Google Releases Gemini 4 Carbon Model, Focused on Long-Horizon Agentic Workflows

Smart East October 29 report, in the early hours of today, Google released its strongest flagship model, Gemini 4 Carbon, which is built specifically for continuously running Agentic workflows, with a focus on improving capabilities in programming, agents and scientific tasks, as well as full-stack web development and multimodal understanding.

The model ranked first with 59 points on the recently updated v5.0 Artificial Analysis Intelligence Index, surpassing a host of models including Claude Opus 5.5, GPT-6 Astra (max), and GPT-6.1 Sol (max).

In terms of pricing, the launch promotional price for Gemini 4 Carbon is $2 per million input tokens and $10 per million output tokens, consistent with Argon's launch promotional price. In addition, cached input tokens can enjoy a 95% discount on the input token unit price, meaning only 5% of the original price needs to be paid.

At the same time, Google's parent company Alphabet announced its financial results for the third quarter of 2026. In this quarter, Alphabet's revenue reached $127.577 billion (about RMB 850.435 billion), up about 25% year over year, achieving 13 consecutive quarters of double-digit growth; net profit reached $26.17 billion (about RMB 174.451 billion), down about 25% year over year; diluted earnings per share were $2.12 (about RMB 14.13), down about 26% year over year.

Among them, Google Cloud's revenue grew 103% year over year to $30.709 billion (about RMB 204.708 billion), becoming the most standout business segment this quarter. In the same period, Microsoft's Azure cloud computing business revenue was about $33.2 billion, up 48% year over year. By comparison, Azure's revenue remains ahead, but Google Cloud's growth rate is more than twice that of Azure.

At the subsequent earnings call, Google further disclosed the commercialization progress of its AI business: Google Cloud's backlog reached $561 billion (about RMB 3.74 trillion), more than half of which is expected to convert into revenue in the next 24 months.

Alphabet's third-quarter capital expenditure reached $51.355 billion (about RMB 342.335 billion), up about 114% from the same period last year. CFO Ashkenazi said that the vast majority of the quarter's capital expenditure was used for the technical infrastructure needed to support AI investments, of which about 65% was for servers and networking equipment, and the remaining about 35% was for data centers.

As of the close on October 28, U.S. Eastern Time, Google Class A shares were at $324.89 per share (about RMB 2,165.73 per share), up about 1.34% from the previous trading day, with a total market value of about $3.97 trillion (about RMB 26.48 trillion). After the earnings release, as of press time, Google Class A's after-hours stock price fell about 9.97% to $292.50 per share (about RMB 1,949.82 per share).

1. Gemini 4 Carbon released, ranking first on the intelligence index

On the day of the earnings release, Google announced the launch of Gemini 4 Carbon. This is another iteration of the Gemini 4 series after the release of Gemini 4 Argon at the end of September and the opening of Argon to developers in mid-October.

According to Google's official introduction, Carbon is its strongest frontier model to date, with a focus on strengthening complex programming, Agent task execution, and cybersecurity defense capabilities.

Compared with model applications that focus on the quality of single responses, Google places more emphasis on Carbon's ability to handle long-horizon tasks: the model can continuously advance complex engineering tasks over several days, and even when it encounters dead ends, it can try to adjust its approach and ultimately complete multi-step work.

Google revealed that internal teams have already used Carbon to handle complex problems in large codebases, including locating and fixing hidden bugs and optimizing underlying systems.

The model has also been improved for fields such as full-stack web development, scientific tasks, and multimodal understanding. In the WebDev Arena web development evaluation, Carbon's Elo score increased by 89 points to 1769, ranking fourth, surpassing GPT-6.1 Sol and Claude Fable 5.1, and trailing only Claude Opus 5.5, GPT-6 Astra, and Claude Sonnet 5.5.

In terms of benchmarks, in programming capability, Carbon ranks first on SWE-sweep and Terminal-Bench 4.0. But on FrontierSWE 2, GPT-6 Astra ranks first with 65.5%, and Carbon's 61.2% is lower than GPT-6 Astra and Claude Opus 5.5.

In the computer operation test, Carbon scored 77.4% on Agents' Last Exam v1.1, higher than GPT-6 Astra and Claude Opus 5.5. In the offline subset of OSWorld-2.1 for operating in real computer environments, Claude Opus 5.5 led with 87.2%, while Carbon was second with 81.5%.

