Who Can Use Gemini 4 Carbon

What Google shipped on October 28

Google announced Gemini 4 Carbon as its strongest model, aimed at cybersecurity work and long coding jobs. The post says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. That figure is Google's citation of Artificial Analysis. Carbon is not a directory listing, so the leaderboard does not print Google's 59. This directory's own snapshot is Intelligence Index v5, taken November 3, 2026. On that snapshot Claude Opus 5.5 scores 58.4 and ranks second among the models the index scored.

The introductory API price matches Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Google says cached input is 95 percent off the input price during that introductory period. After the introductory period, the post's footnote sets the price at $4 per million input tokens and $20 per million output tokens.

Carbon is not in this directory. These are Google's published claims. They are not a CuratedAI test of the model, the index, or the bill.

Access this week is a defender program, not a chat toggle

Google says a small set of cyber defenders gets Carbon this week, with more trusted partners added later through the Fairwind program. The same post says Google is in the US government's voluntary pre-release testing process, and that developers, businesses, and consumers do not have Carbon on product surfaces yet.

Read the program page on its own. As of October 29 it still describes early access to Gemini 4 Argon for vetted partners, including governments, critical-infrastructure operators, and core technology platforms. It does not yet list Carbon. Google also says partners may not share or resell that access. An application form is not a login.

Read the cyber scores as the lab's own account

Google says Carbon leads published tests for finding software flaws and writing patches, and that on an internal check it writes secure code at more than twice the rate of Gemini 4 Argon. The post names three outside benchmarks, all evaluated without guardrails, and gives these scores: 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym, and 71.0 percent on CWE-bench. A separate line says the Chrome team found Carbon more than ten times better than another lab's frontier model at writing high-quality patches. That is Google reporting the Chrome team's result.

On coding outside security, Google says Carbon gained 89 Elo on WebDev Arena and sits at 1,768, fourth on that board. The post also says internal teams have run it for days on large debugging jobs. None of those numbers was remeasured here. This page does not describe how those security tests are run, and it is not a guide to finding or exploiting flaws.

Google's safety notes are also its own. It says successful cheating on Terminal-Bench 4.0 fell to 0.0 percent, with attempts at 0.4 percent, and that high-severity misalignment alerts in internal coding dropped by almost 80 percent. It says unsupported claims fell by about half versus Argon, and that Carbon scored 0.0 percent attack success on Gray Swan's indirect prompt-injection benchmark, where Google had put Argon at 0.7 percent. Treat that as a lab report, not a reason to skip your own review.

Match the job to something you can open

Carbon is the wrong plan for any task you need to run today, unless Google has already put you in the defender group. The other model news from the same day is already usable.

Where the index line is easy to overread

A rank of first on an index you cannot open yet does not move the model you can call this morning. Google's 59 is their citation of Artificial Analysis v5.0. This directory's snapshot is now that same index version, taken November 3, 2026. Do not paste Google's rounded 59 onto the leaderboard. Use the score the snapshot records for a model that is actually listed.

GitHub's own coding note is separate, and it is also theirs. On the Fusion changelog, GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Opus 5.5 in Claude Code at the Max setting, at about half the average API cost per task and about 40 percent less time. That is a router score, not a Carbon score.

The same briefing covers a speech model with a precise name. TechCrunch and the Hugging Face repo call it Nemotron 3.5 Transcribe, a 30 billion parameter model Nvidia says can run on one chip. It is not a listing here yet. The other stories from the day, including the funding and policy items, are in the October 28 briefing.

What to do this week

If you already build or fix software, turn on GitHub Copilot Fusion and keep Gemini 4 Argon for the Gemini work you can run. Apply to Fairwind only if your organization is the sort of defender Google describes. Wait for a public Carbon release before you plan a product on it. For the rest of the day, read the October 28 briefing. Reviewed October 29, 2026, from Google's Gemini 4 Carbon post, the Fairwind page, GitHub's Fusion changelog, and TechCrunch's note on Nemotron 3.5 Transcribe. The index comparison was checked again on November 3, 2026, against this directory's Intelligence Index v5 snapshot.