Google opens Gemini 4 Carbon to all developers; introductory price doubles December 1

Google DeepMind made Gemini 4 Carbon generally available with a /swarm preview in Antigravity, at an introductory price set to double on December 1.

What happened

Google DeepMind made Gemini 4 Carbon available to developers, enterprises, scientists and consumers, two weeks after releasing it to a set of cybersecurity defenders through its Fairwind Program. Google describes Carbon as its strongest model, with frontier performance in cybersecurity capabilities and long-horizon agentic workflows. It positions Gemini 4 Argon, which The Verge notes launched only last month, as the ideal choice for most software engineering and knowledge work, while Carbon targets tasks that test the limits of a model's intelligence, using more tokens than Argon to explore thoroughly and run extended verification loops.

Access runs through the Gemini API in Google AI Studio and Gemini for business for all paying developers and enterprises, at the same introductory price as Argon. Simon Willison writes that the launch price matches GPT-6.1 Sol and Claude Sonnet 5.5 pricing and is scheduled to double on December 1, to the same level as Opus 5.5. Carbon is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. The Verge reports it is not clear when or if Carbon will reach lower tier subscribers or free users.

Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model, and #2 on a new Multimodal Knowledge Agent benchmark built with Artificial Analysis. Google is also previewing /swarm in Antigravity, turning a single prompt into a coordinated team of agents that developers can track in real time, assigning Carbon to roles such as architectural design, coordination and auditing alongside lighter models.

Why it matters

For teams costing out agentic workloads, two variables moved at once. Google says Carbon deliberately spends more tokens than Argon to reason deeply and run verification loops, so per-token comparisons do not settle the question. The Verge makes the same point: even after the introductory period ends, Carbon stays cheaper than frontier rivals per token, though that does not necessarily translate to lower costs because models can use different numbers of tokens for the same tasks. Google's own framing is cost per task, claiming more than 2X lower cost per task than the next best external model on CursorBench 4.0, WANDR and AutomationBench, and more than 6X lower average cost per task than the top model on the MKA benchmark.

The routing guidance is explicit: Argon for most software engineering and knowledge work, Carbon for multi-package migrations, cross-subsystem root-cause investigations and complex architectural redesigns. The /swarm preview in Antigravity formalises a mixed-model pattern, with Carbon on architecture, coordination and auditing, lighter models in parallel isolated workspaces, and adversarial verifiers testing every change.

Security numbers are relevant to anyone shipping agents. Google reports Carbon achieved a 0% prompt injection attack success rate on the Gray Swan IPI benchmark, while stating that prompt injection robustness remains an area for continued improvement. On a 375-task agentic honesty benchmark Google says Carbon passed 92% of tasks versus 70% for Argon, declining to fill in missing experimental data 98% of the time. The Verge notes companies can and do cherry-pick benchmark results.

What we don't know yet

Google's blog does not say what happens to throughput, rate limits or capacity once the introductory period ends, nor whether the Fairwind Program remains the only route to the cyber-specific benchmark leadership on SEC-bench Pro, ExploitGym and CWE-bench v1. The Verge says it is unclear when or if Carbon reaches lower tier subscribers or free users. How much of the cost-per-task advantage holds up outside the three benchmarks Google cites, given The Verge's note that vendors can cherry-pick results? Google says broader benchmarking of the protein structure homolog work is still underway, and that details of the Cloud Vulnerability Research findings will be shared only after responsible disclosure completes. None of the internal figures, including the 100% scanning recall and 95% triage precision in Project Fortify, are described in the sources as independently verified. How does /swarm behave outside Antigravity, and when does the preview become generally available?

Sources: Google DeepMind / Simon Willison / The Verge AI

How this Deep Take is made: it is written with AI assistance from the linked sources, then checked against them sentence by sentence, and every fact links to where it came from.