Google opens Gemini 4 Carbon at $2 per million input tokens

By Nathan Brooks

Google opened Gemini 4 Carbon to all paying developers and enterprises on Tuesday at introductory prices of $2 per million input tokens and $10 per million output tokens. The model, which Google calls its strongest, is available through the Gemini API in Google AI Studio and Gemini for business.

Google is also rolling out Carbon to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and the Antigravity coding platform. It has not said whether lower subscription tiers or free users will receive the model.

The introductory rates match Gemini 4 Argon and end Nov. 30. After that date, the public model will cost $4 per million input tokens and $20 per million output tokens. Anthropic charges $8 per million input tokens and $40 per million output tokens for Claude Fable 5.2.

The Verge put Carbon's introductory rate at roughly one-fifth of OpenAI's GPT-6 Astra and one-quarter of Anthropic's Claude Fable 5.2, which launched a day earlier. Cheaper tokens do not necessarily mean cheaper completed work.

Google still calls Argon "the ideal choice for the majority of software engineering and knowledge work." Carbon uses more thinking tokens than Argon, despite sharing Argon's introductory API price.

Two weeks before the general opening, Google had released Carbon only to a group of cybersecurity defenders through the Fairwind Program, MLQ.ai reported on Oct. 29. Defenders can still apply to the program on a rolling basis.

What does /swarm change?

Alongside the rollout, Google previewed /swarm in Antigravity, turning one prompt into a coordinated team of agents that users can track in real time. Carbon handles architecture, coordination and auditing, while faster, lighter models work in parallel in isolated workspaces and adversarial verifier agents test every change.

Operationally, the preview resembles a semiconductor test cell: one controller, parallel workers and a separate inspection function. Google demonstrated a 3D Eiffel Tower exhibition, a reconstruction of a Google office and a 3D pelican game, with one project running through 10 overnight loops. The company disclosed neither the maximum agent count nor the plans eligible for the preview.

What does Carbon contain?

Carbon uses the same pre-trained sparse mixture-of-experts base as Argon. Its model card attributes the additional capability to extra post-training and reinforcement learning for deep reasoning and long verification loops.

The model accepts text, images, audio and video with a context window of up to 1 million tokens. It outputs text and has a March 2026 knowledge cutoff. Carbon spends more thinking tokens than Argon, yet Google still positions Argon as the general-purpose option for software engineering and knowledge work.

How does Carbon compare?

Google's post places Carbon first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at less than half the cost per task of the next-best outside model. It ranks second on a multimodal knowledge-agent test built with Artificial Analysis, at less than one-sixth of the leader's average cost per task. Google did not identify the rivals behind those claims.

Google's release is broad; its comparative evidence is not. Artificial Analysis had ranked Carbon (high) first on its Intelligence Index v5 at 59 on Oct. 28. Its current leaderboard puts Claude Fable 5.2 (max) at 65, Carbon (high) at 60, Claude Opus 5.5 (max) at 59, GPT-6 Astra (max) at 57 and OpenAI's lower-priced GPT-6.1 Terra (max) at 51.

Per-token prices understate the differences in what a job costs. Artificial Analysis measured $7.76 per index task for Carbon, $26.54 for Fable 5.2 (max), $6.24 for GPT-6 Astra (max) and $0.96 for GPT-6.1 Terra (max). The score and the spending requirement need not agree.

Anthropic's self-run launch table showed Fable 5.2 ahead on coding, science and computer-use tests, while Carbon led AutomationBench. Its results were narrower than the launch rankings:

Anthropic-run benchmark	Claude Fable 5.2	Gemini 4 Carbon	GPT-6 Astra
FrontierSWE v2	69.9%	61.2%	65.5%
FrontierCode 2.0	49.1%	41.1%*	42.8%
Terminal-Bench-Science 0.1	74.6%	64.8%	68.1%
OSWorld 2.1	89.3%	81.5%	79.5%
AutomationBench	51.4%	53.9%	41.4%

Carbon ran at medium effort on FrontierCode 2.0; the other scores used each model's highest effort unless noted. Neither Google nor Anthropic published head-to-head scores against GPT-6.1 Terra.

What does the cyber gate protect?

The strongest cyber build remains gated, and the cybersecurity scores in Carbon's model card reflect Fairwind access. In Google's tests, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent.

Its per-attempt breach rate fell to 0.8% from 3.1% for Argon. Gray Swan's automated cyberattack-query success rate fell to 2.17% from 9.86%.

Outside tester Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests. Google nevertheless says prompt-injection robustness "remains an area for continued improvement."

Google is also taking applications from scientists for six months of free Argon and Carbon access on Antigravity's top usage tier, subject to capacity. Engadget reported that thousands had already been onboarded.

After Nov. 30, Carbon's listed API price doubles, while /swarm retains undisclosed agent and plan limits and the six-month scientist program remains capacity-bound. Its benchmark standing is strong, but measured task economics remain mixed, and current access still excludes free users and the strongest cyber build.