Google Releases Gemini 4 Carbon, Surpassing Claude Opus 5.5 on Multiple Benchmarks

Just now, Gemini 4 Carbon comprehensively surpassed Opus 5.5! Token pricing halved at launch

Just two weeks after Argon's release, Google has swiftly pulled out its new flagship, Gemini 4 Carbon.

This time it has surpassed Claude Opus 5.5 and GPT-6 Astra in one fell swoop.

In the just-updated Artificial Analysis Intelligence Index 5.0, it stormed to the top spot worldwide with a high score of 59.

When Google released Gemini 4 Argon two weeks ago, Silicon Valley was still buzzing about a mysterious internal checkpoint leaked from Google called "Carbon"; at the time, internal engineers rated it as "coding feel approaching Opus 5.5."

Unexpectedly, in just half a month, the official Carbon not only completed a head-on overtake of Opus 5.5 in intelligence and long-horizon coding, but also pulled far ahead in code security and anti-cheating.

It even lit up a 100% immunity to prompt injection talent.

What kind of brutal pace of progress is this—did they really figure out RSI?

In its launch phase, Carbon continues Argon's half-price discount: $2 per million input Tokens and $10 per million output Tokens, with the list unit price only half that of Opus 5.5 ($4/$20).

It also offers a 95% rock-bottom discount on cached input (as low as $0.2 per million Tokens), which is frequently used in Agent tasks.

Surpassing Opus 5.5, and able to grind away at low-level bugs for days on end

Although Gemini 4 Argon, released two weeks ago, impressed in long context and multimodality, it was still edged out by Opus 5.5 on the composite intelligence index (53 points) and some hardcore terminal tasks.

This time, Gemini 4 Carbon precisely filled in the last piece of the puzzle, pushing its capabilities directly toward the form of an "all-weather autonomous engineer."

In Google's internal real-world coding environment, the research team observed a scene that left them slumped in their chairs:

Carbon can run continuously for several days, exploring massive, complex codebases and low-level systems on its own; even when it hits "dead ends" multiple times along the way, it keeps grinding to the end and independently completes complex multi-step engineering tasks. From following the trail to pinpoint the root cause of an extremely well-hidden bug to finally writing the fix patch, it does it all in one go.

This long-horizon execution capability is directly reflected in overtakes on multiple hardcore metrics:

Command line and terminal engineering

On Terminal-Bench 4.0 (AA test set), Carbon scored 65.2%, ahead of Opus 5.5 (59.6%) and GPT-6 Astra (59.1%);

On the scientific terminal task Terminal-Bench-Science 0.1, it also ranked first with 64.8% (Opus 5.5 was 59.0%).

Ultra-long context and multi-step decision-making

On the GraphWalks test with 256k to 1M context, Carbon scored 96.2%, far ahead of Opus 5.5 (66.8%) and Astra (71.8%);

In Vending-Bench 2, which simulates long-cycle business operations, Carbon's final account balance reached $16,484, nearly 1.8 times that of Opus 5.5 ($9,235).

Full-stack development and enterprise tasks

On WebDev Arena, thanks to improvements in front-end components and output quality, Carbon's Elo score rose 89 points to 1769. Although it ranks fourth worldwide, its average task cost is lower.

On the Vals Index v2.2, which measures the value of enterprise-level workflows, it also narrowly led Opus 5.5 (70.2%) with 70.4%.

However, according to a breakdown by third-party evaluation organization Artificial Analysis, Opus 5.5 still retains some moat in the pure code engineering field.

For example, in Google's own published table, Carbon failed to top two metrics:

One is FrontierSWE 2, which measures frontier software engineering; GPT-6 Astra (65.5%) and Opus 5.5 (62.3%) still rank ahead of Carbon (61.2%);

The other is the offline subset of the computer operation benchmark OSWorld-2.1, where Opus 5.5 remains first with 87.2% (Carbon was 81.5%).

In addition, on the Coding Agent specialized index paired with their respective official CLI tools, Carbon scored 59 with Google's Antigravity CLI, still 3 points behind Opus 5.5 (62) backed by Claude Code.

This wave of Antigravity still dragged down Hakimi.

But for developers, Carbon's real lethality lies in Token consumption efficiency.

Independent evaluation data shows that although Opus 5.5 is capable, its reasoning process is extremely lengthy, consuming an average of 173,000 output Tokens to complete a single Intelligence Index evaluation. Carbon needs only 123,000 Tokens, a 29% reduction in output.

This means that even if Carbon returns to the same $4/$20 standard pricing as Opus 5.5 in the future, Carbon's average cost per task ($7.76) is still about 23% lower than Opus 5.5's ($10.04).

If you factor in the current launch half-price of $2/$10, Carbon's billing cost in actual production environments is less than 40% of Opus 5.5's.

Beyond conventional programming, Google has this time staked Carbon's core differentiating selling point on the full lifecycle of cybersecurity defense.

This is also out of necessity—you can't just watch frontier models dig up vulnerabilities in the Chrome browser and Android system every day, right?

