According to Google DeepMind, internal teams have run Carbon continuously for several days while it investigated software problems, worked through unsuccessful approaches and developed fixes across large codebases. The company also reports improvements in scientific reasoning, multimodal understanding and web development.
Independent evaluations provide additional evidence of Carbon's capabilities, although its performance varies across different tasks.
On Artificial Analysis's Coding Agent Index, Carbon scored 59, narrowing the gap with Claude Opus 5.5, which scored 62. The model also reached 65.2% on Terminal-Bench 4.0, compared with Argon's 57.1%, and improved from 69.0% to 77.4% on Agents' Last Exam.
Carbon performed particularly well on long-context reasoning, scoring 96.2% on GraphWalks, compared with 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5. Its WebDev Arena rating increased by 89 points to approximately 1,768, placing it fourth.
Other independent assessments showed similar gains. Carbon ranked first on Arena's Text leaderboard with 1,528 points and second on its Agent Arena evaluation. Vals AI placed the model second overall with 70.36% accuracy, narrowly behind Claude Sonnet 5.5 at 70.85%.
However, Carbon does not lead every evaluation. On FrontierSWE 2, it scored 61.2%, trailing GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%. Anthropic's model also performed better on the OSWorld offline subset, scoring 87.2% against Carbon's 81.5%.
The results suggest that Carbon is highly competitive across general reasoning and agentic tasks, but its advantages are not uniform across every coding and computer-use benchmark.
Cybersecurity is one of the model's most significant areas of improvement.
Google reports that Carbon achieved 88.2% on SEC-bench Pro, which evaluates the ability to identify vulnerabilities in real software. That compares with 83.4% for Argon.
On ExploitGym, which measures whether identified weaknesses can be successfully exploited, Carbon reached 54.8%, up from 42.0%. It also scored 71.0% on CWE-bench, an evaluation of software vulnerability remediation, compared with Argon's 68.0%.
Together, the results show improvements across vulnerability discovery, validation and patch generation. Google noted that these external cybersecurity evaluations were conducted without safety restrictions, meaning performance in controlled deployments could differ.
The company also cited improvements in its internal security testing. Carbon more than doubled Argon's success rate when generating secure code, while Google's Chrome team reported that the model produced substantially better security patches than an unnamed competing frontier model, with an improvement exceeding tenfold.
Wiz, Google's cloud security platform, also reported strong results from Carbon on web penetration-testing evaluations.
These capabilities help explain Google's decision to limit initial access. The company is participating in the U.S. government's voluntary pre-release model testing process and is making Carbon available first to a subset of cybersecurity defenders through its Fairwind Program.
Access will subsequently expand to additional trusted partners. Google intends to offer the model to developers, businesses and consumers through its products after gathering feedback and refining safety protections, but has not announced a broader release date.
Alongside performance improvements, Google has focused on reducing unreliable behavior during complex assignments.
The company says Carbon reduced successful attempts to exploit shortcuts on Terminal-Bench 4.0 to 0.0%, while the rate of attempted cheating fell to 0.4%. In internal coding environments, alerts involving serious model misalignment declined by nearly 80%.
Google also reported that unsupported claims fell by approximately half compared with Argon. Instances of acting beyond assigned boundaries declined 26%, while failures to complete required work dropped 29%.
On Gray Swan's indirect prompt-injection benchmark, Carbon recorded a 0.0% attack success rate, compared with 0.7% for Argon. That result reflects the evaluated attacks rather than a guarantee against all prompt-injection techniques.
Not every reliability measure improved. Artificial Analysis found that Carbon answered 58% of questions correctly on its knowledge evaluation, up from 48% for Argon, but its measured hallucination rate also increased from 4% to 11%.
The combination illustrates a distinction between greater answer accuracy and the frequency of incorrect claims. Carbon produced more correct answers overall, but the evaluation also recorded a higher rate of hallucinations.
Google is introducing Carbon with the same initial API pricing as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount.
After the introductory period, pricing will increase to $4 per million input tokens and $20 per million output tokens. Google has not specified when Carbon's introductory rates will expire.
Artificial Analysis estimated Carbon's cost at $7.76 per task using regular pricing, compared with $10.04 for Claude Opus 5.5 and $7.68 for Argon. That puts Carbon roughly 23% below Anthropic's model on the evaluation while delivering a higher overall Intelligence Index score.
The model's additional reasoning also introduces trade-offs. Vals AI found that some financial and legal research assignments took considerably longer than with Argon, with Carbon generating more reasoning and output tokens to complete those tasks.
For now, Gemini 4 Argon remains the more broadly accessible model. Carbon's initial availability is limited despite its leading benchmark position, and Google has not disclosed its parameter count, architecture or a detailed technical report.
The release gives Google a new top-ranked model only two weeks after its previous flagship, with particularly notable advances in cybersecurity and sustained engineering work. Its wider impact will depend on how those capabilities perform under deployment restrictions and whether the benchmark improvements translate into reliable results for developers and businesses once access expands.
This analysis is based on reporting from Google & the decoder.
Image courtesy of Google.
This article was generated with AI assistance and reviewed for accuracy and quality.
