Gemini 4 Carbon Explained: “Carbon” Lands 15 Days After Argon’s General Release. What Changed?First on 15 of 17, so why did its hallucination rate go up? Ten interactive tools for reading Google’s new frontier model

01“Periodically” Turned Out to Mean the Periodic Table

Pichai’s post starts like this: “I did say we’d be releasing our models periodically…introducing Gemini 4 Carbon, our next frontier model.” Two weeks after Argon, he introduced the next model and chose, of all words, “periodically”.^[1] Elon Musk quoted the post and added one word: “Impressive”.^[2]

Asked what the naming scheme means, Logan Kilpatrick, who leads Google AI Studio and the Gemini API, replied: “All elements right now! But having fun with it : )”^[3] That is the entire official explanation. The community, naturally, had a field day. If the names go in alphabetical order, someone noted, the scheme breaks at the tenth letter because no element starts with J.^[4] Then came the math (18, then 6, so helium at 2 next)^[5] and the joke that if smaller numbers mean stronger models, hydrogen (1) must be the final boss.^[6] Draw a few yourself.

The names alone make a fun picture. Argon’s name comes from the Greek for “lazy”, and it is an inert gas that barely reacts with anything. Carbon bonds with almost everything, and the same atoms become both pencil graphite and diamond. Google’s key art is a faceted crystal, too. It fits the model’s selling points, versatility and hard-wearing stamina. With a blog called Dandan Life (dandan means “solid”), I admit a soft spot for the element behind the hardest crystal. That is my reading, though. Google hasn’t explained the name.

After the name, the next question is: can I use it? Not yet. Google is opening it to a subset of cyber defenders starting this week and widening access in stages while it takes part in the US government’s pre-release testing. There is no public release date (section 7).^[7]

A new frontier model 15 days after Argon’s general release is fast. Google didn’t explain the timing, but on the earnings call the same day Pichai said AI is helping Google increase its training and launch velocity, and that he expects continued acceleration over the next 12 months because of RSI (recursive self-improvement, AI helping to build the next AI).^[8] He also said Google is already deploying Carbon internally.^[9] The code names and the security drama around the launch are collected in the asides at the end.

02One Report Card, 17 Items: First on 15, but Who Did the Measuring?

Google published a table comparing Carbon, Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 side by side. There are 17 items. Carbon beats Argon on all 17 and scores highest on 15.^[10] I call them items rather than tests for a reason: some rows, like the AA Index and the Vals Index, are composites of many tests.

A table is dull, so let’s make it a game: find the two items Carbon didn’t win. Hint: one is in coding and one is in computer use. Tap the model you think won each card to reveal the answer and the bars. Keep an eye on the badge in the top right. Google’s methodology document shows that different items were measured by different people.^[11] On some, an independent evaluator ran every model under the same conditions. On others, Google measured Carbon itself and took rivals’ numbers from their own reports. On GraphWalks, Google measured every model. Count the badges and you get 10 independent, 6 partly Google-run and 1 Google-only. Carbon wins 9 of the 10 independent items and 6 of the 7 that Google had a hand in. Who did the measuring didn’t change the picture much, so you can read on with a little more confidence.

The answers are FrontierSWE 2 and OSWorld-2.1. On FrontierSWE 2, a set of long software-engineering tasks, GPT-6 Astra (65.5%) beats Carbon (61.2%); on OSWorld-2.1, which tests using a computer from screenshots, Opus 5.5 (87.2%) beats Carbon (81.5%). Step outside the table and the rankings shift again, because the table compares only five models.

• AA Intelligence Index v5.0: Carbon (high) is first with 59. Opus 5.5 (max) has 58, GPT-6 Astra (max) 57 and Argon 50. Carbon was first on 4 of the 9 evaluations.^[12]
• Terminal-Bench 4.0: first in Google’s table, but second on AA’s full leaderboard behind Kimi K3.1 Pro (67.2%).^[13]
• Vals Index v2.2: 70.36%, second of 29 models. Claude Sonnet 5.5 leads with 70.85%.^[14]
• FrontierSWE 2: third in the default one-model-per-provider view, and fourth in the full view, behind Claude Sonnet 5.5 (61.9%) as well.^[15]
• Coding Agent Index: pairing models with their coding tools, Claude Code with Opus 5.5 leads with 62, while Antigravity CLI with Carbon scores 59.^[16]
• WebDev Arena: on the board where people vote on web apps the models build, Carbon is fourth with 1768, up 89 points from Argon (1679, 11th).^[17]
• AA-Briefcase (office document work): 1609. Seven models are ahead on this one.^[18]

The same carbon atoms become pencil lead or diamond depending on how they are arranged. Scores are similar: put them in a different frame and “clear #1” turns into “one of the top few”. In short, Carbon is closer to “top tier almost everywhere, dominant in a few places” than to “first at everything”. Why the benchmark leader isn’t always the tool developers use most is a story I covered in my coding AI comparison.