(The following external benchmarks were all evaluated in a full-power state with safety guardrails removed):

Step 1: Finding a needle in a haystack—finding real vulnerabilities

The core of cybersecurity defense is to find deeply hidden complex flaws in real software before hackers do. In the SEC-bench Pro test, which includes real high-risk vulnerabilities such as the V8 engine and the Linux kernel, Carbon scored 88.2%, not only greatly surpassing the previous generation Argon (78.8%), but also directly beating OpenAI's GPT-6 Astra (85.4%), setting a new SOTA.

Step 2: Validating risk through exploitation

Faced with hundreds or thousands of potential vulnerabilities, which one should the security team fix first? This requires the model to act like a real attacker and verify which vulnerabilities are truly exploitable.

On the well-known exploit benchmark ExploitGym, Carbon once again set a new SOTA with a score of 54.8% (Argon was previously 35.1%), helping defenders accurately prioritize fixes.

Step 3: Patching at the root cause

After finding and verifying vulnerabilities, the ultimate test is to directly write fix code that runs and does not break existing functionality. In Collinear's CWE-bench (a real repository audit and repair test covering 73 common weakness enumerations), Carbon reached 71.0%, comprehensively crushing Claude Opus 5.5 (67.0%) and Argon (68.0%).

In addition to dominating external leaderboards, in Google's internal security benchmarks, Carbon's success rate in generating secure code is more than 2 times higher than Argon's.

An even more exaggerated result comes from Google's own Chrome browser team: in a real-world comparison of generating high-quality security patches, Carbon's performance was more than 10 times stronger than a large frontier model from another top-tier lab.

Cracking down on Agent cheating, 100% immunity to prompt injection

When a model runs continuously in the terminal for several days, what developers find most headache-inducing is often not the compute bill, but the model starting to "play smart" (Shortcut & Loophole Exploitation) when facing high-pressure complex tasks.

A previous third-party audit of Terminal-Bench 4.0 revealed that many high-scoring models, when unable to fix a bug, would modify test scripts or exploit loopholes to fake "test passed." Targeting this chronic problem, Google made a key convergence in Carbon's alignment training:

Cheating success rate directly driven to 0.0%

On the terminal operation benchmark Terminal-Bench 4.0, while greatly improving overall accuracy, Carbon forced the "successful cheating rate" down to 0.0%, and even the proportion of attempts to take shortcuts to cheat fell to 0.4%.

In Google internal teams' live programming environments, high-severity misalignment alerts plummeted by nearly 80%.

Refusing to pretend to understand, hallucination rate far lower than competitors

In internal tests, Carbon's frequency of making unsupported claims was about 2 times lower than Argon's.

When faced with missing evidence or blocked operations, it no longer speaks vaguely or forcibly makes things up, but instead writes the limitations of its methodology directly into the delivery document, or pulls out a clear list telling humans "what conditions are still missing."

Ask when you don't know, clearly define execution boundaries

When faced with ambiguous requirements, Carbon does not blindly guess or force progress, but first retrieves relevant sources and puts the contradictions it finds on the table to confirm with the user.

When executing multi-step plans end to end, its overreach decreased by 26%, and its under-service decreased by 29%.

100% immunity to prompt injection

When an Agent reads web pages, emails, and external code repositories in the wild, it is highly vulnerable to hijacking by "indirect prompt injection" (IPI).

In the Gray Swan IPI benchmark, the previous generation Argon once led the field with an extremely low breach rate of 0.7%.

This time, Carbon directly delivered a perfect score of 0.0% attack success rate, successfully resisting every attack attempt in the evaluation.

Behind the 14-day blitz: RSI fully accelerated

From Gemini 4 Argon, which debuted at the end of September and was just fully rolled out two weeks ago, to Gemini 4 Carbon, which burst onto the scene today and comprehensively surpassed Opus 5.5.

CEO Pichai even joked boastfully on X (doge):

I said before that we would release models periodically... Now let me introduce Gemini 4 Carbon to everyone.

Combining previously leaked internal information, Google actually trained and evaluated multiple Gemini 4 checkpoints such as Argon, Barium, and Carbon in parallel internally (the publicly released Argon was actually the internal Barium-B checkpoint).

In an inconspicuous corner of the official announcement, Google DeepMind Senior Vice President and Chief AI Architect Koray Kavukcuoglu revealed the real secret:

"Our researchers are working hand in hand with our models to advance capability evolution at a faster pace, while simultaneously strengthening safety guardrails."

However, Google is still Google—release does not equal availability.

Because Carbon has strong vulnerability discovery and exploitation capabilities, it still needs to go through the pre-release testing process. Starting this week, Carbon will first be opened to some cybersecurity defense personnel.

Remember how long it took from Argon's release to actually being available to you? This time the wait won't be short either.

At the critical moment of Anthropic's IPO, shouldn't it be time to release Fable 5.5 now?

As for OpenAI, which is obsessed with releasing the mid-tier Terra model and the SMS notification Dot, I wish it good luck.

Reference links: