Google has introduced Gemini 4 Carbon, a new flagship artificial intelligence model that took first place in the Artificial Analysis Intelligence Index v5.0. According to the test results, Carbon outperformed Claude Opus 5.5 and GPT-6 Astra, while Google’s biggest improvements are in coding, autonomous tasks, and cybersecurity. There is a catch, however: the model is not yet available to the general public. Its leadership in the rankings therefore does not mean users have gained access to the best AI assistant for everyday work.
Key Takeaways
- Google announced Gemini 4 Carbon on October 28, 2026, two weeks after the launch of Gemini 4 Argon.
- Carbon scored 59 points in the Artificial Analysis Intelligence Index v5.0, compared with 58 for Claude Opus 5.5 and 57 for GPT-6 Astra.
- In cybersecurity tests, Google reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71% on CWE-bench.
- According to Google’s internal evaluations, Carbon makes unsupported claims about half as often as Argon.
- At launch, access is being granted to a limited group of cybersecurity specialists rather than all Gemini users.
- The announced introductory API pricing is $2 per million input tokens and $10 per million output tokens, but the public API has not yet launched.
Gemini 4 Carbon Takes First Place in Artificial Analysis
The most notable result of the new release is its first-place ranking in the Artificial Analysis Intelligence Index v5.0. This is a composite metric that combines results from several independent AI model evaluations.
Carbon scored 59 points in the ranking. By comparison, Claude Opus 5.5 scored 58, GPT-6 Astra scored 57, and the earlier Gemini 4 Argon scored 50.
The gap between Carbon and its closest competitor is just one point. That is enough for first place, but not enough to automatically declare the model the undisputed winner in every practical scenario.
Moreover, the ranking has moved to the updated v5.0 methodology, with new and modified tests. According to Artificial Analysis, approximately 60% of the evaluations in this version use private tasks or hidden answers. This is intended to reduce the risk that models have already seen the test materials during training.
For that reason, it is appropriate to compare model results within the same version of the index. Comparing the new scores with earlier rankings without accounting for methodological changes would be misleading.
Where Gemini 4 Carbon Is Actually Strong
Google positions Carbon not simply as a general-purpose chatbot but as a model for long-running autonomous workflows. These are tasks in which AI must plan work, use tools, check intermediate results, and return to a problem after unsuccessful attempts.
Coding and Complex Multi-Step Tasks
Carbon scored 65.2% on Terminal-Bench 4.0, compared with 57.1% for Argon. This benchmark tests models’ ability to perform tasks in a terminal environment.
On FrontierSWE 2, which evaluates software engineering work, Carbon scored 61.2%. That is a notable improvement over Argon’s 55%, although GPT-6 Astra achieved a higher result in this test at 65.5%.
In other words, Carbon does not win everywhere. Even among the comparisons published by Google, there are scenarios in which competitors remain ahead.
Google also reports that Carbon was able to work on complex engineering problems for several days in internal evaluations. However, such examples should not be confused with a guarantee of similar performance for every user or project.
Cybersecurity Is the Main Focus of the Release
Google devoted particular attention to cybersecurity tasks. Carbon was tested on finding vulnerabilities, checking whether they can be exploited, and creating fixes.
- SEC-bench Pro — 88.2%. Detecting real software vulnerabilities.
- ExploitGym — 54.8%. Checking whether discovered vulnerabilities can be exploited.
- CWE-bench — 71%. Creating fixes for security issues in code.
According to Google, Carbon more than doubled its success rate in generating secure code compared with Argon in internal tests.
At the same time, the company notes that some external cybersecurity evaluations were conducted without the model’s safety guardrails. This matters when interpreting the results: performance in a controlled environment may not necessarily be replicated in a commercial product with active safety mechanisms.
Has Gemini 4 Carbon Become Less Prone to Hallucinations?
One of the most important aspects of the new release is not speed or ranking position, but the reliability of its answers.
Google says that in internal evaluations, Gemini 4 Carbon roughly halved the number of unsupported claims compared with Gemini 4 Argon.
The company also reports an almost 80% reduction in severe warnings about incorrect model behavior in its own coding environments.
These results are encouraging, but they come from Google’s testing. There has not yet been enough independent, large-scale evaluation of Carbon in everyday scenarios because the model is not available to everyone.