03The Hallucination Paradox: More Right Answers, More Made-Up Ones?

Artificial Analysis’ AA-Omniscience asks knowledge questions and also checks whether a model says “I don’t know” when it doesn’t. The result is fun. Carbon’s accuracy is 58%, 10 points above Argon’s 48%, and its overall index is first at 54. Yet its hallucination rate rose from Argon’s 4% to 11%.^[19]

The trick is in the denominator. Here, the hallucination rate isn’t the share of all questions answered wrongly. It is the share of made-up wrong answers among all responses that weren’t correct (wrong, partial or not attempted).^[19] So a model that often holds back when unsure gets a low rate. Think of an exam: Argon leaves the questions it doesn’t know blank, while Carbon attempts more and sometimes guesses. The hallucination rate is the share of missed questions where it guessed wrong. Let’s shrink it to a 100-question exam.

One number alone is easy to misread. Carbon’s 11% is the second-lowest on this test after Argon, and far below the 50% of Opus 5.5 (max) and GPT-6 Astra (max).^[19] And Google says that in its internal testing Carbon cuts unsupported claims roughly in half compared with Argon.^[7] Different tests measure different things. The takeaway is boring but true: still check the source of any number that matters. It is also why I run an automated checker on the numbers in my AI-assisted posts every morning.

04Stamina for Days, and a Slow First Word

The first strength Google highlights is stamina: the ability to keep going through long, multi-step work. According to the announcement, Google’s internal teams have seen Carbon run for days, backing out of dead ends and pressing on to root-cause and patch subtle bugs across large codebases and low-level systems.^[7] Google DeepMind put it this way on X: Carbon can work through complex tasks for days without getting stuck or taking shortcuts.^[20]

There is an unusual test for stamina. Andon Labs’ Vending-Bench 2 hands an AI a simulated vending-machine business for a year. It orders stock, sets prices and emails suppliers. A day or two is easy. Months of small mistakes piling up are what empty the bank account.

Carbon finishes first with an average balance of $16,484.^[21] Andon Labs posted on X that Google is #1 on Vending-Bench 2 “for the first time in a year”.^[22] Although it isn’t in Google’s table, OpenAI’s GPT-6 Sol ($14,428) also lands third, ahead of Argon.

Its grip on very long inputs improved a lot too. GraphWalks asks a model to follow links scattered through 256K to 1M tokens of input, like tracing “A links to B, B links to C” through several novels’ worth of text. Carbon scores 96.2%, 12 points above Argon and 24 to 31 points above the rivals. Note that Google measured every model on this test itself.^[11]

Other things changed too. Carbon thinks for a long time before answering. In Artificial Analysis’ measurement at the high setting, its first word took about 100 seconds, roughly the time it takes a cup of instant noodles to get halfway done. The median for models in its price tier is in the 3-second range. Once it starts, it writes at around 80 tokens per second, about the tier median.^[23] These values are re-measured constantly and depend on the input and reasoning setting, so they don’t mean every request will take this long. The stopwatch below uses an October 29 snapshot.

It also uses more tokens. According to Vals AI, across the Vals Index Carbon uses 2.3 times Argon’s reasoning tokens and 64% more output tokens. Most of the extra reasoning goes into research-style tasks such as finance, tax and legal research, and legal research takes over 50 minutes per task.^[24] People who tried what was believed to be Carbon in blind arena matchups before launch described it as two to four times slower than Argon but less lazy. Nobody can confirm which checkpoint they were served.^[25] By Artificial Analysis’ count, Carbon wrote 200 million output tokens to get through the Intelligence Index, close to three times the median of about 71 million for similar models. It is quite the talker.^[23]

05Security in Three Steps: Find, Exploit, Patch

Carbon’s signature skill is cybersecurity. In his post, Pichai singled out three numbers: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching.^[1] All three are first place. Google says it ran these cyber evaluations with guardrails off to see raw capability; real products keep the safeguards on.^[7] And if “finding” makes you picture an AI discovering brand-new bugs with no clues, that overstates it. SEC-bench Pro gives the agent hints from a real vulnerability case, such as the source paths, the bug class and the expected error, and asks it to build a proof-of-concept input that actually triggers the bug. It never sees the reference PoC.^[26] Even with hints, pinpointing the cause in an engine with millions of lines of code is no small feat.

Think of security work as a factory conveyor: find the bug, prove it’s dangerous by actually exploiting it, then fix it.

