POLICY · MODELS · #2739
Google DeepMind announces Gemini 4 Carbon, frontier model focused on cybersecurity and long-horizon tasks
Google DeepMind announced Gemini 4 Carbon, a new 'frontier' model in the Gemini 4 family designed for long-horizon agentic workflows and improved cybersecurity capabilities. DeepMind reports Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 (59) and posts leading benchmark results (SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%), will be rolled out first to a subset of cyber defenders via its Fairwind Program while participating in U.S. government voluntary pre-release testing, and is priced at $2 per million input tokens and $10 per million output tokens (cached input tokens at a 95% discount).
KEY POINTS
- Google DeepMind announced Gemini 4 Carbon, a new 'frontier' model in the Gemini 4 family designed for long-horizon agentic workflows and improved cybersecurity capabilities.
- DeepMind reports Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 (59) and posts leading benchmark results (SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%), will be rolled out first to a subset of cyber defenders via its Fairwind Program while participating in U.S.
- government voluntary pre-release testing, and is priced at $2 per million input tokens and $10 per million output tokens (cached input tokens at a 95% discount).
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Google DeepMind announced Gemini 4 Carbon, a new 'frontier' model in the Gemini 4 family described as focused on state-of-the-art cybersecurity capabilities and long-horizon (sustained agentic) workflows. DeepMind says Carbon achieved top placements on internal and external indices and benchmarks, will be rolled out first to a subset of cyber defenders via the Fairwind Program while participating in U.S. government voluntary pre-release testing, and is priced at $2 per million input tokens and $10 per million output tokens (with cached input tokens at a 95% discount).
WHY NOW
This matters now because DeepMind has publicly introduced a model they present as materially stronger on cybersecurity-relevant benchmarks and long-horizon agentic tasks, and they are starting a controlled, near-term rollout to cyber defenders and government testing channels. Those two facts change the immediate availability and claimed capability profile compared with prior Gemini releases.
WHO IS AFFECTED
Directly affected initially: the subset of cyber defenders and trusted partners participating in DeepMind’s Fairwind Program and the U.S. government's voluntary pre-release testing. DeepMind also states implications for developers, businesses, and consumers once access is expanded (future/conditional).
CONFIRMED
All confirmed, source-backed facts from the supplied excerpts: - Google DeepMind announced Gemini 4 Carbon as their next 'frontier' model. - DeepMind reports Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 with score 59. - Reported benchmark results: SEC-bench Pro 88.2% (vulnerability finding), ExploitGym 54.8% (exploitation), CWE-bench 71.0% (patching). - DeepMind reports Carbon drives uplift in generating secure code and shows reduced misalignment and fabrication rates in internal live coding environments. - DeepMind states Carbon is designed for sustained agentic workflows and that internal teams have seen it run for days to tackle multi-step engineering tasks. - DeepMind says they are engaged in the U.S. government’s voluntary pre-release model testing process. - DeepMind says Carbon will be made available starting this week to a subset of cyber defenders and will be phased to additional trusted partners via the Fairwind Program. - Introductory pricing is stated as $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. - DeepMind notes Carbon follows Gemini 4 Argon, launched two weeks earlier, and reports Carbon’s WebDev Arena change (+89 elo to #4, 1769).
UNCERTAIN
Key uncertainties and missing or conflicting evidence (explicitly labeled): - Independent validation: the supplied sources are DeepMind/Google announcements; independent or third-party replication of the reported benchmark scores and rankings is not provided in the excerpts. (MISSING) - Scope and participants: the excerpts do not specify which organizations or how many are included in the 'subset of cyber defenders' or the composition/scope of Fairwind Program participants. (MISSING) - Pre-release testing details: no details on the scope, methodology, timeline, or outcomes of the U.S. government voluntary pre-release testing are provided. (MISSING) - Guardrails and limitations: the excerpts state DeepMind is iterating on guardrails but give no technical details on safeguards, failure modes, or specific operational limits. (MISSING) - Real-world performance and safety: internal reports of 'running for days' and reduced misalignment/fabrication are company statements; external evidence of real-world performance, failure rates, or misuse incidents is not provided. (MISSING) - Reliability of multi-step agentic behavior at scale: claims of sustained agentic workflows are internal; the excerpts do not quantify failure modes, costs, or environmental constraints. (MISSING) - No conflicting evidence is present in the supplied excerpts, but absence of external corroboration remains a major gap. (MISSING)
WHAT TO WATCH
Concrete observable signals to watch next (based on supplied sources): - Publication or disclosure of results/details from the U.S. government's voluntary pre-release testing and any public summary or findings. - Independent third-party benchmark replications or papers reporting SEC-bench Pro, ExploitGym, and CWE-bench results for Carbon. - Public case studies, reports, or technical write-ups from Fairwind Program participants or other early-access cyber defenders describing real-world performance, limitations, and guardrails. - DeepMind/Google updates on broader product availability (dates, supported product surfaces) or changes to pricing/terms. - Any DeepMind-published technical details on safeguards, failure modes, or operational constraints for Carbon. - External security researcher/red-team reports or disclosed incidents linked to model-assisted vulnerability finding/exploitation/patching workflows.
