AI AI Toolkit
AI Newsai-models

Google DeepMind 发布 Gemini 4 Carbon 前沿模型

Google DeepMind:Blog(RSS)2026-10-28T17:00:58.000Z

Core Highlights

Google DeepMind released Gemini 4 Carbon on October 28, 2026, calling it "the strongest model yet" with significant breakthroughs in cybersecurity and long-horizon agentic tasks. Carbon builds on Gemini 4 Argon, released two weeks earlier, and is designed for sustained agentic workflows — internal teams observed it running continuously for days on complex multi-step engineering tasks across large codebases and low-level systems, a capability that hints at a new class of autonomous software maintenance.

What It Does

Carbon delivers leaps across three security stages — discovery, verification, and remediation: SEC-bench Pro 88.2%, ExploitGym 54.8%, and CWE-bench 71.0% (all SOTA). It ranks fourth on WebDev Arena (1768 Elo, +89 points). On Gray Swan's indirect prompt-injection benchmark, attack success rate is 0.0%; Terminal-Bench 4.0 cheating rate drops to 0.0%. Those last two numbers matter as much as the raw scores, because they speak to whether the model can be trusted to act without being tricked or gaming the test.

Technical Details

On alignment, unsupported claims are down about 2x versus Argon, high-severity misalignment alerts fall nearly 80%, over-acts drop 26%, and under-serves drop 29%. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with 95% off cached input; after the intro period, input rises to $4 and output to $20. The pricing curve is clearly designed to pull early adopters in and then normalize once dependence sets in, a familiar enterprise-software playbook.

How It Compares

In Chrome team tests, Carbon produced high-quality patches over 10x better than "a large frontier model from another lab." It also tops multiple external security benchmarks and the Artificial Analysis Intelligence Index v5.0 (score 59, first place), signaling broad leadership. The implicit claim is not just that Carbon is good, but that it is decisively ahead on the dimensions enterprises care about most: security and endurance.

Industry Impact and Use Cases

For defenders, Carbon rolls "discover-verify-remediate" into an end-to-end capability and is gradually opened via the Fairwind Program, including voluntary pre-release testing with the US government. It is a reminder that frontier models demand careful governance on both offensive and defensive sides, and that the same capability that finds your vulnerabilities can also be turned against you if released without guardrails.

Practical Takeaways

For security and engineering leaders, the takeaway is that "autonomous patching" is moving from research demo to usable tool, but adoption should be staged rather than all-at-once. Start by letting Carbon propose patches in a sandbox where humans approve before anything ships, then expand the autonomy envelope only as trust accumulates from clean runs. The prompt-injection and cheating-rate zeros are the numbers to watch, because they indicate whether the model can be trusted not to be steered or to game the metrics — both of which matter enormously once the model acts on its own. Budget-wise, the introductory pricing is clearly a land-and-expand move, so model the post-intro cost before committing architectural dependence. And for defenders broadly, the same model that finds your holes can be pointed at your stack by an attacker, so the net effect of frontier security models is to raise the baseline expectation of what "defended" means.

Wider Context

More broadly, Carbon reflects where frontier labs are now competing: not just on chat benchmarks but on autonomous, security-relevant work that enterprises will actually pay for. That raises the stakes for governance, because a model good enough to patch your systems is also good enough to be weaponized if it leaks. Google's staged rollout through the Fairwind Program and government pre-release testing is a template other labs will likely copy, signaling a move toward coordinated, somewhat regulated releases of the most capable systems. For the security community, the net effect is a rising baseline: defenders get better tools, but so do attackers, so the absolute security of any given organization depends on adoption, not just model quality. The era of "release it and see" is giving way to "release it carefully and watch."

Bottom Line

Carbon's real headline is not its score but its trustworthiness under autonomy — the property enterprises will actually pay for. Adopt it in stages, watch the safety metrics, and model the post-intro cost before you depend on it.