Google makes Gemini 4 Carbon available to developers and enterprises

Google says it has made Gemini 4 Carbon, which it describes as its most capable model, available to developers, enterprises, scientists and consumers. Carbon is accessible via the API in Google AI Studio and in Gemini for business, and is rolling out to Google AI Ultra subscribers in Google Antigravity (Google's AI development workspace), the Gemini app and AI Mode in Search.

For developers and enterprises in Africa, the roll-out raises immediate practical questions about performance, cost and controls: where the model runs, how fast responses will be, how much cross-border data transfer will add to bills, and what safeguards apply to sensitive uses such as vulnerability research.

Google set an introductory price that matches Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at 95% off the input token price.

Google's blog says Carbon was first staged through its Fairwind Program for a limited set of cybersecurity defenders. The blog reports that Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task compared with the next best external model, and ranks #2 on the Multimodal Knowledge Agent (MKA) benchmark while being more than 6X lower in average cost per task compared with the top performing model. These named benchmarks are independent tests that measure different aspects of model performance in practical workflows.

Google previewed a coordinated multi-agent workflow feature, /swarm, in Antigravity to run teams of agents on long-horizon tasks. Agentic workflows are teams of AI agents coordinating on extended tasks; Google says Carbon is intended for roles that demand deep reasoning, verification and auditing, while being paired with faster, lighter models for parallel work.

Google highlighted Carbon's cybersecurity performance during the staged release. In Project Fortify, Google's internal autonomous scanning, triage and remediation system, the company reported 100% recall on scanning and 95% precision on triage in its internal benchmark when using Carbon. Google Cloud Vulnerability Research used Carbon to build proofs of concept that showed two bugs were exploitable and helped reclassify them as critical; Google says details will be shared after the responsible disclosure process completes. As part of Wiz's Scan for Good program, teams using Carbon uncovered zero-day vulnerabilities in under two hours. Separately, the drug-toxicity firm Axiom Bio said it is moving production scientific analyses and agents to Carbon and Argon for translational research.

Google Cloud advertises 43 global regions and 130 zones, and offers specialised AI zones with high-capacity GPUs and TPUs. That global footprint matters for any organisation considering Carbon for latency-sensitive, regulated or large-scale agentic workloads.

For organisations and developers in Africa, the announcement raises immediate practical questions. Questions to watch for in Google's documentation include:

• Will Gemini 4 Carbon and Antigravity run from Google Cloud regions in Africa, and if not, which nearby regions will host the model?
• What latency and network egress costs should African customers expect when models run outside the continent?
• Will pricing or committed-use options be offered to African enterprises to reflect cross-border costs or usage patterns?
• What operational safeguards, access controls and responsible-disclosure processes will apply to African customers using Carbon for vulnerability research and agentic workflows?

The announcement makes Carbon broadly available, but regional hosting, pricing details for African customers and the guardrails for dual-use cybersecurity tasks remain unclear.

Newzroom dot Africa has reached out to Google / Google DeepMind and Google Cloud for comment and is awaiting a response. This story will be updated when they respond.

Source: Google DeepMind