WHY IT MATTERS
This matters because Carbon claims substantial advances in automated vulnerability finding, exploitation validation, and patching plus sustained multi-step agentic capabilities, which could materially change cyberdefense tooling and risk profiles.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Google DeepMind announced Gemini 4 Carbon as its next “frontier” model and described it as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities.
SUPPORTEDChecked 2026-10-28 · 2 supporting
The Google blog post introduces Gemini 4 Carbon as the next frontier model and calls it the strongest model with state-of-the-art cybersecurity and long-horizon capabilities.
SUPPORTS Google DeepMindMEDIA · 2026-10-28The DeepMind blog post repeats the same description of Carbon as the strongest frontier model with cybersecurity and long-horizon capabilities.
DeepMind reports Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
SUPPORTEDChecked 2026-10-28 · 2 supporting
DeepMind reports leading benchmark results for Carbon: SEC-bench Pro 88.2% (finding vulnerabilities), ExploitGym 54.8% (exploitation), and CWE-bench 71.0% (patching).
SUPPORTEDChecked 2026-10-28 · 2 supporting
DeepMind says Carbon will be made available first to a subset of cyber defenders via its Fairwind Program and that it is participating in the U.S. government’s voluntary pre-release model testing process.
SUPPORTEDChecked 2026-10-28 · 2 supporting
The Google post states Carbon will be made available to a subset of cyber defenders through the Fairwind Program and mentions engagement with the U.S. government’s voluntary pre-release testing.
SUPPORTS Google DeepMindMEDIA · 2026-10-28The DeepMind post repeats the rollout plan and mention of U.S. government voluntary testing.
SOURCES & TIMELINE
4Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities. SVP, Google DeepMind and Chief AI Architect, Google Your browser does not support the audio element. Today, we’re announcing the next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon ranks #1 on the recently updated v5.0 of Artifici…
Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities. SVP, Google DeepMind and Chief AI Architect, Google Your browser does not support the audio element. Today, we’re announcing the next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon ranks #1 on the recently updated v5.0 of Artifici…
On today’s Q3 2026 earnings call, Google and Alphabet CEO Sundar Pichai recapped a strong quarter, with our full-stack approach driving momentum across the business. Below is an edited transcript of his remarks. Q3 was a strong quarter for Alphabet. Our deep, full-stack approach is driving momentum across the business, with many opportunities ahead. Alphabet revenues grew 25% year-over-year. We’re now at a half-tri…
Just two weeks after Gemini 4 Argon, Google DeepMind released Gemini 4 Carbon, its next frontier model. Carbon tops the Artificial Analysis Intelligence Index v5 with 59 points, up from Argon's 50. Google says Carbon is built for multi-day, autonomous development workflows. It's closing the gap fast in Artificial Analysis's Coding Agent Index (59 points vs. Argon's 50) but still trails Claude Opus 5.5 (62 points). …
Gemini 4 Carbon is outperforming the top models from OpenAI and Anthropic on some benchmarks. Google has released its latest frontier model, Gemini 4 Carbon, offering “significant gains” in coding, agentic tasks, and cybersecurity. In an announcement on Wednesday , Google highlights Carbon’s top spot on the Artificial Analysis Intelligence Index (v5.0) benchmark, with Claude Opus 5.5 and GPT-6 Astra trailing behind.…