Google Expands Gemini 4 Carbon Access Beyond Cybersecurity Defenders

Google announced on November 10, 2026 that Gemini 4 Carbon is available to paying developers and enterprises and is rolling out to Google AI Ultra subscribers. The frontier model’s introductory API price is $2 per million input tokens and $10 per million output tokens, including tokens used for reasoning.

Paying developers can use Carbon through the Gemini API in AI Studio and Android Studio, while enterprises can access it through Gemini for business. For Ultra subscribers, it is rolling out in Google Antigravity, the Gemini app and AI Mode in Search.

For scientists, Google is expanding its research program to offer selected current and future participants both Argon and Carbon in Antigravity’s highest usage tier free for six months. Applications remain open, with rolling onboarding subject to eligibility, capacity and selection.

An October 28, 2026 announcement had introduced Carbon through a deliberately limited Fairwind rollout to a subset of cybersecurity defenders, with wider access promised later. The specialized cyber configuration remains gated through Fairwind applications and rolling admission.

Higher Token Use Behind the Price

Carbon builds on Argon’s pre-training foundation and architecture, with additional post-training for deep reasoning, extended verification and multi-step agent workflows, according to the model card. It accepts text, images, audio and video and produces text, with up to 1 million tokens shared by input, reasoning and output. The listed knowledge cutoff is March 2026; the card also identifies hallucinations, possible slowness or timeouts and greater token consumption at higher thinking levels as limitations.

Developers can select low, medium or high thinking levels, with medium as the default. The developer documentation reports that Carbon generally spends more thinking tokens per turn than Argon, increasing token use and latency, and positions Argon for interactive coding, routine feature work and fast production loops. Moving an existing Argon request requires changing the model ID rather than the request schema, although output limits and client timeouts may need adjustment.

Carbon is excluded from the Gemini API free tier, and its pricing page says every processed token is billed, including thinking tokens and specified internal grounding instructions and declarations. After the introductory period ends on November 30, 2026, input and output prices become $4 and $20 per million tokens, respectively, on December 1, 2026. Matching Argon’s introductory per-token rates therefore does not guarantee matching per-task costs: Carbon’s extra reasoning tokens are billable.

How /Swarm Coordinates Agent Teams

Google Antigravity’s /swarm preview is rolling out in the Antigravity app and command-line interface. A coordinator creates specialized workers in isolated Git worktrees so their edits do not collide, while adversarial verifiers run tests and challenge code changes. Agents communicate through a shared append-only message board, and idle agents can wait without consuming tokens.

A live board, member roster and graph timeline let users inspect tool activity, steer individual agents and replay how a team reached its result. Roles can use different models, allowing Carbon to handle coordination, architecture or auditing while lighter models perform routine implementation. The Antigravity team reports five weeks of internal testing on multi-package migrations, cross-subsystem investigations and architectural redesigns, alongside a warning that parallel multi-agent systems consume quota much faster than standard chat.

Benchmark Results and Cyber Safeguards

The model card’s results, current as of October 2026, report 59 for Carbon on Artificial Analysis Intelligence Index v5.0, compared with 50 for Argon, and 96.2% versus 84.2% on the 256K-to-1M GraphWalks long-context test. Carbon’s 61.2% on FrontierSWE 2 exceeded Argon’s 55.0% but trailed the listed 65.5% for GPT-6 Astra and 62.3% for Claude Opus 5.5. The card’s cybersecurity results reflect access through Fairwind rather than the general public configuration.

Google’s safety report says Carbon did not cross its material-capability-change threshold relative to Argon, so a full new critical-capability assessment was not required. It carries forward the determination that, without safeguards outside the model, Carbon cannot be ruled out from reaching Cyber Uplift Level 1, Google’s threshold for cyber capabilities that could enable additional severe-scale harm from high-impact attacks. Google assesses that strengthened safeguards reduce the residual risk to an acceptable level.

In Gray Swan testing detailed in the report, Carbon’s automated cyberattack query success rate was 2.17%, compared with 9.86% for Argon. The per-objective rate, which records whether a task was breached at least once across retries, remained 40% after 200 attempts at each of 25 high-severity tasks. The contrast shows greater operational friction rather than the elimination of breaches: individual attacks became less reliable, but persistent retries still succeeded on a substantial minority of objectives.

Carbon also recorded a 0% attack-success rate on Gray Swan’s external indirect prompt-injection benchmark, yet the safety report says such attacks remain possible. A benchmark shutout therefore does not establish immunity for agents that process untrusted content.

The broader release makes Carbon a paid option for deeper reasoning and coordinated agent work, while Fairwind preserves a separate gate for specialized cyber capabilities. The central trade-off is whether its quality gains justify additional token use and latency, and whether the documented safeguards remain effective as a much wider user base tests those agent workflows.