Google's newest frontier model tops one influential leaderboard and leads most of Google's own tests. Almost nobody can use it yet.
Gemini 4 Carbon Arrives Behind a Security Gate
Google DeepMind published Gemini 4 Carbon on October 28 and called it the company's most capable frontier model. The post carries Koray Kavukcuoglu's name. He's a DeepMind SVP and Google's chief AI architect, and that byline tells you how seriously the company wants this one taken.
The framing is narrower than a usual flagship launch. Carbon is pitched at two jobs: network defense and long-running agent work that stretches across hours or days. Most of Google's post stays on the security work that Fairwind members already do, and treats general coding as a second theme rather than the headline.
Then there's access. Carbon is limited to cybersecurity defenders inside the Fairwind Program, a curated group that has received earlier Google models before public release. Developers, enterprises, and consumers are told to wait. Google says it will widen access "as soon as possible" once it has collected feedback and iterated on guardrails. No date was given.
So the strongest model on one leaderboard isn't for sale to the rest of us. That gap is the story. What does a top score mean when the model behind it can't be bought?
The Security Benchmarks Google Is Quoting
Google published three scores for Carbon's security work. The model reached 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. All three are vendor numbers, reported by Google, and none of them had outside verification on the day of the announcement.
Carbon also took the top spot in 15 of the 17 benchmark rows Google printed. The two losses are worth naming. OpenAI still leads on one software engineering test, and Anthropic leads on one computer-use evaluation.
Fifteen of 17 sounds crushing. It's still a table the model's maker assembled from tests it picked, and every benchmark sheet is a marketing document to some degree. Read it as a claim, not a verdict.
Where Artificial Analysis Ranked Gemini 4 Carbon
The third-party read is tighter. Artificial Analysis runs its own evaluations instead of trusting vendor sheets, and it put Carbon at 59 on Intelligence Index v5.0 with high reasoning enabled. That puts it in first place.
Look at the runners-up, because the margin is what matters here. Claude Opus 5.5 scored 58 on the same index. GPT-6 Astra took 57, and Gemini 4 Argon, the model Carbon displaces at the top, sat at 50.
Two points cover the podium. A different mix of tests could shuffle those three. Anyone chasing the absolute best number should check what the index weights, how many attempts each score allows, and whether the same settings are reachable through the API before switching vendors.
Independent testing also found trade-offs Google didn't lead with. Carbon is clearly stronger than Argon at coding and knowledge work. On some specialist tasks it trails rivals. And on tests of factual recall, it hallucinates more often than Argon did. Fewer invented facts would have been the expected direction. This went the other way.
Gemini 4 Carbon Pricing Starts at $2 and $10
API pricing is $2 per million input tokens and $10 per million output tokens. Cached input carries a 95% discount, which works out to roughly $0.10.
A footnote lists the later rate: $4 in, $20 out. Google doesn't say when the introductory window closes, which is the sort of detail that decides a migration budget. That missing date is the most expensive unknown in the post. A budget built on $2 per million tokens doubles the day the window ends, and no team wants to learn that from an invoice.
There's a precedent to lean on. Argon launched at the same $2 and $10, and its promotional pricing runs through November 30. Carbon's window probably looks similar. Google has no reason to give a newer model a longer runway than the last one got.
Fifteen Days Separate Argon and Carbon
Argon opened to everyone on October 13, and Carbon landed 15 days after that. Argon itself debuted on September 30 under the same Fairwind-first arrangement, which makes this a rhythm rather than a one-off.
Argon's documentation is the only guide to what Carbon might support. It describes shared context beyond 1 million tokens, a default output cap of 65,536 tokens, and inputs spanning text, images, video, audio, and PDFs, plus function calling, code execution, and structured output. Carbon's announcement says none of that. Whether the same limits carry over is unconfirmed, and nobody at Google has published a spec sheet for Carbon.
For anyone building on the API, that's the practical question. Context size decides which documents fit in a single call, which for legal and research teams is the whole ballgame. The output cap decides how much a single request can return. Google left both unstated, which is a strange omission for a model sold on long agent runs that need room to think.
What Google Says Gemini 4 Carbon Does Over Days
Google says internal teams ran Carbon continuously for several days. The described work is unglamorous and hard: finding bugs in large codebases, dropping approaches that fail, and writing patches that survive review.
Varun Mohan, the former Windsurf CEO who now leads Google Antigravity, said his teams use Carbon there for their hardest tasks. He described access widening through Fairwind first, then following Antigravity to a broader audience.
Long-running agents are the claim outsiders can't check. Most agent products still buckle after a few steps, when a failed tool call sends the model into a loop that burns tokens without producing a patch. A demo is easy. A model that holds its footing for days inside code it didn't write is a much bigger promise, and no external team has tested it.
The Missing Technical Report
Google shipped no technical report with Carbon. No parameter count. No word on whether the architecture differs from Argon's. For a release billed as the company's strongest, the paper trail is thin.
That's normal for a gated launch. It does mean every capability claim rests on Google's write-up plus one third-party index.
What to Watch Before Gemini 4 Carbon Opens Up
Three things will tell you more than this week's scores. Whether the guardrail work produces a public date is the first, since nothing else matters until access opens. Whether the $4 and $20 rates arrive with the wider launch. And whether outside labs reproduce the hallucination gap, or find it runs the other way.
Until then, the summary is short. Google holds the top score on one respected index, ships to a closed security group, and hasn't said when that changes. If you're budgeting an API migration, plan on the $4 and $20 rates, not the intro price.
