Two frontier models, fourteen days apart, one price tag. Google shipped Gemini 4 Argon on October 13, then announced Gemini 4 Carbon on October 28. On paper the newer model wins every category. In practice, only one of them is something a normal developer can call today, and that gap is the entire story.
Argon is globally available now through the API in AI Studio and the Gemini platform, inside Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode for Ultra and Pro tiers. Pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off. Google says it ranks first on the Vals Index, which scores models across finance, coding, legal, and tax work.
The Bracket: One Ships, One Waits
Carbon is the stronger model. Google calls it the strongest yet, with state-of-the-art cybersecurity and long-horizon capabilities, and it tops the updated Artificial Analysis Intelligence Index v5.0 at 59. It also posts 88.2 percent on SEC-bench Pro for vulnerability discovery, 54.8 percent on ExploitGym, and 71.0 percent on CWE-bench for patching. Those are not incremental numbers.
None of that is available if you are building a product this week. Carbon goes to a subset of cyber defenders first, then to more trusted partners through the Fairwind Program, widening only after Google iterates on guardrails. The company is also in the United States government's voluntary pre-release testing process. General access is promised "as soon as possible," the softest deadline in the industry.
The price is identical, which matters more than it looks. Carbon is not a premium tier you can buy into. It is a capability gate that money cannot open yet, so Argon is the only Gemini 4 you can benchmark against your own workload.
Scorecard: Five Codebases, Five Verdicts
Launch benchmarks are advertising. Merged pull requests are evidence. Google published internal engineering results for Argon that are unusually specific, and they are the best proxy for what the model does on real code.
On Chrome, Argon helped separate core webpage handling from a legacy folder of roughly 6 million lines inside Chromium's 40 million line codebase. The refactor touched hundreds of files and compressed months of work into weeks. The payoff is user-facing: live pages that move between tabs, side panels, and floating windows without reloading, removing a whole category of browser crash.
On Google's Agent Development Kit for Python, the team handed Argon a skills rulebook and a one-paragraph goal to cut framework overhead. Working autonomously overnight, it produced upstream merges that cut in-memory session preparation by 40 percent in long sessions and made cold imports 21 percent faster.
On xTB, a heavily cited open-source quantum chemistry package, Argon made structure prediction up to 2.3 times faster and conformer search 1.8 times faster, with changes contributed upstream. On Google's open Gemma models, it delivered a 2.85 times inference speedup on an A100 through vLLM, pushing Gemma 4 E2B from about 37,900 to over 108,000 tokens per second with full token parity.
Here is the scorecard, judged on shipped outcomes rather than claims:
- Real codebase surgery: Argon wins outright. A 6 million line extraction inside Chromium is not a demo.
- Performance engineering: Argon wins. Four upstream merges with measured speedups and no quality regression.
- Enterprise document work: Argon wins on evidence. Box reports 14 percent better accuracy than Gemini 3.1 Pro; Kepler says it matched analyst-built financial models on over 99 percent of figures.
- Cost per completed task: Argon wins on the only independent datapoint. Quadrillion says it finishes research evals 1.5 times faster than peer models at a quarter of the cost per task.
- Cybersecurity: Carbon wins, and it is not close. That spread is why Carbon exists as a separate release.
Where Argon Actually Loses
Being fair to Carbon means conceding what Argon cannot defend. Carbon doubles the success rate for generating secure code over Argon on internal benchmarks, jumps 89 Elo points to 1,768 and fourth place on WebDev Arena, and has run for days inside Google to root-cause subtle bugs across large codebases. Argon is a very good agentic coder. Carbon is built for tasks that outlive a coffee break and several dead ends.
Argon also hits a familiar wall on sustained autonomy. Its wins are impressive but bounded: an overnight run, a targeted refactor, a single performance goal. Google's framing of Carbon for "sustained agentic workflows" is an admission that Argon taps out where Carbon keeps going.
The Verdict
If you need a frontier model in production this month, the comparison is not close, because there is only one entrant. Argon at $2 and $10 per million tokens, with a 95 percent cached input discount, is the most defensible price-to-capability offer Google has made. Teams doing refactors, performance work, document analysis, or financial modeling should test it now, while introductory pricing holds.
If you run security operations, wait for Carbon and expect to wait. The Fairwind gate is real, the government testing is real, and 88.2 percent on SEC-bench Pro is worth the delay. But do not pretend the two are competing for the same purchase order today.
The honest summary: Google shipped a workhorse and announced a specialist two weeks later at the same price. One you can benchmark tonight. The other you can only read about. That is how launch seasons work now, with the scoreboard public long before the product is.
Comments