It's the first time in the history of the world.
Google has just unveiled a brand-new King of Fighters model.Gemini 4 CarbonA breath of fresh air.World number oneGet it back!
In Google's official list of 14 core benchmarks, carbon is down.12 first。
In the latest v5.0 global smart index by the authoritative third-party Artificial Analysis, Carbon topped the global ranking with 59 points, fully pressurized Opus 5.5 and GPT-6 Astra!
Arena Community Blind Synchronized List, which takes one step at a timeWorld No. 1 with a total of 1528 texts、Agent is No. 2 in the world。
The big Vs exclaimed: "Google has achieved the biggest blockbuster of the year!"
Some people say that maybe it's because Google has already implemented the RSI that carbon is so great.
Unfortunately, the Gemini 4 Carbon is not yet open to everyone and only allows the use of "National Security Defenders".
Google's internal team has reportedly seen carbon in a complex codebase.It's been going on for days.I've been dying of bugs, patching myself.
In addition, in the Chrome browser security team's observations, it generated high-quality bug patches, directly larger than another front-end model.More than 10 times the force!
Interestingly, Carbon will maintain the same introductory price as Argon:
For every 1 million tokens entered: $2
$10.00 for every 1 million tokens.
Compared to the GPT-6 Astra and Opus 5.5, this is a "white price"!
Now, the internet is full of amazing demonstration demos.
According to some, it was a surprise to see Google return to the top of the list.
It has been said that carbon is more beautiful in nature than all the 3D cities produced by the big models.
Some reviewers say the black hole made by carbon is shocking.
The following is a summary of the assessments.
Lumina allows it to write a one-off with full optical shadow interaction.3D X-ray wind turretThe scene.
The same suggestion is placed in four-dimensional contrast, and the gap is obvious.
In the upper left corner of the tower, the body of the Argonne tower is thin and rough, while the upper right corner of the tower is of carbon.Multilevel weight ratio, water quality and garden layoutHe jumped to the top.The completeness is directly equivalent to Claude Opus 5.5.。
to the Scripture.SVG is riding a bicycleTesting, Carbon takes the front-end engineering details to new heights.
It not only paints a pavilion full of details, but also provides support.20 to 100 RPM without a rate of changeIt's a great way to learn how to play the game.
In the bottom right corner."CAD engineering blueprint inspection"Button, the whole picture will be switched to instantaneousDeep blue X-ray mechanical bone anatomyIt is also possible to calculate the ratio of the torque to the torque of the torque.
The community developer Raphi-2Code has also unveiled a demonstration of the Dragonfly open-screen dynamics.
Many of the observers pointed out in the same way:Carbon thinking takes two to four times as long as Argon, but it's no longer rigged.。
Peter Gostev, the head of the Arena Ability Assessment, was even more direct:"Carbon is much stronger than Argon, and I really wish they'd just jumped right over that version of Argon".
In the past, Google has long been praised by both OpenAI and Anthropic.
But the birth of Gemini 4 Carbon directly shattered the voice of questioning.
Carbon was ranked number one in 17 tests, with Madness 15 ranking first.
First of all, it demonstrates the suppression of integrated intelligence.
In the AA Index v5.0 test, Carbon scored 59 points ahead of Opus 5.5 and GPT-6 Astra. In the Vals Index v2.2 it also took first place with 70.4%.
Second, it demonstrates an unrivalled combination of long text and logic.
In the GraphWalks (256k to over 1M below) test, Carbon ran a staggering 96.2%! In contrast, GPT-6 Astra was only 71.8%, and Opus 5.5 was only 66.8%.
This means that when it comes to handling bulk information and finding logical correlations across documents, carbon is almost zero error.
In addition, it is a master of science.
In the Riemann Bench test in science and mathematics, Carbon led by 80.0%, far ahead of its competitors.
Terminal-Bench 4.0 jumped to 65.2% in terminal operations and multi-step programming, while the R&D rating of Terminal-Bench-Science jumped to 64.8%, which also doubled as the world's number one.
In Vending-Bench 2, an autonomous long-cycle simulation, Carbon scored 1.8 times as much as Claude Opus 5.5.
Of course, the battle for specific engineering tasks is still ongoing, for example in the Frontier SWE 2 and OSWorld-2.1 offline subset, where the GPT-6 Astra and Claude Opus 5.5 each retain an advantage.
On the Arena Agent list, Carbon jumped directly to the top of the list.World No. 2The Sonnet 5.5 Max and the GPT-6 Astra Max are all behind you.
