Gemini 4 Carbon Opens to Developers: Tops CursorBench at Half Its December API Price
Carbon leads three agentic benchmarks at half its December price; /swarm debuts in Antigravity
Google made Gemini 4 Carbon broadly available on November 10, 2026, opening the model to paid API developers, enterprise customers through Gemini for Business, and AI Ultra subscribers. "Today, we're making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers," wrote Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect. The release arrives at an introductory price of $2 per million input tokens and $10 per million output tokens — exactly half the standard rate of $4 and $20, respectively, that takes effect December 1.
Three Agentic Benchmark Rankings
According to Google's availability announcement, Carbon ranks first on three real-world agentic evaluations: CursorBench 4.0, which measures coding-agent performance; Zapier's AutomationBench, which tests office automation tasks across sales, marketing, operations, support, and finance workflows; and Perplexity's WANDR research evaluation. All three rankings are drawn from Google's own launch data.
An independent evaluation by Artificial Analysis placed Carbon first among 32 models tested on AutomationBench-AA — the firm's own variant of the Zapier benchmark — with a score of 77.8 percent. At Carbon's October 28 announcement, Google also published a 53.9 percent score on Zapier's public AutomationBench leaderboard, the highest score for any single model at that time. Google says Carbon achieves more than twice the cost efficiency per solved task compared with the next-best external model when measured at standard pricing.
Google's launch data shows Carbon scoring 60.4 percent on CursorBench 4.0, ahead of competing frontier models in that evaluation. At Proximal Labs' FrontierSWE v2 benchmark — which measures 34 ultra-long-horizon engineering tasks each with a 20-hour computational budget — Google's launch data places Carbon at 61.2 percent. Anthropic's Claude Fable 5.2, released November 9 the day before Carbon's general availability, scored 69.9 percent on FrontierSWE v2 according to Anthropic's launch data, placing it ahead of Carbon on that benchmark.
Cybersecurity Capabilities and Phased Access
In external evaluations reported by SecurityWeek and Help Net Security, Carbon scored 88.2 percent on SEC-bench Pro for finding vulnerabilities in real-world software, 54.8 percent on ExploitGym for determining whether vulnerabilities can be exploited, and 71.0 percent on Collinear's CWE-bench for writing security patches. Carbon also more than doubled Argon's success rate at generating secure code in Google's internal testing, and Google's Chrome team found Carbon outperformed a large rival frontier model by more than ten times on high-quality patch generation.
In safety testing conducted by Gray Swan, Carbon blocked 99 percent of attack trajectories in simulated cyberattack scenarios, with a per-attempt breach rate of 0.8 percent, down from 3.1 percent for Gemini 4 Argon. Carbon's automated cyberattack query success rate reached 2.17 percent compared with 9.86 percent for Argon, and the model achieved a 0 percent attack success rate on Gray Swan's Indirect Prompt Injection benchmark.
Google is staging broader cybersecurity access through its Fairwind Program, introduced alongside the October 28 preview, which provides phased access to trusted cyber defenders before widening to additional partners on a rolling basis. Separately, Wiz's Scan for Good responsible-disclosure program — which uses Gemini 4 Argon to help security researchers find zero-day vulnerabilities in production software — has demonstrated how the broader Gemini 4 model family is being applied in live security research.
Reasoning Quality and Hallucination Rates
On the Artificial Analysis Intelligence Index v5, Carbon leads with a score of 59, placing it first among frontier models in that composite evaluation. The score has been confirmed independently by Tom's Hardware, AF.NET, and benchmarklist.com.
Artificial Analysis's AA-Omniscience v1.1 evaluation, which measures factual recall and hallucination rates across economically relevant domains, records Carbon at an 11 percent hallucination rate at the high thinking level. Gemini 4 Argon records a 4 percent hallucination rate on the same evaluation — the lowest among publicly available models tested at the time of Argon's October launch.
/swarm Brings Multi-Agent Teams to Antigravity
Alongside the Carbon release, Google previewed /swarm in Antigravity, its agent-focused development platform. The command turns a single prompt into a coordinated team: a coordinator agent spawns named worker agents in isolated Git worktrees, while adversarial verifier agents run tests and challenge proposed code changes on a shared message board. Workers and verifiers run in parallel, with the coordinator deciding how to integrate outputs.
The /swarm preview rolled out November 10 across both the Antigravity application and its command-line interface. Google says the feature is designed to draw on Carbon's long-horizon coding strengths, where complex engineering tasks benefit from multiple agents independently testing and challenging each other's work before changes are committed.
Pricing and Developer Access
Carbon's introductory API pricing through November 30, 2026, is $2 per million input tokens and $10 per million output tokens. Beginning December 1, pricing moves to $4 per million input tokens and $20 per million output tokens — meaning developers building now have through the end of November to test and optimize applications at the current introductory rate.
Google's API for Carbon supports a thinking_level parameter with low, medium, and high settings that control the amount of internal reasoning the model applies before producing a response, allowing developers to trade response speed against analytical depth on a per-call basis.
Paid API developers can access Carbon through Google AI Studio and the Gemini API immediately. Enterprise access is available through Gemini for Business. AI Ultra subscribers receive access via the Gemini app. Access for sensitive cybersecurity use cases remains gated through the Fairwind Program, which is admitting additional vetted partners on a rolling basis.