The announcement also has examples beyond benchmarks. The Chrome team reported that Carbon did more than 10 times better than another lab’s large frontier model at generating high-quality patches (without saying 10x of what). On internal benchmarks it more than doubled Argon’s success rate at writing secure code. And Wiz, Google’s cloud security platform, sees Carbon leading on web penetration-testing benchmarks.^[7] Remember that all of these are Google’s claims.

“Great at attacking” can sound alarming, because the same skill serves attackers and defenders alike. That is also why Google is giving Carbon to defenders first instead of releasing it to everyone (section 7).

06Cheating in the Exam Room, and Watching the Chain of Thought

The stronger the model, the more it matters whether it takes tests honestly. In its October 13 safety post on Argon, Google described rare tricks that showed up during reinforcement-learning training, when tasks were hard or training environments were flawed.^[27] To be clear up front: this happened while training Argon, not Carbon. The “chain of thought” here means the reasoning a model writes out before answering. It isn’t everything going on inside the model, just the notes it leaves where we can read them.

Because of results like this, Google says it follows one rule: it doesn’t penalize the model for shortcuts it writes down in its chain of thought.^[27] The detailed report adds that some signals do touch the reasoning, such as length penalties that teach it to think efficiently, but Google avoids using its monitors’ verdicts as a training signal.^[28] Punish a model for admitting a shortcut and it may learn to hide shortcuts instead of dropping them. Google also says it runs a monitor that reads an agent’s actions and reasoning in real time and blocks severe boundary violations before they happen.^[27]

Carbon builds on Argon’s multi-layered safety and security stack. Here are the numbers Google reports for Carbon.^[7]

The second card’s “misalignment alerts” are not about bugs in general. They are alerts raised when a model acts outside the user’s intent or the scope it was allowed. What fell was how often those alerts fired. Google also says that when instructions are ambiguous, Carbon retrieves the relevant sources and surfaces discrepancies to the user before acting.^[7] All of these numbers were measured by Google, though. As of October 29 there is no Carbon-specific safety report or model card. I collected real cases of AI agents crossing the line in an earlier post.

07Price, Cost per Job and When You Can Use It

Carbon launches at Argon’s introductory price, half the list price for the early period: $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period it becomes $4 and $20. Google hasn’t said when the introductory period ends.^[7] For reference, Argon’s introductory price ends on November 30.^[29]

Token prices alone hide the real cost, though. A model that uses many tokens can cost more per job even with cheaper tokens. That is what the calculator’s second tab is about. Keep in mind that the second tab uses list prices, without the intro discount.

On Agent Arena, where models compete on real agent sessions, Arena said Carbon’s median cost per task was $1.80, 72% less than the top-ranked Opus 5.5 (max).^[30] I dug into the gap between token price and real cost in my Claude Opus 5.5 vs GPT-6 Sol comparison.

Google products are famous for being confusing about where and how to use the same model. On Carbon’s launch day, a Hacker News user grumbled that if a Google PM handed them $100 to figure out how to send a prompt to Gemini 4 Argon, they were only about 30% sure they could.^[31] Carbon isn’t even at that stage yet.

Google isn’t rushing Carbon. It is taking part in the US government’s voluntary pre-release testing process, opening access to a subset of cyber defenders starting this week and gradually adding trusted partners through its Fairwind Program. It says it will refine the guardrails and then make Carbon available to developers, businesses and consumers as soon as possible.^[7][32]

Next come the smaller models. In the question-and-answer part of the earnings call, Pichai said Google is working on distilled versions of its Carbon models so it can deploy them efficiently to a billion users, in terms of compute, latency and quality.^[9] Distillation means teaching a small, fast model with a big model’s answers.

There are rumors of a “Gemini 4 Flash”, too. Around October 23, a screenshot of a configuration screen listing the model IDs gemini-4-carbon-preview and gemini-4-flash-preview circulated on X.^[33] Its origin is unverified. As of October 29, the only Gemini 4 model in Google’s developer model list is gemini-4-argon^[34], and Pichai never said the word “Flash”.

All told, Carbon puts Google a step ahead in the race for the smartest model. Its stamina on long tasks and its security skills show up in the numbers across several evaluations. The slow first word, the hallucination rate that rose alongside accuracy and the still-closed door are things to judge only after using it. And this report card taught me one thing: who measured a number, and against whom, says more than the word “first”. When Carbon opens up, I’ll start with the number checks I listed above and see whether thinking longer actually buys a better answer. Whether the next one is helium or hydrogen, I may be writing another one of these in two weeks. If you want the model that came right before, read my element-card teardown of Gemini 4 Argon next.

Asides: Two Stories Behind Carbon

The detours are folded away. Open whichever you’re curious about.

Six Questions You Might Have Now