Carbon is used to measure the ability of a model to hear people's voices when a user changes their needs or objects in the meantime.+18.6% high in first place overall。
The cost of a single mission is $1.80, compared to the first Opus 5.5 Max.That's down 72% overall.。
In the Text Arena text overall list, Carbon is based on1528 points to the top of the world；
On the WebDev Arena rankings, Carbon jumped 89 Elo points to the top four globally.
If running scores are just an appetizer, then Gemini 4 Carbon's performance on cybersecurity and ultra-long-cycle proxy missions is what really shocks the industry.
In a very special description on Google's official blog: "Our internal team saw Carbon running continuously for days, finding new paths to solve the most hidden bugs in large code libraries and underlying systems after a dead end".
This is a highly autonomous digital employee.
In its cybersecurity offensive testing, Google unveiled three highly representative data that perfectly match the three core steps of "searching, testing, fixing" security teams to address vulnerabilities.
It's the first one.Looking for a vulnerability: SEC-bench Pro cuts 88.2%
It's the first time that a JavaScript engine has been able to capture a real high-risk vulnerability.SEC-bench ProIn the test, carbon was obtained.88.2% ranked firstIt is estimated that the GPT-6 Astra was compressed by 85.4 percent and Argon by 83.4 percent.
The second,ExploitGym takes 54.8% of the new high
The Linux kernel and the Chrome V8 engineExploitGymIt requires the model to write down the risks of using script verification in a real environment.
In this hardcore Red Team test, Claude Opus 5.5 was only 34.5%, GPT-6 Astra was 42.4%, and Carbon rose to 54.8%!
The third line.Fixed vulnerability: CWE-bench to 71.0%
However, the security issues seem to have been resolved and people are not allowed in.CWE-benchIt's about stopping this phony fix: in 120 real-world vulnerabilities that cover the 6 major languages and 73 vulnerabilities,The original attack must fail and the original functionality test must pass 100%.。
Under this rigorous set of standards, Carbon is attempting to pass rates for the first time.Pass@4 up to 81%)Take the first one again.
In Google's internal output line testing, Carbon generated a higher success rate for security code than Argon.It's more than doubled.The core team at Chrome tested more feedback, which generated high-quality bug patches than larger models from other leading edge labs.More than 10 times the force.。
That's why it's called "the strongest cybersecurity model of all time".
Extreme alignment and security: not only genius but also good AIImagine if Carbon's ability to "find vulnerabilities and write attack code" fell into the hands of hackers, it would be a disaster.
Google is well aware of this and is building carbon as a fuel source."The most accurate and obedient model of all time".It has been added an unprecedented "Golden Clock" to it.
It's a 0% success rate!
Gray Swan, who hides malicious instructions in web pages and files, indirectly injected the test, inducing Carbon to deliver directly after 15 consecutive attempts.0.0% attack success rateIt's a record of zero failures.
This means that no matter how sophisticated the language is, no matter how sophisticated the logic is, Carbon doesn't care.
Refusing to cheat, ending illusions
What developers are most afraid of is AI.It's the only way to get around traffic.This is the first time in the history of our country that we've been able to do this, and we've been able to do that.
Two weeks ago, when the Every team was testing Argon, they picked up a live stream, and in 38 test runs, Argon had the same effect.17 timesThe first step is to open and close the necessary code review steps.You can't look at the code and pretend it's over.That's it.。
This is a major correction of Carbon's approach to "cheating and sneezing". In a Terminal-Bench 4.0 special inspection, Carbon was found to be "smart".Successful cheating is squeezed to 0.0%.The rate of attempted cheating has also dropped to0.4%。
In a real programming environment inside Google, Carbon is triggered by theHigh-Danger Alarm Drop by Nearly 80%，Unsubstantiated assertions cut in halfCross-border operations and service contractions also declined by 26% and 29%, respectively.
Extreme candor and transparency
In the AA-Omniscience assessment of knowledge honesty, Claude Opus 5.5 Max and GPT-6 Astra Max still have problems with the confidentiality.About 50% of the probability of forced blind guessingIt's not just carbon.Illusion blind guessesControlledIt's 11.4%.。
When will we be able to use a model like this?The answer is that you can't use it now, and even most businesses can't.
Because it was so powerful, especially its "hole-digging" capabilities, Google adopted a "conservative" publishing strategy.
When Argon was first released, at least Google AI Pro subscribers and paid API developers were able to taste it for the first time.
This week, carbon.Open only to "a small fraction of trusted cybersecurity defenders", the list on Fairwind.
Regular users and independent developers have been kicked out of the door.
In just two weeks, Google refreshed the newly-established recognition of Argon with Carbon.
This terrifying pace of transformation has put pressure on OpenAI and Anthropic.
Next, look at the other two.