Gemini 4 Carbon is now available to everyone — and Google’s /swarm turns one prompt into a team of agents

Google made Gemini 4 Carbon generally available 61 minutes ago, opening its strongest model to all paying developers, enterprises and Google AI Ultra subscribers — and previewing /swarm, a feature in Antigravity that turns a single prompt into a coordinated team of agents. Carbon had been limited to a small circle of cybersecurity defenders since its October 28 debut.

What we checked: Google’s announcement, published November 10, 2026 at 11:00 a.m. ET, cross-checked against VentureBeat’s same-day coverage and Google’s Antigravity blog. Google’s #1 benchmark claims are the company’s own; public tables for two of the three benchmarks did not yet show Carbon at publication time.

• What: Gemini 4 Carbon, Google’s strongest model, is now generally available; Antigravity is previewing /swarm, a multi-agent orchestration feature.
• When: announced November 10, 2026, at 11:00 a.m. ET — two weeks after Carbon’s October 28 debut, which was limited to cybersecurity defenders in the Fairwind Program.
• Price: $2 per million input tokens and $10 per million output tokens through November 30, 2026; $4 and $20 after. A 95% cached-input discount applies.
• Benchmarks: Google claims #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than 2x lower cost per task than the next best model. On Artificial Analysis’s Intelligence Index, Carbon scores 60 at high effort versus 65 for Claude Fable 5.2 at maximum effort.
• Access: paid Gemini API in AI Studio and Android Studio, Gemini for business, and Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.
• Cyber proof point: Wiz used Carbon to find previously unknown vulnerabilities in two major SaaS providers in under two hours; Google says Carbon blocked 99% of attack trajectories in red-team testing.

In under two hours, a security team found vulnerabilities nobody knew existed in the software of two major companies — flaws that would normally take weeks of expert work to surface. The team was Wiz, working through Google’s Scan for Good program, and the tool doing the finding was Gemini 4 Carbon, Google’s strongest model. Until today, almost nobody outside a small circle of cybersecurity defenders was allowed to touch it.

That changed this morning. Google opened Carbon to everyone who pays: developers through the Gemini API, enterprises through Gemini for business, and Ultra subscribers across the Gemini app, AI Mode in Search and Antigravity. It is the widest release of Google’s most capable model, and it lands a day after Anthropic’s Claude Fable 5.2 launch — at a quarter of Fable’s price.

What happened#

Carbon is Google’s frontier model for the hardest jobs: long-horizon agentic workflows, complex multi-step coding, open-ended scientific research and cybersecurity. It was unveiled on October 28 and handed only to selected defenders through DeepMind’s Fairwind Program. Today’s announcement is the general release of that same model — plus a new feature that may matter more than the model itself.

That feature is /swarm, now in preview inside Google Antigravity. One prompt becomes a coordinated team of agents: Carbon takes the roles that demand the deepest reasoning — architectural design, coordination, auditing — while faster, lighter models handle parallel implementation in isolated workspaces, with adversarial verifiers testing every change. Developers can watch the team work in real time.

The price is the headline for buyers. Carbon starts at $2 per million input tokens and $10 per million output tokens — the same introductory rate as Gemini 4 Argon — with a 95% discount on cached input. But the rate is introductory in the literal sense: after November 30 it doubles to $4 and $20. Through the end of the month, Carbon costs a quarter of Claude Fable 5.2’s $8 and $40; after that, half.

Google positions Carbon above Argon, not beside it. Argon, the company says, “remains the ideal choice for the majority of software engineering and knowledge work”; Carbon is built for tasks that test the limits of a model’s intelligence, using more tokens to explore thoroughly, reason deeply and run extended verification loops.

Why it matters#

The benchmark story is where the two launches collide. Google claims Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than twice lower cost per task than the next best external model. On Zapier’s AutomationBench, Carbon scores 53.88% at an estimated $2.12 per task, ahead of Fable 5.2 at 51.45% and $4.64. On Terminal-Bench 4.0, Carbon reaches 65.2%, ahead of Fable’s best listed run at 61.6%.

But the race splits by task. On FrontierCode 2.0, which grades real repository changes against the standard for merging them, Fable reaches 49.1% against Carbon’s 40.6%. On Artificial Analysis’s broad Intelligence Index, Fable 5.2 leads at 65 on maximum effort, with Carbon at 60 on high effort, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. A team picking an agent for terminal-heavy diagnostics could reasonably choose differently from one trying to get patches through review.

There is also a caveat worth stating plainly. VentureBeat, reviewing the launch materials, found that the public CursorBench table shows Fable 5.2 leading at 60.0% with no Carbon row, and Anthropic’s WANDR table leaves Carbon blank — so two of Google’s three #1 claims could not be verified against public data at publication time. The Zapier result, by contrast, is publicly listed. Treat the headline claims as Google’s until the tables catch up.

The cyber numbers are the most concrete in the announcement. In simulated attack testing, Carbon blocked 99% of attack trajectories from Google’s adaptive red-teaming agent, and its safeguards cut the per-attempt breach rate to 0.8%, down from 3.1% on Argon. On a 375-task agentic honesty benchmark it passed 92% of tasks, against 70% for Argon — declining to fill in missing experimental data 98% of the time. External testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and more than 13,000 requests.

And the scientific claims are striking, if early. Researchers used Carbon in Antigravity to build AlphaProof Code, an agentic framework that mathematically proved key security properties of the code behind Signal’s cryptographic protocol in weeks — work Google says historically takes expert teams years. On protein structure prediction, Carbon increased the homologs retrieved across seven low-homolog protein families by 40-fold on average.

What’s next#

Three things to watch. First, the price clock: the $2/$10 rate expires November 30, and teams that adopt Carbon now should measure what the same workloads cost at December rates before the introductory window closes. Second, /swarm: the preview is Google’s answer to the agent-orchestration race, and its real test is whether developers hand Carbon the architecture role while cheaper models do the implementation — the division of labour Google itself recommends. Third, access: Google is expanding its program giving 10,000 scientists free access to Antigravity’s highest usage tier for six months to include Carbon, with applications still open.

Carbon remains a proprietary, hosted model — no downloadable weights, no self-hosting. For teams that need to run inference on their own infrastructure, that is a hard limit no price cut removes. But for everyone else, Google’s strongest model just went from a gated cyber-defender tool to a $2-per-million-token API call. The frontier got cheaper today; whether it got better for your workload is a measurement, not a launch claim.