Google spent much of 2026 playing catch-up in the AI race. A month and two frontier models later, it is sitting at the top of the charts again — sort of.
Gemini 4 Carbon has the numbers to back up the comeback talk. But like Argon at launch, the public can't use it yet, and with OpenAI and Anthropic shipping at a similar pace, the crown might not last long.
Google unveiled Gemini 4 Carbon on October 28, the same day Alphabet announced its third-quarter earnings. The model is built for cybersecurity and long-running coding work, and it tops 15 of 17 benchmarks on Google's own charts. Independent evaluator Artificial Analysis places Carbon, using high reasoning, first on its Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50.
Carbon tops 15 of 17 Google benchmarks and leads Artificial Analysis at 59 points, but the public cannot call it yet.
Access, for now, is limited to selected cybersecurity defenders in Google's curated Fairwind program. Google says a broader release will come "as soon as possible" after it gathers feedback and iterates on guardrails. A request for gemini-4-carbon on October 30 returned HTTP 503. Argon, which Google opened more broadly on October 13, remains the model customers can actually evaluate.
The timing requires distinguishing Argon's announcement from its public launch. Google introduced Argon on September 30, initially giving approved cybersecurity partners access through Fairwind; its wider October 13 rollout followed. Carbon arrives 15 days after that expansion.
Argon already targeted demanding engineering, professional research and security assignments. Google described agents assisting with memory optimization and C/C++ migrations to Rust, with sensitive code changes undergoing testing and human review. Cybersecurity was central too: Argon was designed to discover flaws, validate their significance and produce fixes.
The big distinction between Carbon and Argon is Google's reported improvement on those existing domains, particularly sustained engineering work, exploitation testing and agent reliability. Google didn't release a technical report, nor say whether Carbon uses a new model architecture. Nor did it disclose a parameter count.
"Internally, our teams have seen Carbon run for days," Google wrote in the announcement, authored by Koray Kavukcuoglu, Google DeepMind SVP and Google's chief AI architect. The company describes the model continuing through unsuccessful approaches while diagnosing bugs and developing patches in substantial codebases.
Where a run lasts days rather than seconds, Carbon is not marginally ahead — it is in another weight class. On GraphWalks, spanning 256k to 1M tokens of context, Carbon scores 96.2% against Opus's 66.8% and Astra's 71.8%. On Vending-Bench 2, which simulates long-horizon autonomous operation, Carbon's final account balance reached $16,484 against Opus's $9,235 — roughly 1.8 times the outcome.
The crown is real, and it is already partitioned. Carbon leads four of the nine evaluations in AA's Intelligence Index v5.0 outright — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, AutomationBench-AA — and does not lead the other five. On the Coding Agent Index, Claude Code with Opus 5.5 at max effort scores 62 against Carbon's 59 running Google's own Antigravity CLI. Three points is not a rounding error; it is Anthropic's moat holding under direct assault.
Google's own comparison chart concedes more. On FrontierSWE 2, GPT-6 Astra leads at 65.5% and Opus 5.5 sits at 62.3%, ahead of Carbon's 61.2%. On the offline subset of OSWorld-2.1, Opus holds first at 87.2% against Carbon's 81.5%. In pure software engineering and computer operation, the model that supposedly lost the crown is still wearing it.
Then there is the strangest number in the release, and the one buyers should internalize. Carbon takes 97.94 seconds to first token. Argon takes 47.20. Carbon starts every reply twice as slowly — and finishes a complete coding-agent task in 29.8 minutes against Argon's 34.7, at $7.74 per task against $11.73.
Sit with that triad. Slower to begin, faster to finish, cheaper per completed job. The metric most developers watch — time to first token — has become actively misleading for agentic work. A model that thinks longer before speaking and then does more per turn will look worse on every dashboard built for chat and better on every dashboard built for delegation.
