AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Gemini 4 Carbon: Our next frontier model

Collected Oct 28, 2026

Google DeepMind announced Gemini 4 Carbon on its blog, describing it as the next frontier model and its strongest to date, with what it calls state-of-the-art cybersecurity and long-horizon capabilities. The announcement is attributed to Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. Carbon arrives two weeks after Gemini 4 Argon, which the post frames as evidence of how quickly the lab is moving the frontier.

The headline benchmark claim is a first-place score of 59 on the recently updated v5.0 of the Artificial Analysis Intelligence Index, an independent aggregator that combines several evaluations into a single number. Alongside that, DeepMind reports a cluster of cybersecurity results: 88.2% on SEC-bench Pro for finding vulnerabilities in real-world software, 54.8% on ExploitGym for validating whether a flaw is actually exploitable, and 71.0% on Collinear's CWE-bench for producing patches that fix root causes. It also reports gains in coding and agentic tasks more broadly.

The cybersecurity lifecycle framing is worth unpacking. Finding, prioritizing and fixing vulnerabilities are distinct jobs: detection without validation produces noise, and validation without remediation leaves risk open. Carbon is positioned across all three, which is why the three benchmark scores are presented together. DeepMind says these external results were produced without guardrails applied, a common way to measure raw capability before deployment restrictions are layered on.

On coding, Carbon scores 1769 on WebDev Arena, a jump of 89 elo points that places it fourth on that leaderboard. Elo here is a relative ranking derived from head-to-head comparisons, so a gain of that size at the top of a crowded board is substantial rather than incremental. DeepMind says the model will be offered at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input billed at 95% off the input rate. That cache discount matters for agentic workloads, which tend to resend large context repeatedly.

Long-horizon work is the other pillar. DeepMind says its internal teams have watched Carbon run for days on multi-step engineering problems, persisting through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems. Internal benchmarks also showed more than double the success rate versus Argon at generating secure code. Two partner claims accompany this: the Chrome team reportedly found Carbon more than 10x better than another lab's large frontier model at producing high-quality patches, and Wiz, a cloud security platform, is cited as observing Carbon leading on web penetration testing benchmarks relevant to its Scan for Good initiative.

A notable section covers alignment and honesty. DeepMind says it trained Carbon to handle complex or high-pressure tasks appropriately, and reports that successful cheating attempts on Terminal-Bench 4.0 fell to 0.0%, with attempts themselves down to 0.4%, while overall accuracy improved. Internal live coding environments showed high-severity misalignment alerts dropping by almost 80%. Carbon is said to cut unsupported claims roughly in half relative to Argon, to surface discrepancies to the user instead of guessing when requirements are ambiguous, and to reduce overreach by 26% and under-service by 29%. On prompt-injection resistance, it reports 0.0% attack success on Gray Swan's Indirect Prompt Injection benchmark, against Argon's previously leading 0.7%. The safety stack mirrors Argon's: training-time measures, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring.

Rollout is deliberately staged. DeepMind says it is participating in the U.S. government's voluntary pre-release testing process. Starting this week, Carbon goes to a subset of cyber defenders, with access widening gradually through the Fairwind Program to additional trusted partners. DeepMind says it is collecting feedback and refining guardrails before bringing the model to developers, businesses and consumers through its product surfaces, without giving a date.

My reading is that the sequencing is the story as much as the numbers. A model trained for offensive-adjacent cybersecurity work is being handed first to defenders under supervision, which limits early exposure while the lab gathers real-world signal. The trade-off is availability: teams outside that cohort, including most developers, have no timeline yet, and benchmark scores obtained without guardrails may not translate directly into what shipped APIs will permit. The pricing also applies to an introductory period, so cost planning should assume it can change.

Why it matters: Security and platform teams evaluating automated vulnerability detection, triage and patching now have a concrete benchmark profile to compare against, but the practical impact waits on general availability. Developers building agentic coding workflows should watch whether the long-horizon and honesty improvements survive guardrails once the model reaches public APIs, since that is where multi-day autonomous runs and reduced fabrication would change day-to-day engineering work; that dependency is my inference, not a stated commitment. The staged defender-first release also hints at how frontier labs may sequence dual-use capability in future launches.

Read at Google DeepMind

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt