Google releases Gemini 4 Carbon, topping the Artificial Analysis Intelligence Index with 59 points

Why is the shelf life of Google's flagship model only 15 days?

At 3 hours before Alphabet's Q3 earnings call on October 28, US time, Google released its new-generation flagship model Gemini 4 Carbon, which topped the latest Artificial Analysis Intelligence Index with 59 points, surpassing Claude Opus 5.5 and GPT-6 Astra to become the global No. 1.

The model took first place, but the earnings report was not so celebratory.

Alphabet's Q3 free cash flow deteriorated to negative $9.065 billion, negative for the second consecutive quarter, and the stock fell more than 10% after hours at one point.

Just two weeks earlier, Google had officially opened Gemini 4 Argon to paying users.

Some netizens commented directly: "Carbon is much stronger than Argon. I really wish Google had just skipped Argon back then, that way this comeback would be more convincing."

Other netizens joked about Google's naming approach: "From Argon to Carbon in two weeks, next is Iron, and then it's time for heat treatment. LOL."

Judging from the data already made public, the progress of Gemini 4 Carbon is indeed quite obvious.

Unfortunately, Google once again chose a phased rollout, so ordinary users can only watch the new champion on the leaderboard while continuing to use Argon, which netizens had just suggested "just skip."

Hakimi returns to global No. 1

After about half a year, Google has finally returned to the top of the Artificial Analysis Intelligence Index.

The last time Gemini topped the list was in April this year, when Gemini 3.1 Pro tied for first with GPT-5.4 and Claude Opus 4.7. OpenAI and Anthropic then successively launched stronger models, and Google did not regain the top spot until the release of Carbon.

In Artificial Analysis's latest Intelligence Index v5.0, Carbon scored 59 points, 9 points higher than Argon's 50, 1 point higher than Claude Opus 5.5, and 2 points higher than GPT-6 Astra.

Among the 9 tests that make up the index, Carbon took first place in 4, and is especially strong in complex reasoning, scientific research, knowledge Q&A, and automated tasks.

Among them, the most noteworthy is the improvement in coding ability.

In Artificial Analysis's other Coding Agent Index, which specifically measures AI coding ability, Carbon scored 59 points, second only to Claude Opus 5.5's 62 points. Google has never been particularly outstanding in such tests in the past, and this time it has finally caught up.

In the Terminal-Bench 4.0 terminal coding test, Carbon's score also rose from Argon's 57.1% to 65.2%. In web development, Carbon's score on WebDev Arena increased by 89 Elo points compared with Argon, rising to fourth place with 1768 points (Google's official announcement says 1769 points; this article follows the original Arena leaderboard).

Long-context processing ability has also improved significantly. In the GraphWalks test, Carbon's score rose from Argon's 84.2% to 96.2%, surpassing GPT-6 Astra's 71.8%. This test examines the model's ability to track complex associations in contexts as long as one million tokens.

However, in coding agents, some software engineering and computer operation tests, Claude and GPT still maintain the lead.

Beyond benchmark scores, Google specifically emphasized Carbon's ability to execute complex tasks.

The official announcement said that Carbon was specifically optimized for long-running agent workflows. In internal tests, it has been able to run continuously for several days, repeatedly trying different methods, and locating and fixing complex bugs in large codebases and low-level systems. This means that Carbon is expected to keep pushing work forward when encountering difficulties and gradually take on tasks that previously required engineers to participate for long periods.

Another key upgrade is the model's cybersecurity capability.

According to official documents, in the SEC-bench Pro vulnerability discovery test, Carbon scored 88.2%; in the ExploitGym exploit verification test, it reached 54.8%; and in the CWE-bench vulnerability fix test, it reached 71.0%, all achieving leading results.

Google also revealed that in internal secure code generation tests, Carbon's success rate was more than twice that of Argon.

