Google introduces Gemini 4 Carbon, reserving first access for cyber defenders
Google reports stronger vulnerability detection, patching and safeguards. Selected defenders get access this week; developers and consumers still have no firm release date.
Google announced Gemini 4 Carbon on October 28, 2026, positioning it for long-running engineering and cybersecurity work, with defenders getting first access. Google reports leading benchmark results for finding vulnerabilities, validating exploits and patching flaws, but says those cybersecurity evaluations ran without guardrails. The staged release lets Google gather feedback and refine safeguards before broader availability; access for developers, businesses and consumers has no firm date, so the model’s reported performance and eventual rollout remain to be tested.
01
Google reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench; all three cybersecurity evaluations ran without guardrails.
02
Selected cyber defenders are due to get access this week, with trusted partners added gradually through the Fairwind Program; Google has set no date for wider access.
03
Google reports Carbon scored 59 on the Artificial Analysis Intelligence Index 5.0 and reached fourth place on WebDev Arena at 1769; these are company-cited results.
Google’s next flagship AI model is headed to cybersecurity defenders before ordinary customers. Announced on October 28, 2026, Gemini 4 Carbon is designed for engineering work that can stretch across days. Google calls it its strongest model, but plans to gather feedback and refine safeguards before opening access to developers, businesses and consumers.
Engineering work measured in days
Carbon is pitched as more than a tool for answering coding questions. Google says its internal teams have seen the model work for days on complex tasks, moving past failed approaches. It has traced subtle bugs and patched them across large codebases and low-level systems. Google says Carbon is designed for extended workflows in which an AI agent carries out a series of actions.
Google also reports gains outside that internal engineering work. It places Carbon first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. For web development, Google reports an 89-point improvement on WebDev Arena, taking Carbon to fourth place at 1769. These are rankings cited in Google’s announcement, rather than performance independently tested for this story.
Finding flaws, proving risk, writing fixes
Cybersecurity is the release’s clearest application. Google presents Carbon as covering three connected jobs: finding a software weakness, checking whether it can actually be exploited, and fixing its underlying cause. Validation helps defenders decide which flaws deserve priority. The company’s headline external benchmark results cover each step, but Google explicitly says those evaluations ran without guardrails.
Google’s reported cybersecurity scores
88.2%Finding vulnerabilities
Google reports 88.2% on SEC-bench Pro, which tests the discovery of vulnerabilities in real-world software.
54.8%Validating exploitation
Google reports 54.8% on ExploitGym, a benchmark for exploiting vulnerabilities to validate risk.
71.0%Patching flaws
Google reports 71.0% on Collinear’s CWE-bench, which assesses fixes addressing vulnerabilities’ root causes.
Inside Google, the company says Carbon more than doubled Argon’s success rate at generating secure code on internal benchmarks. Its Chrome team also reported more than a tenfold improvement over another lab’s large frontier model at producing high-quality patches. Google did not name that competing model in the announcement, leaving the comparison without a specific rival readers can identify.
Safety changes before wider access
Alongside those capability claims, Google describes training intended to reduce shortcuts and inappropriate behavior under pressure. Carbon retains Argon’s layered safeguards, including real-time misuse defenses, monitoring of its reasoning, and resistance to prompt injection—attempts to steer a model through malicious instructions embedded in material it encounters.
Google reports zero successful cheating on Terminal-Bench 4.0, with cheating attempts at 0.4%, while overall accuracy improved. On Gray Swan’s Indirect Prompt Injection benchmark, it reports that no evaluated attack succeeded. That result applies to the attacks tested, not a guarantee against every future attempt to redirect the model.
In live coding work with internal teams, Google says high-severity misalignment alerts fell almost 80%. It also reports roughly half as many unsupported claims as Argon in internal testing. The intended behavior is concrete: disclose missing evidence or blocked actions instead of inventing a result, and surface conflicting information to the user before acting.
The immediate rollout remains limited. Google says selected cyber defenders will receive access this week, followed gradually by additional trusted partners through its Fairwind Program. It is also participating in the U.S. government’s voluntary pre-release testing process. Broader access depends on further feedback and guardrail work; Google has not set a firm date.
Introductory pricing
Input: $2 per million tokens.
Output: $10 per million tokens.
Cached input: 95% off the standard input-token price.
