AI
“Google just unveiled Gemini 4 Carbon, its most powerful AI model yet, with a major focus on cybersecurity." (plus caption details: #1 on Artificial Analysis Intelligence Index at 59; 88.2% SEC-bench Pro; 54.8% ExploitGym; 71% CWE-bench; more than doubled predecessor's secure-code success rate internally; can work autonomously for days;…”
Plain restatementOn 28 October 2026 Google announced a new frontier model called Gemini 4 Carbon, described by Google as its strongest model, with cybersecurity and long-horizon agentic work as the headline capabilities. The post further states a set of benchmark scores, a staged rollout beginning with selected cyber defenders, and introductory API pricing of $2 and $10 per million input and output tokens.
This one checks out in its main points. Google really did announce a model called Gemini 4 Carbon on 28 October 2026, really does call it its strongest model, and really does pitch it on cybersecurity, and the quoted scores all appear in Google's own announcement. The claim that it ranks first on the Artificial Analysis Intelligence Index with a score of 59 is confirmed by that evaluator independently, and its board still showed Carbon in first place as of 30 October 2026. Two pieces of context are missing from the post. Reporting says Google ran the three cybersecurity tests with the model's safety guardrails switched off, and the $2 and $10 prices are an introductory rate that is reported to double afterward. The post also lists only good news: the same independent evaluator found Carbon leads just four of the index's nine tests and makes more factual errors than the previous Gemini model on one knowledge test. The security scores, the claim about working for days on its own, and the doubled secure-code result all come from Google's own testing, and because the model is not publicly available, nobody outside Google can check them yet.
Google [drifted from the evidence:] just unveiled Gemini 4 Carbon, its [drifted from the evidence:] most powerful AI model [drifted from the evidence:] yet, with [drifted from the evidence:] a major focus on cybersecurity." [drifted from the evidence:] (plus caption details: #1 on Artificial Analysis Intelligence Index at 59; 88.2% SEC-bench Pro; 54.8% ExploitGym; 71% CWE-bench; more than doubled predecessor's secure-code success rate internally; can work [drifted from the evidence:] autonomously for days; resisted 100% of [drifted from the evidence:] evaluated prompt injection attacks in Gray Swan's benchmark; [drifted from the evidence:] access this week limited to select cyber defenders; pricing [drifted from the evidence:] starts at $2 [drifted from the evidence:] per million input tokens and $10 per million output tokens)
[added by the neutral restatement:] On 28 October 2026 Google [added by the neutral restatement:] announced a new frontier model called Gemini 4 Carbon, [added by the neutral restatement:] described by Google as its [added by the neutral restatement:] strongest model, with cybersecurity [added by the neutral restatement:] and long-horizon agentic work [added by the neutral restatement:] as the headline capabilities. The post further states a set of benchmark [added by the neutral restatement:] scores, a staged rollout beginning with selected cyber defenders, [added by the neutral restatement:] and introductory API pricing [added by the neutral restatement:] of $2 and $10 per million [added by the neutral restatement:] input and output tokens.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Google did announce Gemini 4 Carbon on 28 October 2026 through its official blog and official DeepMind accounts. This is not a rumor, a leak, or a fabrication.
- Google describes Carbon as its strongest model, and frames cybersecurity and long-horizon agentic work as its headline strengths. The post's characterization matches Google's own framing.
- The #1 ranking at 59 on the Artificial Analysis Intelligence Index v5.0 is confirmed by the independent evaluator itself, and the live board still shows Carbon first at 59 as of 30 October 2026.
- The three security figures, 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, appear in Google's announcement with the same descriptions the post gives for each test.
- The staged access description is correct: Google said access would begin that week with a subset of cyber defenders and expand through its Fairwind Program. The post does not present the model as generally available.
- The introductory API price of $2 per million input tokens and $10 per million output tokens matches what Google published.
- Google did report a 0.0% attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark, which is the same thing as resisting all evaluated attacks in that benchmark.
- The post attributes its source to Google and names Google's methodology page, so a reader is told where the numbers come from.
- Named outlets report that Google ran the three cybersecurity benchmarks without safety guardrails, and that the ExploitGym figure corresponds to a ten-hour runtime budget. The post gives the three numbers with no indication that they were produced in an unguarded configuration, which is not the configuration a user of the shipped model would be operating.
- The security scores, the secure-code doubling and the prompt injection result are all Google-run or Google-internal. The post labels its source as Google, which is honest, but it presents these numbers in the same flat list as the independently confirmed index ranking, so a reader cannot tell which figures were checked by an outside party and which were not.
