First on one board, not on a product

The Google blog timestamps Gemini 4 Carbon at 17:00 UTC on October 28, 2026. The page says the model builds on Gemini 4 Argon, and that Argon launched two weeks ago. Access is a separate paragraph from the boards. Starting this week, Carbon goes to a subset of cyber defenders, and then to additional trusted partners through the Fairwind program. The same paragraph says the company is taking part in the U.S. government's voluntary process for pre-release model testing. Access for developers, businesses, and consumers waits while guardrails are still being adjusted. The page uses the phrase as soon as possible. A calendar date for that access does not appear.

Each rank stays with the board that printed it. On the recently updated v5.0 of the Artificial Analysis Intelligence Index, the page says Carbon ranks first, at 59. On WebDev Arena, the page says Carbon gains 89 Elo points and lands at fourth, with a score of 1768. First and fourth name different boards. The figure 59 and the figure 1768 also name different boards. The page does not say what a score of 59 is measured against, and it does not print a maximum for that index in the sentence that announces the rank.

The cybersecurity section publishes scores. The page says those external benchmarks were evaluated without guardrails, then lists three results. SEC-bench Pro is 88.2 percent, and the page labels that result as finding vulnerabilities. ExploitGym is 54.8 percent, and the page labels that result as exploitation. CWE-bench is 71.0 percent, and the page labels that result as patching, naming Collinear as the source of the bench. The percentages are the published results. That section does not describe how exploitation or patching is carried out. An internal sentence stands apart from those three benches: more than a 2x uplift in the success rate for generating secure code, compared with Argon. The page also reports a result from the Chrome team. When generating high quality patches, Carbon did more than 10 times better than another lab's large frontier model. The other lab is unnamed. The page says Wiz observes Carbon leading on benchmarks for web penetration testing, and it presents Wiz as the company's cloud security platform. That sentence is an observation on the page. A test list is not attached.

Price sits outside the boards and carries a footnote. The introductory price matches Argon: 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input tokens are priced at 95 percent off the input token price. The footnote says that after the introductory period expires, the price becomes 4 dollars per million input tokens and 20 dollars per million output tokens. The footnote does not say which day the introductory period ends, and it does not say whether the cached-token discount continues after that change.

The safety figures come from this same blog. On Terminal-Bench 4.0, successful cheating falls to 0.0 percent, and attempts to cheat fall to 0.4 percent. The same sentence says overall accuracy improved. It does not print the accuracy figure. In live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80 percent. Unsupported claims are described as cut by approximately 2x over Argon. The page does not turn that 2x into a percentage, and this column does not convert it either. Overreach falls by 26 percent, and under-service falls by 29 percent. On the Gray Swan indirect prompt injection benchmark, the page says Argon previously recorded a 0.7 percent attack success rate and Carbon reaches 0.0 percent, resisting every evaluated attack attempt. The rate covers the attempts in that evaluation. The page does not describe how those attempts were constructed.

Long runs are an internal observation. The page says teams have seen Carbon run for days, persisting through dead ends on complex engineering tasks, including finding the cause of subtle bugs and patching them across large codebases and low-level systems. That is the company's account. The page does not record a count of days or a count of interruptions. The page also places Carbon on the same multi-layered safety stack as Argon. The list is model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. A further sentence says alignment was pushed so the model would respond to complex or high-pressure tasks in the way the page describes. That sentence does not add a new external board.

The blanks on the page can be read together. There is no end date for the introductory price. There is no opening date for access through the product surfaces. There is no name for the other lab in the Chrome comparison. The page does not say whether the score of 59 used the same unguarded setting as the three security benches. The line about evaluation without guardrails is attached to those external benches. The picture with this column is the blog header: a crystal on a blue field, with the words Gemini 4 Carbon beside it.