Knowledge work testing is an area where Carbon performs quite prominently. On AutomationBench, it scored 53.9%, notably higher than GPT-6 Astra and Claude Opus 5.5. In the Vending-Bench 2 simulated business test, Carbon ultimately achieved a simulated account balance of $16,484, higher than GPT-6 Astra.

Carbon ranked first in both chart understanding and visual similarity evaluations, and its context handling pages were far higher than those of GPT-6 Astra and Claude Opus 5.5. The same is true in the science and mathematics domains: on RiemannBench, Carbon scored 80.0%, 4% higher than Argon; on Terminal-Bench-Science 0.1, Carbon scored 64.8%, ahead of GPT-6 Astra.

Cybersecurity is another focus of this upgrade. Test data released by Google shows that Carbon scored 88.2% on the SEC-bench Pro vulnerability discovery evaluation, reached 54.8% on the ExploitGym vulnerability exploitation evaluation, and reached 71% on the CWE-bench vulnerability remediation evaluation.

Google says that in internal testing, Carbon's success rate in generating secure code was more than 2 times that of Argon; the Chrome team also reported that on high-quality patch generation tasks, Carbon's performance exceeded that of a frontier model from another AI lab by more than 10 times.

In addition to capability improvements, Carbon also made adjustments for model reliability and Agent execution boundaries. Google says that in internal testing, Carbon's high-severity misalignment alerts decreased by nearly 80%, and unsupported statements decreased by about half; cases of the model exceeding its authority decreased by 26%, and cases of under-execution decreased by 29%. In the indirect prompt injection attack evaluation, Carbon's attack success rate was 0%, meaning it successfully resisted all attack attempts covered by that test.

Carbon adopted a phased rollout strategy this time. Google first opened the model to some cybersecurity defenders, and through the Fairwind program gradually expanded access for trusted partners, before later opening it to developers, enterprises, and consumers. Google says the company is still participating in the U.S. government's pre-release model testing process and will continue to adjust safety measures based on early user feedback.

At today's earnings call, Google CEO Sundar Pichai announced that Google's model API currently processes nearly 32 billion tokens per minute, more than 4 times the level a year ago and up 40% from the previous quarter. At the same time, the Gemini app reached 1 billion users, making it the fastest-growing product in Google's history.

Google is also pushing AI Agents from standalone tools into enterprise business processes. Google says that over the past 12 months, Gemini Enterprise has been used by more than 2,600 customers to process at least 100 billion tokens, with nearly 75 customers using 10 trillion tokens; nearly 90% of Fortune 100 companies have already used Gemini Enterprise.

2. Google Cloud revenue doubles, with $561 billion in backlog orders

Google Cloud became the most noteworthy business of Alphabet in the third quarter.

Financial reports show that Google Cloud's third-quarter revenue reached $30.709 billion (about RMB 204.708 billion), up 103% year over year.

In the same period, the business's operating income reached $11.121 billion (about RMB 74.133 billion), up about 209% year over year; the operating margin rose from 23.7% in the same period last year to 36.2%.

At the earnings call, Google Cloud CEO Thomas Kurian said that the cloud business's growth came from an integrated AI product portfolio covering chips, models, data, security, and agent platforms, while also benefiting from demand across different product lines, industries, and customer groups. Google Cloud currently has 18 product lines each with annual revenue exceeding $1 billion.

Looking at specific businesses, Google Cloud Platform (GCP)'s AI infrastructure, AI solutions, and traditional core cloud services all achieved growth, and TPU system sales also became an important source of revenue contribution.

Kurian revealed that Google is expanding TPU sales from its own cloud platform to customers' own data centers and other service providers to broaden product coverage.

Google disclosed that more than 100 important customers are already using Google TPU systems. In addition to existing partners such as Anthropic, the company is also working with model companies such as Mistral and Physical Intelligence, as well as financial institutions, universities, and public sector customers. Google also announced TPU system delivery partnerships with Oracle and CoreWeave AI, and a collaboration with Meta to advance TorchTPU, in order to lower the barrier for developers to migrate PyTorch models to the TPU platform.

The significance of the TPU business is that Google can sell customers a complete computing system including hardware, compute, storage, networking, and system software, rather than only charging by usage through cloud services.

