Gemini 4 Carbon opens to all paying developers: first on three agentic benchmarks at under half the per-task cost of the next external model
Synopsis
Google released Gemini 4 Carbon for developers, enterprises, scientists and consumers, pitched at long-horizon agentic workflows and cybersecurity. It ranks first on three independent benchmarks, CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. API pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.
Interpretation
Google opened Gemini 4 Carbon to paying developers, enterprises, scientists and consumers, positioning it for complex tasks that spend more tokens exploring, reasoning and running extended verification loops, while Gemini 4 Argon remains the choice for most software engineering and knowledge work. Two weeks earlier Carbon was available only to a set of cybersecurity defenders through the Fairwind Program, and the prior public frontier release in this family was Gemini 4 Argon. Carbon ranks first on three independent benchmarks, CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model.
In Google Antigravity, a preview called /swarm turns a single prompt into a coordinated team of agents tracked in real time, with Carbon assigned roles that demand the deepest reasoning such as architectural design, coordination and auditing. The arrangement pairs Carbon with faster, lighter models across parallel isolated workspaces, and adversarial verifiers test every change. The preview runs in Google Antigravity, while Carbon is also available through the API in Google AI Studio and Gemini for business and rolls out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.
Perspective
The report covers the product availability, benchmark rankings, pricing and safety testing results Google published; free research access requires an application and is subject to eligibility, capacity and selection, with onboarding on a weekly rolling basis.
Protein homolog retrieval improved 40-fold on average across 7 low-homolog protein families from the CASP17 blind test set, but broader benchmarking is still underway; prompt injection robustness remains an area for continued improvement, and advanced automated red teaming shows models could be further improved against sophisticated attacks in some distributions.