Google releases Gemini 4 Carbon, previews /swarm

Google released Gemini 4 Carbon to paying API users at $2 input and $10 output per million tokens until November 30, and previewed /swarm teams.
What it means for founders
- Budget for the December price. Cost models built on the $2 and $10 rate break after November 30. Plan on $4 and $20, and on more tokens per task than Argon, since Carbon is built to reason longer.
- Route by task. Copy Google's pattern in your own agent stack now: a strong model coordinating and checking, cheaper models doing bulk edits.
- Platform risk for coding agent startups. An orchestrator inside Google's coding tool covers ground many agent wrappers sell. Watch whether Carbon in Antigravity, and the `/swarm` teams built on it, reach users below the Ultra tier.
- Test the cost claims. Treat the cost per task figures as a hypothesis until outside evaluators or your own workloads confirm them.
What we don't know yet
Google has not said when `/swarm` leaves preview, what API rate limits apply to Carbon, or whether it will reach cheaper consumer plans.
Related: Gemini 4 in refinement, free Gemini users moved to Flash-Lite and our Gemini page.
The story
Google DeepMind opened Gemini 4 Carbon to paying developers and enterprises on November 10, two weeks after giving it to security defenders through its Fairwind Program. Google calls Carbon its strongest model, places it above Gemini 4 Argon for "complex, multi-step coding and debugging and open-ended scientific research" and says its cybersecurity skills are its best yet. Argon, Google says, stays better for most software engineering and knowledge work; Carbon spends more tokens exploring, reasoning and running extended checks.
Price, access and limits
Paying developers get Carbon through the Gemini API and AI Studio, enterprises through Gemini for business. Until November 30, 2026 it costs the same as Argon, $2 per million tokens of input and $10 per million of output, according to Google's announcement, and a footnote says the rate then doubles to $4 and $20. For consumers, Carbon is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity.
Per the model card, Carbon builds on Argon, takes text, images, audio and video, and has a combined input and output window of up to 1 million tokens and a March 2026 knowledge cutoff. Google warns of occasional timeouts and heavier token use at higher thinking levels.
The /swarm preview
Google is also previewing `/swarm` in Antigravity, its agentic coding app. Per the Antigravity team, a prompt goes to a coordinator agent that launches parallel specialized workers, each in its own git worktree so edits cannot collide. Adversarial verifiers run tests and challenge each diff. All agents post to one shared board that allows additions but no edits, and an idle agent costs zero tokens until another needs its input.
Agents can run different models: Google suggests Carbon for coordinator, architecture, security audit and verification roles, lighter models for routine edits. A side panel shows the live board, a roster of agents you can step into and steer, and a replayable timeline. The preview is rolling out in the Antigravity app and CLI; Google warns swarms use quota much faster than normal chat.
Benchmark results, all from Google
Google says Carbon tops three workflow tests, Zapier's AutomationBench, Perplexity's WANDR and CursorBench 4.0, at under half the cost per task of the next best outside model. On the Multimodal Knowledge Agent benchmark built with Artificial Analysis, Google places it second, at over six times lower average cost per task than the leader. Google calls these benchmarks independent, but no benchmark maker has confirmed the results.
Sources
Primary sources
Reporting
Tools in this story
We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.
Google's assistant, built into Workspace and Android
More in Models & Labs
- Nvidia's open source push faces test after Huang exit

For founders: Hosting terms. The pledge that Hugging Face stays chip neutral is Nvidia's, made in a post under Huang's name.
TechCrunch · 5d ago - Google releases Gemini 4 Hydrogen, its new fast modelGemini 4 Hydrogen is Google's fast, cheaper model: $1.50 per million input tokens, a 1M context, and now inside Search.Google DeepMind · 6d ago

- Anthropic releases Fable 5.2 and Mythos 5.2

For founders: Price the switch by task cost. At twice the Opus 5.5 rate, Fable needs to finish jobs with fewer tokens or retries to save money.
The Verge · 8d ago - Meta launches Muse Ember for high-volume agent work

For founders: Price your agent by the task. Ember's advantage depends on short outputs and cheap cache hits, so a pipeline that reuses a long system prompt gains the most.
TechCrunch · 9d ago - OpenAI's GPT-6.1 Terra halves per-token API rates

For founders: Rerun your agent costs. At $1 in and $5 out per million tokens, high-volume chores like invoice entry and form filling get cheaper.
OpenAI · 23d ago