Moreover, cutting hallucinations in half does not mean eliminating them entirely. When working with facts, financial calculations, legal documents, or advertising analytics, AI outputs still need to be verified.
Why First Place in a Ranking Does Not Guarantee the Best AI for Work
Benchmarks are useful for comparing model capabilities under identical conditions. But they do not fully reflect work with real requests, ambiguous instructions, large documents, or specialized business processes.
For example, a model may perform well on mathematics and coding benchmarks while falling behind a competitor in text editing, spreadsheet analysis, or following complex output-format requirements.
The published comparisons contain concrete examples of such differences. On OSWorld-2.1, which evaluates interaction with computer interfaces, Carbon scored 81.5%, while Claude Opus 5.5 scored 87.2%.
Therefore, an overall ranking should be treated as a preliminary selection tool rather than a final verdict on AI quality.
To compare different approaches to using AI models, UAGEEK has already published an article on ChatGPT or Gemini: Which AI Is Better for Different Tasks. And the Artificial Analysis platform can be used to evaluate model quality, speed, and cost before practical testing.
Who Can Access Gemini 4 Carbon and How Much Will It Cost?
Despite the official announcement, Gemini 4 Carbon did not receive a full public launch on October 28.
Google began granting access to a limited group of cybersecurity specialists. The next stage is expected to involve a gradual expansion to trusted partners through the Fairwind Program.
The company also announced its participation in a voluntary pre-release model testing process involving the US government. Broader access for developers, businesses, and ordinary users is promised later, but the announcement gives no specific date for a general release.
Introductory API pricing is expected to match Gemini 4 Argon:
- $2 per 1 million input tokens.
- $10 per 1 million output tokens.
- A 95% discount on cached input tokens.
After the introductory pricing period ends, Google plans to raise rates to $4 per million input tokens and $20 per million output tokens.
These amounts are announced prices, not confirmation that the public API was available at the time of the announcement.
What Gemini 4 Carbon Could Change for Marketing and Online Business
The clearest potential business benefit of Carbon is the automation of complex processes that require a sequence of actions rather than a single answer.
For marketing teams, these could include checking large datasets, analyzing technical website issues, working with code, preparing analytical reports, and supporting internal automations.
However, none of the benchmarks cited proves that Carbon is already better than its competitors at optimizing advertising campaigns, creating high-converting creatives, or conducting SEO research.
Once access opens up, companies should evaluate the model on their own controlled tasks, comparing accuracy, the number of corrections needed, token costs, execution time, and error frequency.
For those who use Gemini in everyday work, the development of Google’s ecosystem also matters. UAGEEK previously covered Gemini integration with Gmail, Google Photos, and Search. However, Carbon’s announcement does not mean the new model is already running in all those services.
Conclusion: Gemini 4 Carbon Leads the Tests but Is Not Yet the Everyday Favorite
Gemini 4 Carbon has demonstrated significant progress over Argon and taken first place in the updated Artificial Analysis index. Its results in cybersecurity, coding, and long-running autonomous tasks look particularly convincing.
But its lead over the closest competitors in the overall ranking is small, and some specialized benchmarks are still led by other models. Carbon is also currently available only to a limited group of partners.
Calling it the best AI for every user would therefore be premature. Independent practical evaluations after a broader launch will provide a clearer answer, especially on response reliability, operational stability, and the real cost of complex tasks.
Sources for Verification
Sources for Verification
Google — Introducing Gemini 4 Carbon, October 28, 2026. Official model announcement, benchmark results, internal reliability assessments, access conditions, and API pricing.
Official Gemini 4 Carbon Announcement
Google — Q3 2026 Earnings Call. Confirmation of Carbon’s launch, its position in Artificial Analysis, and its initially limited availability.
Google Leadership Commentary
VentureBeat — Google unveils Gemini 4 Carbon. Analysis of comparative benchmarks, the Artificial Analysis Index v5.0 update, and the limits of interpreting the results.
Gemini 4 Carbon Benchmark Comparison
The Hacker News — Gemini 4 Carbon Leads Cybersecurity Benchmarks. Independent coverage of the model’s performance in detecting, validating, and fixing vulnerabilities.
Carbon’s Cybersecurity Results
DataCamp — Gemini 4 Carbon. Comparison of Google’s reported metrics with independent assessments and an explanation of benchmark limitations.
Carbon Testing Overview
