Is Google's new /swarm mode worth the quota?
In the one like-for-like test I found, Gemini 4 /swarm finished 36% sooner and used twice the quota. For agentic work I would take slow, steady and cheap.
I would not run Google's new /swarm mode by default. For agentic development I want stability more than raw speed, and swarm trades quota for speed at a rate I don't think is worth paying. What would change my mind is a large job that splits cleanly into separate pieces and has a real deadline.
On 10 November 2026 Google previewed /swarm in Antigravity, its agent-based development environment, alongside Gemini 4 Carbon1. A coordinator splits a task across named worker agents, each with its own conversation, and a verifier reads the diffs and runs the tests2. Google's documentation is blunt about the price: every worker is its own agent, so a swarm "uses quota much faster than a single conversation"2. Quota is the allowance a subscription gives you, separate from any per-token API bill. Antigravity refreshes it every five hours on Pro and Ultra plans, with a weekly limit on top3.
The only like-for-like number I found is thin. On 13 November one developer, @spnova05, posted results from running the same prompt on Gemini 4 Hydrogen twice. The swarm took 9 minutes 14 seconds, used 10% of quota and wrote 3,735 lines. A single agent took 14 minutes 26 seconds, used 5% and wrote 2,084 lines4. That is 36% less time for twice the quota and 79% more code. It is one run from one person, the post does not say which quota window the percentages measure, and line count says nothing about quality. I treat it as a rough direction and no more.
Swarm result as a percentage of the solo run on Gemini 4 Hydrogen (one developer's single run, @spnova05 on X, 13 Nov 2026; percentages calculated from the posted figures)
Others report the same shape without a control run. On 13 November @imrahul165 wrote that 25% of the weekly quota went in one day of using /swarm with Hydrogen5. On 11 November @malsabbagh_io ran seven or more tasks at once on Carbon and reported that after 6 or 7 hours of continuous use 65% of the weekly Ultra quota was gone, while still calling the swarm interface impressive6. The author of paddo.dev, who benchmarks models on real code, ran four parallel Carbon sessions on a $99.99 Ultra plan on 11 November. Every session stopped about two hours in, with a message that quota would reset in 2 hours 5 minutes7. Those were parallel sessions rather than /swarm itself, but the write-up reports that the swarm menu labels the mode "Preview & Quota Heavy"7.
That is where stability comes in. Swarm's own documentation says a worker that runs out of quota pauses, and does not restart on its own when your quota refreshes2. The point of an agent is that it keeps going while I am not watching, and a design that stalls halfway through is the wrong shape for that. The paddo.dev write-up guesses that one session at a time would probably last a window, and says that was not tested7. Artificial Analysis also rates Hydrogen as very verbose. Its score of 54 at the top setting took 304 million output tokens against a median of 77.3 million8, and Carbon takes 30.9 minutes per task against a median of 13.19. A model that already spends a lot per task is a poor thing to multiply.
Speed matters when a person is waiting. An agent that works through the night and the weekend is not waiting on anyone. If the ratio in that one run held, halving the quota per job gives roughly twice as many jobs from the same allowance. Price can help too, depending on the provider. On 14 November DeepSeek's pricing page showed half price outside its weekday peak windows, so deepseek-flash output falls from $1.20 to $0.60 per million tokens10. I could not find a time-of-day discount at Google. Its API pricing page of the same date lists Batch, which runs asynchronously with a 24-hour target, and Flex, a preview tier with variable latency, both at half the standard price, so Hydrogen output costs $3.75 rather than $7.50 per million tokens111213. Antigravity's plans page describes five-hour and weekly quota windows and mentions no off-peak rate3. The mechanisms differ, but both discount work that can wait. Carbon's API price of $2 and $10 per million input and output tokens also doubles to $4 and $20 after 30 November1.
Google's answer is that swarm does not have to be expensive. It says Carbon can coordinate while Hydrogen agents do the work and the checking, without driving up token costs14. It also says Orchestrator, a separate feature that switches between Carbon and Hydrogen inside one session, used 24% less quota and finished 19% faster on Terminal-Bench 4.0 while reaching close to Carbon's success rate14. Those are Google's own figures and I have not seen them independently confirmed. They describe Orchestrator rather than swarm, and the gain came from switching models inside one conversation. The swarm test above already used the cheaper Hydrogen model in both runs and still burned twice the quota.
I am not against swarm. Google's guidance is to use it when a task divides cleanly across separate files, packages or questions2, and the demos show what that looks like. @shengzheyao, whose profile reads "Antigravity@Google DeepMind | Creator of Antigravity CLI", posted a run of 23 agents, with a Carbon coordinator and verifiers over 16 Hydrogen builders, that produced 16 small apps from one prompt15. The post does not give the quota it used. If a migration is large, splits cleanly and has to land tomorrow, paying double for the speed is reasonable.
Until someone is waiting on the result, I will take the agent that is still running on Friday over the swarm that finished by lunch.
Nothing here is financial, legal or investment advice. Every figure is published by its named source and carries the date it was published; figures may since have changed.
References
-
Google, Gemini 4 Carbon is now available ↑
-
Google Antigravity Docs, Swarm (/swarm) ↑
-
@spnova05 on X, Hydrogen in Three.js, Swarm vs Solo (13 Nov 2026) ↑
-
@imrahul165 on X, Testing Gemini 4 Hydrogen in Antigravity via /swarm (13 Nov 2026) ↑
-
@malsabbagh_io on X, Antigravity Swarm UI is impressive (11 Nov 2026) ↑
-
paddo.dev, Two Hours of Carbon: Gemini 4 Carbon vs Argon on Real Code ↑
-
Artificial Analysis, Gemini 4 Hydrogen: Intelligence, Performance & Price Analysis ↑
-
Artificial Analysis, Gemini 4 Carbon: Intelligence, Performance & Price Analysis ↑
-
DeepSeek API Docs, Models & Pricing ↑
-
Google AI for Developers, Gemini Developer API pricing ↑
-
Google AI for Developers, Flex inference ↑
-
Google, Introducing Gemini 4 Hydrogen ↑
-
@shengzheyao on X, Gemini 4 Hydrogen just landed in Antigravity CLI (12 Nov 2026) ↑