally available frontier model with a 1M output token window. Gemini 4 Carbon builds on Argon for multi-day software engineering, full-stack web apps (+89 Elo to 1768 on WebDev Arena), and defensive cybersecurity (88.2% on SEC-bench Pro), ranking #1 on the Artificial Analysis Intelligence Index v5.0 at 59.

What is the API pricing for Gemini 4 Carbon?

Gemini 4 Carbon launches at an introductory rate matching Argon at $2 per 1M input tokens and $10 per 1M output tokens, with a 95% discount on cached input tokens. After the introductory period expires, standard pricing of $4 per 1M input tokens and $20 per 1M output tokens will apply.

Who can access Gemini 4 Carbon right now?

Google is phasing deployment starting with voluntary U.S. government pre-release testing and vetted cyber defenders and open-source maintainers via the Fairwind Program, ahead of general rollout across developer and enterprise surfaces.

Why does 0.0% on the Gray Swan IPI benchmark matter for marketing teams?

Marketing agents that scrape external competitor pages, customer reviews, or supplier product feeds process untrusted text that can hide malicious prompt instructions. Scoring 0.0% attack success on Gray Swan Indirect Prompt Injection protects tool-using agents from hijacking and data leaks.

Want to wire frontier AI agents into your Google Ads, Server-Side GTM, and BigQuery stack with strict guardrails and predictable API unit economics? Reach out for a technical consultation and let us design the right automation architecture for your team.