Google says Gemini 4 Carbon will launch at $2 per million input tokens and $10 per million output tokens—half the $4/$20 rates displayed by Artificial Analysis. Those higher rates are scheduled to follow the introductory offer, VentureBeat reports, but Google has not specified when that offer ends.
The distinction comes from Google’s October 28 announcement of Carbon, its new model for demanding software engineering and cybersecurity work. Google’s announcement gives the introductory prices. VentureBeat identifies a footnote specifying subsequent…
Carbon’s advertised token prices match those of its predecessor, Gemini 4 Argon, at both stages. Their pricing calendars are not interchangeable, however. VentureBeat reports that Argon’s introductory period ends November 30, with…
The price difference is straightforward for a fixed amount of usage. VentureBeat illustrates it with 10 million uncached input tokens and 1 million billed output tokens: $30 at the introductory rates, versus…
Read full developmentHide full development⌄
Google says Gemini 4 Carbon will launch at $2 per million input tokens and $10 per million output tokens—half the $4/$20 rates displayed by Artificial Analysis. Those higher rates are scheduled to follow the introductory offer, VentureBeat reports, but Google has not specified when that offer ends.
The distinction comes from Google’s October 28 announcement of Carbon, its new model for demanding software engineering and cybersecurity work. Google’s announcement gives the introductory prices. VentureBeat identifies a footnote specifying subsequent rates of $4 per million input tokens and $20 per million output tokens. The evaluator’s profile displays those higher figures rather than the opening offer.
Carbon’s advertised token prices match those of its predecessor, Gemini 4 Argon, at both stages. Their pricing calendars are not interchangeable, however. VentureBeat reports that Argon’s introductory period ends November 30, with higher rates taking effect December 1. Carbon has no equivalent published deadline. That makes Argon’s date unsuitable for budgeting a Carbon workload.
The price difference is straightforward for a fixed amount of usage. VentureBeat illustrates it with 10 million uncached input tokens and 1 million billed output tokens: $30 at the introductory rates, versus $60 afterward. That example excludes other charges and assumes the model consumes exactly the same number of tokens. It is a rate comparison, not a prediction of a customer’s bill.
Google also advertises a 95% discount for cached inputs—previously processed prompts. Applied to the introductory input rate, that produces a price of $0.10 per million cached input tokens. VentureBeat notes that Carbon’s announcement does not specify its caching mechanics, so developers cannot assume it will behave identically to Argon’s automatic caching implementation.
Equal token rates do not establish equal spending on an assignment. Carbon’s release materials do not provide the token consumption, retries or execution time needed to compare its cost per successful task with Argon. Google pitches Carbon for sustained workflows involving many actions; its internal teams say the model has worked for days while tracing bugs and developing patches across large codebases.
Artificial Analysis supplies one narrower efficiency measure: an average cost of $7.76 per Intelligence Index task. Its profile also records 200 million output tokens across the index evaluation, compared with a displayed median of 64 million. Those figures describe the evaluator’s test workload. They are not a quote for completing a company’s engineering project or evidence that the introductory rates make Carbon cheaper per successful job.
The performance case behind the pricing is stronger than Argon’s on the general Intelligence Index. Artificial Analysis gives Carbon at high reasoning a score of 59, first among 52 models in its displayed comparison. VentureBeat places that result ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50. All four figures refer to the revised v5.0 index.
That overall lead is not a clean sweep. VentureBeat reports that Carbon improves on Argon in professional knowledge work but remains behind several rivals on AA-Briefcase. It also records a higher hallucination rate than Argon on a factual-knowledge test. These differences qualify the headline ranking: the highest composite score does not mean the best result on every component.
For now, most buyers cannot turn the advertised rates into a public trial. Google says selected cyber defenders will receive access first, followed by additional trusted Fairwind Program partners. Wider access for developers, businesses and consumers is planned after feedback and further work on safeguards. Artificial Analysis marks Carbon as having no publicly available API; Google has not set a broad-release date.
gemini 4 carbon's cybersecurity benchmarks are pretty impressive. it scores 54.8% on exploitgym, compared to 42.4% for gpt-6 astra and just 34.5% for opus 5.5 it also leads on cwe-bench and sec-bench. that's a pretty substantial lead on exploitgym in particular.
Sources
blog.googleGemini 4 Carbon: our next frontier model
Reader comments
Newest comments first. Replies stay oldest first.