Google announced Gemini 4 Carbon on October 28, 2026, and described it as its strongest model yet. The announcement came with a first-place claim on a respected third-party index, an API price of $2 per million input tokens, and a catch that matters more than either: almost nobody can use it. (Source: Runtime Wire)
This is the second Gemini 4 model to arrive behind a gate. The pattern is becoming a template for how frontier labs ship models with strong security skills, and it changes what a launch day means for ordinary users. Below is what Google said, which numbers are vendor-reported, how Carbon compares with the Argon model, which became generally available two weeks earlier, and what to watch for next.
Quick summary: Google announced Gemini 4 Carbon on October 28, 2026 (some Indian outlets date it October 29). Google reports it ranks first on Artificial Analysis Intelligence Index 5.0 with a score of 59, and reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These are vendor-reported figures. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Initial access is limited to a subset of cyber defenders in the Fairwind Program, with no date for broader release.
What Google announced
Google DeepMind SVP Koray Kavukcuoglu wrote the accompanying blog post, according to Runtime Wire. Google said Carbon improves on coding and cybersecurity tasks. Initial access goes to selected defenders in the Fairwind Program, so most developers cannot call the model. (Source: Runtime Wire)
Google says it will gather feedback, keep working on guardrails and extend access to additional trusted partners, and that it is taking part in the US government's voluntary pre-release model testing. It has not said when developers, businesses or consumers will get access. (Source: LetsDataScience)
The numbers, and who produced them
Every figure below was published by Google and has not been independently verified, as far as the coverage shows. Google says Carbon ranks first on Artificial Analysis Intelligence Index 5.0 with a score of 59. The coverage I could open does not list rival scores, so I cannot tell you the margin. (Sources: Free Press Journal, LetsDataScience)
On security-related benchmarks Google reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym and 71.0% on CWE-bench (patching vulnerable code), and says Carbon produces secure code at more than twice the success rate of Gemini 4 Argon on internal benchmarks. On the coding side, the Free Press Journal reports a gain of 89 Elo points on WebDev Arena to 1769, which places it fourth. (Source: Free Press Journal)
| Item | Gemini 4 Argon | Gemini 4 Carbon |
|---|---|---|
| Announced | September 30, 2026 (Fairwind partners) | October 28, 2026 (subset of Fairwind defenders) |
| Broader access | Reported on the paid Gemini API from October 13 | No date announced |
| Introductory API price | $2 / $10 per million input / output tokens | $2 / $10 per million input / output tokens |
| Intelligence Index 5.0 | Not stated in sources | 59, ranked first (Google-reported) |
| Secure-code success | Baseline | More than 2x Argon on internal benchmarks (Google-reported) |
The Argon column is partly drawn from secondary reports: the matching introductory price and the "two weeks after Argon" framing come from the Free Press Journal (the gap fits Argon's October 13 general availability, not its September 30 announcement), while the October 13 general availability date comes from aggregator coverage that I could not trace to Google's own documentation. Treat that cell as indicative, not official. (Source: Free Press Journal)
Why a top-ranked model is still locked
Carbon is aimed at cybersecurity and long-running work, and models that can find and fix software flaws are dual-use: useful to defenders and, in the wrong hands, to attackers. That is why Google routes early access through a program for vetted defenders. This post stays at the level of what was announced and does not go into how any of this works technically.
Runtime Wire also makes a fair point about the ranking itself: Artificial Analysis' index is a broad measure of intelligence and does not evaluate cybersecurity performance, so a first-place score on it says little about the security claims. The security benchmark numbers are separate and, again, Google's own. (Source: Runtime Wire)
Price versus availability
Several sources report introductory API pricing of $2 per million input tokens and $10 per million output tokens, matching Argon's introductory rates. Argon's $2 and $10 introductory rate is reported to end November 30, rising to $4 and $20 on December 1 (Mixed News; Artificial Analysis lists $4 and $20). Google has not said whether Carbon follows the same schedule. A price with no public endpoint is a signal, not an offer. (Source: Free Press Journal)
- A first-place rank on a general index does not prove the model is best at your task.
- Vendor-reported security scores need independent replication before anyone treats them as settled.
- Argon's path from Fairwind to the paid API suggests, but does not guarantee, a similar path for Carbon.
- If you are not a security professional, nothing changes for you today.
When a lab announces a model you cannot use, look for four things before getting excited: a public model ID, a general-availability date, an independent benchmark, and a stated price after the intro period. Carbon has a price and a rank, but not the other three.
01When was Gemini 4 Carbon announced?
October 28, 2026 in most coverage. The Free Press Journal's article is dated October 29, likely a publication-date or time-zone difference.
02Who can use Carbon right now?
A subset of cybersecurity defenders in Google's Fairwind Program. Google has not said when developers, businesses or consumers will get access.
03What does Carbon cost?
Introductory API pricing is reported at $2 per million input tokens and $10 per million output tokens, the same as Argon's introductory rates. Argon's introductory rate is reported to end November 30, with $4 and $20 after that. Carbon's schedule is unconfirmed.
04Is Carbon really the top AI model?
Google reports first place on Artificial Analysis Intelligence Index 5.0 with a score of 59. That is a vendor claim about one index, and the index does not evaluate cybersecurity performance.
05Why is access restricted?
Models strong at finding and fixing software flaws are dual-use, so Google is starting with vetted defenders while it tests safeguards.
Locked models are hard to evaluate, but the ones you can use are easy to compare. LumiChats puts 40+ premium AI models in one subscription, so you can test the same prompt across several of them.