The cost advantage is structural, not promotional. Carbon consumes 123,000 output tokens per Intelligence Index task; Opus 5.5 at max effort consumes 173,000 — 29% more. At standard pricing that is $7.76 per task against $10.04, roughly 23% cheaper for comparable overall intelligence. Opus is not expensive because Anthropic prices it high. It is expensive because it talks more.
This matters because the price-war reading of Carbon is wrong. Yes, Google is offering the same $2/$10 introductory rate Argon launched with, and a 95% discount on cached input. But Artificial Analysis ran every chart at standard pricing — $4/$20, identical to Argon's standard rate and to Opus's. At those rates Carbon delivers a nine-point index gain over Argon at essentially the same cost per task: $7.76 versus $7.68.
Google is not buying the crown with price. It is buying the right to define which workloads the frontier is judged on.
Consider what Google chose to lead with: cybersecurity. On SEC-bench Pro, Carbon hits 88.2% against Argon's 78.8% and Astra's 85.4%. On ExploitGym it reaches 54.8% where Argon managed 35.1%. On CWE-bench it scores 71.0% against Opus's 67.0%. The Chrome team reports Carbon generating high-quality security patches more than ten times better than another top lab's frontier model.
That is a workload selection, made deliberately, in a domain where long-horizon execution and million-token context are the binding constraints — precisely where Carbon's architecture wins and Opus's doesn't.
Google also claims progress on agent misbehavior. On Terminal-Bench 4.0, the company says Carbon cut the cheating success rate to 0.0% and the attempt rate to 0.4% while accuracy improved, and that high-severity misalignment alerts in internal coding environments fell by roughly 80%. On the Gray Swan IPI benchmark, Carbon recorded a 0.0% attack success rate against indirect prompt injection, after Argon's 0.7%.
One caveat sits inside the AA-Omniscience result. Carbon's accuracy rose 10 points over Argon to 58%, and its score of 54 leads the index — but its hallucination rate rose from 4% to 11%. That is still the second lowest of any model, behind only Argon, compared with 50% for both Claude Opus 5.5 and GPT-6 Astra.
The strategic question for every lab now is not who holds the crown. It is which workloads a lab is willing to lose. Anthropic appears willing to lose long-horizon science and security in exchange for coding-agent supremacy. Google appears willing to lose interactive latency and SaaS automation — Carbon and Argon tie at 77.8% and 77.5% on AutomationBench — in exchange for agentic endurance.
OpenAI, notably, is absent from this fight's headline. GPT-6 Astra sits third on the index at 57 and still owns FrontierSWE 2, but it leads nothing that either competitor is contesting.
One date should anchor the comparison. Argon's introductory pricing expires November 30, when it doubles to $4/$20. Carbon's introductory period has no announced end. When both models sit at standard pricing — which is where Artificial Analysis already measured them — the apples-to-apples moment arrives, and Carbon's nine-point index advantage over Argon costs the same per task.
On Alphabet's earnings call, Sundar Pichai framed Carbon as the foundation for a broader shift toward agents. "When you see the new model's ability to automate and execute complex tasks at high quality, it's clear that the agentic capabilities everyone imagined are becoming real," he said. He added that Google is distilling Carbon to deliver an agent experience to about 1 billion users, and that agents will serve as "a unifying layer across our products," extending beyond search and the Gemini app to Chrome, Android and Workspace.
That transition has also exposed a slowdown in Google's traditional search advertising revenue. Alphabet's search advertising and related businesses grew about 12 percent year-on-year in the third quarter — down 5 percentage points from the 17 percent growth recorded in the previous quarter. Chief Business Officer Philipp Schindler attributed the slowdown to exchange rate effects as well as the integration of AI features into search, and said Google is expanding ads in AI search results to more users and queries.
So the crown changed hands by a single point, at a price nobody is currently paying, for a model nobody can currently call. That sounds like a hollow victory until you look at the shape of the win. Carbon loses coding agents by three points and wins long-horizon execution by thirty. It starts twice as slowly and finishes 14% faster for 34% less money.
The composite index says 59 to 58. The workload split says the frontier now has two thrones, and Google just chose which one it wants.