- The post's list reports only favorable results. The independent evaluator that confirmed the #1 ranking also reported that Carbon leads only 4 of the 9 index evaluations, is not leading in coding and knowledge work generally, and that its hallucination rate on AA-Omniscience rose from 4% for Argon to 11% for Carbon. None of that appears in the caption, and the caption instead states that unsupported claims fell, using Google's internal measure.
- The $2 and $10 figures are introductory rates that multiple reports say revert to $4 and $20. "Pricing will start at" is literally correct and does signal a starting point, but a reader is not told that the standard rate is double.
- The one-point lead over Claude Opus 5.5 at 58 comes with a large difference in verbosity. Artificial Analysis reported Carbon consuming roughly 200 million output tokens for a full index run against a median of 64 million, which bears directly on what the ranking costs to achieve.
- Whether the cybersecurity scores hold under ordinary guarded conditions is unknown. No independent party has run SEC-bench Pro, ExploitGym or CWE-bench against this model, and with no public endpoint or weights, no one outside Google can.
- The claim that the model can work autonomously for days on complex engineering tasks rests on Google's internal observation. No external demonstration artifact or third-party reproduction was located.
- The claim that secure-code success more than doubled over Argon is a Google internal benchmark result with no published methodology located.
- The Gray Swan result is Google-reported and specific to indirect prompt injection in the evaluated set. Whether Gray Swan itself has published a corresponding independent result was not established in this check.
- No model card or technical report for Carbon was located as of 30 October 2026, so the safety evaluation detail behind the launch is not publicly checkable. Google states it is participating in the US government's voluntary pre-release testing process, but no output of that process is public.
- The durability of the #1 ranking is inherently short-lived. Index v5.0 launched one day before the announcement, and the lead is one point.
Google's own announcement post of 28 October 2026 exists and says what the post says it says. Google describes Gemini 4 Carbon as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities, states that Carbon ranks first on the updated v5.0 of the Artificial Analysis Intelligence Index at 59, and reports leads on SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation, and CWE-bench at 71.0% for patching. The same post states that starting that week Google will make Carbon available to a subset of cyber defenders and will phase access to additional trusted partners through its Fairwind Program, and that it is engaged in the United States government's voluntary process for pre-release model testing. Artificial Analysis, an independent evaluator, separately published its own result on 28 October and its live Intelligence Index page as of 30 October 2026 shows Gemini 4 Carbon (High) at 59, first, ahead of Claude Opus 5.5 at 58. So the index ranking is not merely a Google assertion. Artificial Analysis adds detail the post does not carry: Carbon holds the top score in 4 of the 9 index evaluations, is not leading in coding and knowledge work generally, and its hallucination rate on AA-Omniscience rose from 4% for Argon to 11% for Carbon even as accuracy improved. The cybersecurity scores are Google-run. Named outlets report that Google ran those security evaluations without safety guardrails, and that the 54.8% ExploitGym figure corresponds to a ten-hour runtime budget. The Gray Swan figure is Google-reported: a 0.0% attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark, against 0.7% for Argon. The secure-code doubling and the days-long autonomous operation are Google internal observations. Reporting also notes the model had no public API endpoint, model identifier, or model card at launch, with the Gemini API model list still showing Argon.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did Google really announce a new AI model called Gemini 4 Carbon?
Yes. Google announced Gemini 4 Carbon on 28 October 2026 through its official blog and DeepMind accounts, describing it as its strongest model with cybersecurity and long-horizon agentic work as headline strengths.
Is the claim that it ranks #1 on the Artificial Analysis Intelligence Index confirmed?
Yes, this part is independently confirmed. Artificial Analysis, which is separate from Google, published its own result showing Carbon first at a score of 59, and its live board still showed Carbon in first place as of 30 October 2026.
Are the cybersecurity benchmark scores trustworthy as stated?
The scores themselves are accurately quoted from Google's announcement, but reporting says Google ran those three cybersecurity tests with the model's safety guardrails turned off, which is not how a user would normally run the shipped model. No independent party has reproduced these results.
Does the post leave out any unfavorable information?
Yes. The same independent evaluator that confirmed the top ranking also found Carbon leads only 4 of 9 index tests, is not ahead in coding or general knowledge work, and has a higher hallucination rate than its predecessor on one knowledge test. None of this appears in the original post.
Is the $2 and $10 pricing the full story?
The figures match what Google published, but multiple reports say these are introductory rates that are expected to double afterward, and the post does not mention this.