At the same time, TPU system sales will also bring a cash flow cycle different from traditional cloud services: Google needs to purchase hardware and build inventory first, then deliver to customers and recognize revenue. CFO Anat Ashkenazi explained at the meeting that inventory build-up will tie up cash first, and the related cash will only gradually flow back as delivery and customer payments progress.

What better reflects the future scale of the business than single-quarter revenue is the cloud business backlog. Ashkenazi disclosed that Google Cloud's backlog reached $561 billion in the third quarter, most of which came from regular GCP contracts. The company expects that more than 55% of the backlog will be recognized as revenue in the next 24 months.

This figure means that Google Cloud has already accumulated enormous future contract demand, but backlog is not equivalent to realized revenue, and the actual pace of recognition still depends on contract execution, customer usage, and delivery capacity.

Due to tight supply, the company began using third-party computing resources at the end of the third quarter and expects to expand the scale of use in the fourth quarter. These additional resources may put pressure on cloud business margins in the short term.

In the third quarter, Google Search and other revenue was $63.350 billion (approximately RMB 422.297 billion), up 12% year over year. Pichai said that search query volume growth reached its highest level in the past five years, and AI investment is driving overall query growth, including commercial-related queries.

At the product level, Google is integrating AI Overviews, AI Mode, and the traditional search experience. Pichai said that queries in AI Mode are on average about three times as long as traditional search, and about one-fifth of AI Mode queries worldwide use multimodal inputs such as voice, images, or documents; Google's monthly visual search volume has exceeded 50 billion.

These changes mean that users are completing information retrieval and decision-making through longer, more complex questions. But for the search business, which relies on advertising revenue, changes in query patterns also bring new monetization issues.

At the earnings call, Google Chief Business Officer Philipp Schindler explained that most queries in AI Mode are incremental queries that can bring new commercial intent and future monetization opportunities.

Google is expanding ads in AI Overviews and AI Mode to more users and queries, and for the first time extending the coverage of AI Mode ads to international markets, while launching multi-turn interactive ads in the United States.

Google also disclosed the impact of AI on advertiser tools. In the third quarter, Gemini models helped advertisers generate more than 1.5 billion campaign assets, up more than 7 times year over year. Tools such as AI Max are helping advertisers handle longer, more complex, and multilingual search queries.

However, the commercialization of AI search is still in an adjustment phase. Schindler acknowledged during the Q&A session that the monetization level of AI Mode is relatively low, which has had some impact on the growth rate of the search business. Management stated that it is gradually expanding ad coverage and prioritizing user experience and advertiser return on investment.

IV. Free cash flow turns negative, capital expenditure surges 114%

Along with business growth comes continuously rising infrastructure investment.

Alphabet's capital expenditure in the third quarter reached $51.355 billion (approximately RMB 342.335 billion), up about 114% from the same period last year. CFO Ashkenazi said that the vast majority of capital expenditure in the quarter was used for technical infrastructure needed to support AI investment, of which about 65% was for servers and networking equipment, and the remaining about 35% was for data centers.

Google maintained its full-year 2026 capital expenditure guidance of $195 billion to $205 billion and expects capital expenditure to increase significantly in 2027.

Increased capital expenditure is not the only source of funding pressure. Ashkenazi pointed out that the expansion of AI infrastructure will also continue to affect the income statement through higher depreciation expenses and data center operating costs, including energy and related expenditures. Google also needs to build inventory in advance for TPU system sales, which will also tie up cash.

In the third quarter, net cash generated by Alphabet's operating activities was $42.290 billion, lower than capital expenditure of $51.355 billion. Under the company's disclosed basis, free cash flow for the quarter was negative $9.065 billion; free cash flow for the past 12 months was $19.747 billion.

Ashkenazi said that Google is positioned to monetize AI investment through multiple businesses such as search, cloud, and subscription services, and expects this investment cycle to achieve returns above the cost of capital. But she also acknowledged that technical infrastructure investment and TPU inventory build-up will continue to put pressure on free cash flow.

From the financial report and the earnings call, we can see that Google is entering a stage where it needs to manage demand, capacity, and capital returns at the same time. Google Cloud's backlog continues to increase, and model invocation volume and enterprise Agent applications continue to expand, all of which require corresponding computing resources to support. At the same time, infrastructure construction requires investment in advance, while the related revenue and cash returns are distributed across different time periods.