In addition, in Gray Swan's indirect prompt injection test, Carbon also achieved a 0% attack success rate. This test specifically examines whether AI can resist malicious instructions hidden in web pages, emails, and other external materials.

With attackers allowed up to 15 attempts, Carbon withstood all test attacks. By comparison, Argon's attack success rate was 0.7%, Claude Opus 5.5's was 1%, and GPT-6 Astra's was 8.5%.

One more interesting point: Carbon is officially recognized as an "honest model."

In the past, some AI models might exploit loopholes in the test environment in order to complete tasks, and even obtain high scores through means that did not meet the task requirements. But Google claims that Carbon's cheating success rate in the Terminal-Bench 4.0 test has dropped to 0%, and its cheating attempt rate has also fallen to 0.4%.

In internal testing, the number of unsupported assertions made by Carbon was reduced by about half compared with Argon, and high-severity behavioral deviation alerts fell by nearly 80%. Cases of the model exceeding its authority decreased by 26%, and cases of insufficient task completion decreased by 29%.

As for pricing, Carbon is temporarily following Argon's discounted pricing.

It charges $2 per million input tokens and $10 for output, with a 95% discount on cached input. After the promotional period ends, the prices will rise to $4 and $20, respectively.

Artificial Analysis's tests also show that, calculated at the standard API prices after the promotional period ends, Carbon's average cost to complete a comprehensive evaluation task is $7.76, basically on par with Argon's $7.68 (but with a comprehensive score 9 points higher).

By comparison, Claude Opus 5.5's single-task cost is $10.04, making Carbon about 23% cheaper; GPT-6 Astra costs only $6.24, about 20% lower than Carbon.

The only problem is that not many people can actually use Carbon yet.

Google announced that this week it will first open the model to a small number of cybersecurity defense organizations, and then gradually expand it to trusted partners through the Fairwind program. Ordinary consumers, developers, and enterprise customers will still need to keep waiting, and a specific opening date has not yet been announced.

Stronger, but also slower

Although Carbon has not yet been fully opened, before its official release, some developers have already tested it through an early access program and Arena.

Judging from the public feedback so far, Carbon is indeed considerably stronger than Argon—especially in 3D content generation.

On October 21, X user Lumina shared a set of test results, including generating a voxel pagoda and drawing a pelican riding a bicycle with SVG.

He believes that the quality of 3D content generated by Carbon clearly exceeds that of Argon, and its overall performance across various tasks is also good. But there is one very prominent problem: it is too slow.

Lumina said that depending on the specific task, Carbon's generation speed is about 2 to 4 times slower than Argon's.

Another netizen, Lentils, who has been following Carbon's early tests, gave a similar assessment: 3D capabilities have clearly improved, front-end and UI design are still not ideal, and the model has become slower.

However, he also found a positive change: Carbon is not as "lazy" as before, and is more proactive when executing tasks, making it less likely to do things perfunctorily.

The speed issue is not an illusion on the part of individual developers. Independent evaluation agency Vals AI also found that Carbon's capability improvements come with a clear speed cost.

In financial and tax agent tests, Carbon's average task completion time was 2.25 times and 2.12 times that of Argon, respectively. Among them, the accuracy of the financial agent even declined slightly from 67.66% to 67.16%.

In the legal research test, although Carbon took first place, it needed nearly 52 minutes per task on average, about 20 minutes more than Argon.

Another controversy comes from Carbon's accuracy issue.

A netizen said bluntly on X: "Honestly, as long as Google hasn't significantly reduced the model's hallucination rate, I'm not interested no matter how many more points it scores than other models on benchmarks."

Artificial Analysis's AA-Omniscience test shows that Carbon's correct answer rate for knowledge Q&A increased from Argon's 48% to 58%, but its hallucination rate rose from 4% to 11%.

The hallucination rate here measures, among questions the model did not answer completely correctly, the proportion in which it gave a wrong answer instead of admitting that it did not know.

