跪拜 Guibai
← Back to the summary

Gemini 4 Carbon Tops the Charts but Stays Behind a Security Gate

Gemini 4 Carbon Release: Cybersecurity Benchmarks Are Impressive, But You Can't Use It Yet

? One-sentence summary: Google's "periodic table" army adds a new member with a fortnightly cadence — Gemini 4 Carbon tops the overall evaluation (AA Index v5.0, 59 points), sets new records on all three cybersecurity benchmarks, but is only available to a small group of security defenders at launch. Most developers can only look at the scores and wait.

Introduction: The familiar Google rhythm

Around October 14 you had just figured out how to read Gemini 4 Argon, and two weeks later its successor arrived. On October 28, Google DeepMind officially released Gemini 4 Carbon, described as "our strongest model," focused on two things: cybersecurity and long-horizon agent tasks.

Don't rush to the API console — you won't find it. This time Google is doing something new: all scores are public, all prices are public, but the model itself is not. It's first given to a batch of cybersecurity defenders in the Fairwind project (DeepMind's trusted partner program). Developers, enterprises, and consumers have to wait for "as soon as possible."

Want primary sources?

First overall, but only by a little

Look at the hardest numbers first. Carbon scored 59 points and ranked first on Artificial Analysis (an independent evaluation organization) Intelligence Index v5.0. For comparison: Claude Opus 5.5 is 58, GPT-6 Astra is 57, and the previous generation Argon is 50.

Note this is v5.0 — AA just made the leaderboard a notch harder overall, so the 59 is worth more than the number itself. But half a glass of cold water is warranted: Carbon's lead over second place is only 1 point, and it only takes first in 4 of the 9 subcategories. The "who is #1" topic is moving from "leap-frogging" to "hand-to-hand combat."

Gemini 4 Carbon official benchmark table: comparing Argon, GPT-6 Astra, Claude Opus 5.5, etc. Caption: Official full-suite benchmark comparison, Carbon column (left) leads in composite index and security benchmarks, but coding items like FrontierSWE are still beaten by GPT-6 Astra.

The coding line best illustrates the "split" in being first. On Terminal-Bench 4.0 (command-line tasks) Carbon gets 65.2%, a clear gap over Opus 5.5's 59.6%; but on FrontierSWE 2, GPT-6 Astra (65.5%) beats Carbon (61.2%); and on AA's Coding Agent Index, Opus 5.5 still presses Carbon 62 to 59. Plainly: first on overall scores, but for writing code it depends on the scenario.

Cybersecurity: this time they really put money in

Carbon's real gap is in cybersecurity. The official line splits the defense lifecycle into three steps, each with an external benchmark, and explicitly notes these were "guardrails-off evaluations" (testing with safety rails turned off; numbers are more real but also more risky):

SEC-bench Pro: vulnerability discovery benchmark comparison Caption: The 88.2% score on vulnerability discovery shows a visible gap from competitors.

ExploitGym: exploit verification benchmark Caption: 54.8% looks low, but this is inherently hard — judging "can it be exploited" requires more reasoning than "is there a hole."

CWE-bench: patch generation benchmark Caption: 71.0% means about seven in ten verified vulnerabilities generate directly usable fixes, the most labor-saving step for security teams.

There are also some vendor self-reported numbers: in internal evaluations, secure code generation success rate is more than 2x Argon; Chrome team reports a "10x better" patch quality comparison against another lab's frontier model (the opponent is anonymous, unverifiable, so take it with a grain of salt); and Wiz, the cloud security company Google acquired, observes it leading the Web pen-testing leaderboard — though Wiz's testimony is from an affiliated party, so weigh it accordingly.

Safety alignment: managing "laziness and cheating"

More interesting than benchmark numbers are the alignment details, because they directly affect whether you'll trip over issues in daily agent use:

Gray Swan indirect prompt injection: attack success rate comparison Caption: Prompt injection (embedding malicious instructions in content to hijack the model) is the top security threat in the agent era; 0.0% is a full block under official reporting.

These are all vendor self-reports, no third-party verification — but they point to an industry signal: top vendors' competitive focus is shifting from "being smarter" to "being dumber less often." When the composite score difference is only 1 point, whoever lies less, oversteps less, and gets injected less is more likely to be chosen by enterprises.

Pricing and long-horizon tasks: for "agents that run for days"

Pricing maintains Argon's intro price: $2 input / $10 output per million tokens, with a 95% discount on cached input. Compare that to OpenAI's GPT-6.1 Terra from a week ago ($1/$5, aimed at high-volume, cheap computer use). The two companies' positions are clear: Terra grabs the "high volume at low price" everyday tasks; Carbon guards "expensive but can handle big tasks" long-horizon work.

For long-horizon capability, the official scenarios are specific: internal teams run Carbon on day-scale multi-step engineering tasks — root-cause analysis in large codebases and low-level systems, cross-module bug fixes, and it can self-correct when it hits a wall. On WebDev Arena it gained 89 elo to 1768 (4th), and full-stack web development output quality is also on the official optimization list.

Community reactions: scores impressive, complaints real

The Hacker News post reached 765 points and 625 comments. A rough sample of community sentiment is about 10% positive, 25% mixed, and 65% skeptical. The skepticism isn't about the scores but three more mundane things:

A few EAP users who touched the model leaked some scattered experiences: 3D-related generation is noticeably better than Argon; frontend/UI output is still weak; overall slower but "less lazy." The sample is tiny, so treat it as reference.

Bottom line

Put Carbon on the table: what stands is the composite first place (independently re-verified by third party), the three SOTA cybersecurity benchmarks, and the unchanged price; what needs a cold seat is the vendor-reported alignment improvements and the "10x better than others" internal comparison; what's still missing is the official release date — the company only said "as soon as possible."

It means different things to different people: security folks are the first invited audience, so watch for Fairwind entry tickets; for agent and coding developers, Opus 5.5 and GPT-6 Astra remain the more practical choices in their respective scenarios; wait to test Carbon after it opens. And for those choosing based on the composite #1 — a 1-point gap, look at your scenario-specific sub-scores first.

At the pace of two weeks per generation, the next "element" is probably already on its way. More than asking how strong Carbon is, the more worth-watching question: when "can't use it at launch" becomes standard practice among top vendors, how much meaning is left in a leaderboard first?

References