On a Tuesday in November, Google opened its most capable model, Gemini 4 Carbon, beyond the small circle of cybersecurity firms that had tested it for two weeks, placing it before paying developers and top-tier subscribers. It is a familiar human pattern: a powerful tool first guarded, then shared, with the terms of sharing written into the price itself. The introductory rate expires at the end of November, and the cost of thinking harder is quietly being set in public view. What remains withheld, the cyber edition that requires an application, reminds us that the most dangerous capabilities a
Google Opens Gemini 4 Carbon to Paying Developers, With Introductory Pricing Set to Double After Nov. 30
The most sensitive capability stays gated even as the general model goes wide.
Mark
So what actually changed on Tuesday?
Mimi
Carbon went from a two-week trial with a few cybersecurity firms to general availability for paying developers through the API, plus AI Ultra subscribers in the Gemini app, AI Mode and Antigravity.
Mark
And the price?
Mimi
$2 per million input tokens and $10 per million output, matching Argon. But that's introductory. After Nov. 30, Google says it doubles.
Luke
Worth noting that's Google's statement of its own plans. Nothing in the reporting says what the post-November rate will look like in practice beyond "doubles."
Mark
Why does the token usage matter?
Mimi
Google says Carbon reasons and verifies repeatedly, so it consumes more tokens per task. The same per-token price can mean a bigger bill.
Luke
Which sits awkwardly beside the claim of cost per task more than 2X lower than rivals. That's Google's figure, and we don't see the method.
Mark
The Wiz zero-day result sounds dramatic.
Mimi
Two large software vendors, vulnerabilities found in under two hours, then patched. Google says that usually takes weeks.
Luke
But Wiz used the cyber edition, not the model released this week, and Wiz is owned by Google. The vendors aren't named and the result isn't independently verified here.
Mark
And the benchmarks?
Mimi
Top of CursorBench 4.0, WANDR and AutomationBench, second on the Multimodal Knowledge Agent benchmark, which Google co-built with Artificial Analysis.
Luke
Leaderboard claims come from the company's release. Nothing here shows how large the margins are.
Le Pouls
- Gemini 4 Carbon has moved from a restricted two-week trial to paying API developers, ending a closed period that only a few cybersecurity firms had experienced.
- Introductory pricing of $2 per million input and $10 per million output tokens creates a rush to lock in work before the rate doubles after Nov. 30.
- Wiz, a Google-owned security firm, reported finding previously unknown zero-day flaws in two large vendors' software in under two hours, a task that usually takes weeks.
- The specialized Carbon for cyber edition remains behind a Fairwind application, leaving the most sensitive capability in controlled hands while the general model spreads.
- Google is steering developers toward Carbon for demanding work while keeping cheaper Argon for volume, and the coming price step will test how much that premium is worth.
On a Tuesday in November, Google opened its most capable model, Gemini 4 Carbon, beyond the small circle of cybersecurity firms that had tested it for two weeks, placing it before paying developers and top-tier subscribers. It is a familiar human pattern: a powerful tool first guarded, then shared, with the terms of sharing written into the price itself. The introductory rate expires at the end of November, and the cost of thinking harder is quietly being set in public view. What remains withheld, the cyber edition that requires an application, reminds us that the most dangerous capabilities are still handed out with a gatekeeper's hand on the door.
On Tuesday, Google moved Gemini 4 Carbon, its most capable model, out of a two-week restricted trial and into the hands of paying developers worldwide. Until now, only a handful of cybersecurity firms had access. Developers can now call it through the API in Google AI Studio and Gemini for business.
Consumers can reach it too, though only at the top tier. Subscribers to Google AI Ultra, at $100 a month, can use Carbon in the Gemini app, AI Mode in Search, and Antigravity, Google's agentic development platform. Pro and free users must wait.
The pricing comes with a clock. Introductory API rates match the earlier Argon model at $2 per million input tokens and $10 per million output tokens. They expire Nov. 30, after which Google says the price doubles, so developers have a narrow window to lock in their work.
Google positions Carbon for hard problems such as complex programming, deep debugging and scientific research. The model uses more tokens because it reasons and checks its own output repeatedly. For everyday tasks, Google still points to Argon.
The launch included a security demonstration. Wiz, which Google owns, ran Carbon against two major software vendors, uncovered unknown zero-day vulnerabilities in under two hours, and helped patch them. The specialized cyber edition used for that work still requires a Fairwind application; the released version is the general model.
Google also highlighted benchmark wins, research access for 10,000 scientists, and multi-agent demonstrations. The most sensitive capability stays gated even as the general model goes wide. Attention now turns to whether the cyber edition opens further, and how developers respond when the price changes.
On Tuesday, Google moved its most capable model, Gemini 4 Carbon, out of a two-week restricted trial and put it in front of paying developers around the world. Until then, access had been limited to a handful of cybersecurity firms. According to the company, the model is now callable through the API in Google AI Studio and in Gemini for business.
Consumers can get to it too, but only at the top of the ladder. Subscribers to Google AI Ultra, the $100-a-month tier, can use Carbon inside the Gemini app, in AI Mode in Search, and in Antigravity, Google's agentic software development platform. Pro-tier and free users are left out for now.
The pricing carries a deadline. Google is holding introductory API rates at the level of the earlier Gemini 4 Argon model: $2 per million input tokens and $10 per million output tokens. That rate expires Nov. 30, and the company said the price doubles afterward. Developers have a narrow window to lock in work at the lower rate.
Google is pitching Carbon for hard problems rather than everyday chores. It says the model suits complex programming, deep debugging and scientific research, and that it uses more tokens because it reasons and checks its own output repeatedly. For routine work, Google still recommends Argon, the first and less advanced Gemini 4 model. The company describes a two-tier line: Argon for volume, Carbon for high-precision work in legal and financial analysis, full-stack development and 3D design.
The launch came with a security demonstration. Wiz, the cybersecurity company Google owns, ran Carbon against two large software vendors and surfaced previously unknown zero-day vulnerabilities in under two hours, then helped the affected companies patch them. Google said work of that kind has typically taken weeks. There is a caveat: Wiz used Carbon for cyber, a specialized edition that still requires an application through the Fairwind program. The version released this week is the general-purpose model.
Benchmark results released alongside the launch put Carbon at the top of several third-party leaderboards, including CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Google also claims a cost per task more than 2X lower than rival external models. On security-specific tests such as SEC-bench Pro, ExploitGym and CWE-bench v1, the company said the model can identify, exploit and remediate vulnerabilities.
Google also pointed to gains in protein structure prediction, a step in drug development. It is granting 10,000 scientists six months of free access to support their research, and says thousands have already been onboarded.
On multimodal work, Google and Artificial Analysis built the Multimodal Knowledge Agent benchmark, where Carbon placed second while keeping average cost per task roughly 6X below the top model. Using Antigravity's /swarm feature, the company showed the model coordinating multiple agents to rebuild a 3D model of the Eiffel Tower's construction from historical photographs and engineering plates, and to assemble an interactive textile visualizer from academic research.
The most sensitive capability stays gated behind an application even as the general model goes wide. Watch whether the Fairwind-restricted cyber edition expands, and how developers react when the Nov. 30 price step arrives.
Citations marquantes
Carbon suits complex programming, deep debugging and scientific research, and uses more tokens because it reasons and verifies its output repeatedly; Argon remains recommended for routine work.— Google
Envie de l'histoire complète ?
Lire l'original sur BigGo Finance ↗