As mentioned earlier, Google also specifically emphasized in its official announcement that Carbon's unsupported assertions were reduced by half compared with Argon; the tasks and statistical criteria of the two tests are different.

Many developers are also dissatisfied with Google's release approach.

On Hacker News, some netizens complained that Google rushes to announce a new model every time but delays giving a clear opening date. By the time Carbon is truly opened to the public, OpenAI and Anthropic may have already released the next, stronger model.

One netizen offered a rather cutting analogy: he felt that Google's release approach this time was simply like a person showing off everywhere: "I have a girlfriend, it's just that she's at another school, and none of you know her."

The model took first place, but the accounts are not settled yet

A few hours after Carbon's release, Google's parent company Alphabet announced its third-quarter financial results.

If you look only at revenue and the growth of the cloud business, Google's results are actually quite impressive.

In the third quarter, Alphabet's total revenue reached $127.577 billion, up 25% year over year. Of that, Google Cloud revenue was $30.709 billion, up 103% year over year, with revenue more than doubling compared with the same period last year. Operating profit for the cloud business also rose from $3.594 billion in the same period last year to $11.121 billion, with the margin increasing from 23.7% to 36.2%.

User scale is growing as well. Pichai revealed on the earnings call that the Gemini App has reached 1 billion users, making it the fastest-growing product in Google's history; Google's own model API processes nearly 32 billion tokens per minute, more than four times the level a year ago.

However, the reaction from the capital markets was rather cold.

During regular U.S. trading on October 28, Alphabet's stock price was still up 1.34%. But as the earnings report was released, the stock quickly turned lower, with after-hours losses at one point exceeding 10%.

Carbon had just helped Google take the global No. 1 spot, but that seemingly failed to shore up investor confidence either.

One reason is that the advertising business fell short of expectations. In the third quarter, Google advertising revenue reached $82.012 billion, up about 11% year over year, but about $1.41 billion below analyst expectations.

On the earnings call, Google executives said that although AI Mode has brought rapidly growing query volume, the advertising monetization level per query is still lower than that of traditional search, which has therefore also weighed to some extent on the revenue growth rate of Google's search business.

At the same time, Google is spending money faster and faster.

In the third quarter, Alphabet's capital expenditure reached $51.355 billion, and free cash flow fell to negative $9.065 billion, negative for the second consecutive quarter.

Although this quarter's cash flow was also affected by the payment of a $5.2 billion EU fine and factors such as stockpiling TPU systems, the pressure of infrastructure investment is indeed real.

Alphabet expects full-year capital expenditure in 2026 to reach $195 billion to $205 billion, and to continue increasing significantly in 2027.

To meet continuously growing computing demand, Google has even begun using third-party computing power, and the chief financial officer expects this to put pressure on cloud business margins in the short term.

On the call, Pichai also revealed Carbon's next development plans. Google is developing a distilled version of Carbon and hopes to deploy it in products with a user scale of 1 billion while balancing computing cost, response latency, and model quality.

This also indirectly shows that there is still a considerable cost and engineering challenge between making a model the global No. 1 and putting the global No. 1 model into the hands of 1 billion users.

Of course, Google has its own advantages. It already has a huge user base, a cloud business with continuously growing profitability, and TPU systems that can be sold externally. Even if the strongest model cannot yet be fully opened up, Google still has many channels to turn AI capabilities into revenue.

Moreover, Google has already begun to enjoy the productivity dividends brought by AI internally. Pichai revealed that in recent weeks, Google's Chrome, ads, and search teams have used Argon to write and optimize code, increasing the amount of code delivered by 35% and reducing errors by 30%.

Now, Google has also begun deploying Carbon internally. Pichai believes that the new model's capabilities have improved markedly again, and that as more employees gain access and computing support, Google's internal productivity is expected to improve further in the coming year.

Carbon has indeed restored Google's technical face and put it back in an advantageous position in competition with OpenAI and Anthropic.

But what the capital markets need is probably more than just a pretty leaderboard.