Google opens Gemini 4 Carbon to everyone with agent swarm preview

Who can access Gemini 4 Carbon now?

Google DeepMind vice president Koray Kavukcuoglu announced that Gemini 4 Carbon is now available globally to developers through the Gemini API and AI Studio, to Google AI Ultra subscribers in Antigravity, and to AI Ultra and Pro subscribers through the Gemini app.

The company also enabled Carbon in AI Mode in Search for Google AI Ultra and Pro subscribers.

This is the second phase of a staged rollout. When Google DeepMind first introduced Carbon on 28 October, access was restricted to cybersecurity defenders through the Fairwind program.

Kavukcuoglu wrote that the early preview helped the team refine the model based on real-world security use cases before the wider release.

What does Carbon cost compared to Argon?

Carbon launches at the same introductory pricing tier as [Gemini 4 Argon](/blog/google-gemini-4-argon-general-availability-api-gemini-app-pricing/): $2 per million input tokens and $10 per million output tokens. After 30 November 2026 the price rises to $4 and $20 respectively.

Cached input tokens are priced at 95% off the input token rate, and Google says the model supports a context window of up to 2 million tokens.

The knowledge cutoff is March 2026. Google says Carbon can use more tokens to maximise performance at higher thinking levels, and notes that occasional slowness or timeouts may occur.

Carbon is available to Google AI Ultra subscribers through Antigravity with additional features including extended thinking capabilities and the /swarm preview.

How does Carbon perform on benchmarks?

Google reports that Carbon ranks first on CursorBench, an evaluation of code-editing ability; on WANDR, a benchmark for web-agent navigation; and on AutomationBench, which tests autonomous task completion.

The model also scores first on the Artificial Analysis Intelligence Index v5, a multi-category ranking.

These are vendor-reported results, and independent verification of Carbon on standard academic benchmarks such as MMLU, GSM8K and HumanEval has not yet been published. Google DeepMind says a detailed evaluation and model card are available on its model card page.

In earlier testing for the cybersecurity preview, Google said Carbon achieved state-of-the-art scores on vulnerability discovery and secure coding tasks, based on its own benchmark sets.

What is the /swarm feature in Antigravity?

Google is previewing /swarm, described by the company as a way for one model to coordinate with itself as multiple agents.

The feature, available in Antigravity for Google AI Ultra subscribers, lets developers define tasks that Carbon decomposes across parallel agent instances. Kavukcuoglu called it an early look at how Google sees AI agents working together.

Antigravity already has [cheminformatics skills](/blog/deepmind-antigravity-cheminformatics-skills/) and other domain-specific abilities that can be composed with the new coordination capability.

The /swarm concept echoes ideas from earlier research on [multi-agent systems](/blog/four-gemini-4-argon-agents-beat-minecraft/), where four Gemini 4 Argon agents completed a Minecraft survival challenge from a random starting position in 27 hours.

Google says /swarm is in preview and the API and behaviour may change before general availability.

What safety measures does Carbon include?

Google says Carbon was built with its Safety by Design framework and includes new child safety commitments. The company published a model card alongside the release that details capability evaluations across multiple risk domains.

The model card notes that Carbon, like Argon, may reach capability thresholds in cyber and biological domains that Google classifies as critical.

Kavukcuoglu wrote that Google cannot rule out emerging risks and that the full Safety and Security Report is still forthcoming. The risk framework mirrors the approach Google [described in Argon safety report](/blog/gemini-4-argon-safety-report-cyber-bio-risk-chain-of-thought-monitoring/) from October, which flagged similar concerns.

Google also says Carbon implements input-output guardrails, chain-of-thought monitoring for hazardous reasoning, and automated red-teaming pipelines. The company updates the model card as additional evaluations are completed.

What does the broader availability mean?

The GA release puts Carbon into direct competition with Anthropics upcoming IPO and its Claude models, OpenAIs GPT series, and Metas Llama family.

Carbons pricing matches Argons tier, which was set to compete with GPT-5.5 and Claude 4.5 Opus on cost per token.

Googles strategy of restricted preview followed by broad release mirrors the Argon launch cycle from September and October. The shorter gap between preview and GA may reflect confidence in the models safety evaluation results.

Early adopters in the Fairwind program reported that Carbon handled red-team cybersecurity exercises with fewer bypasses than Argon, according to Googles own summaries.

The /swarm preview signals that Google sees multi-agent orchestration as a key differentiator for its cloud platform. If the coordinated-agent approach gains traction with enterprise developers, it could shift the competitive dynamic from individual model capability toward agent-infrastructure lock-in.