Gemini 4 Carbon Ships AI Agent Swarms — Does Its Cost Advantage Hold Up?
Google opened Gemini 4 Carbon to all paying developers on 10 November and previewed /swarm, a multi-agent mode in Antigravity, at $2/$10 per million tokens. Independent scoring puts it fourth at 60, with $7.76 and 24.9 minutes per index task.
Google spent two weeks teasing Gemini 4 Carbon to a closed group of security defenders before opening it to everyone on 10 November. The headline of the public launch is not the model. It is a button called /swarm, which turns a single prompt into a team of agents that divide the work, check each other, and keep going overnight.
That framing is worth taking seriously, and worth checking. A launch post is a claim, not a measurement. So this piece keeps Google's numbers, then puts them next to the independent ones, and lets the gap speak.
What actually shipped, and who can touch it
Carbon is now available through the Gemini API in Google AI Studio and through Gemini for Business, for paying developers and enterprises. Google AI Ultra subscribers get it in the Gemini app, in AI Mode inside Search, and on Antigravity. Whether cheaper tiers or free users ever see it is, as of launch, unstated.
The positioning matters more than the availability list. Google says its previous frontier model, Argon, remains the sensible default for most software engineering and knowledge work, and that Carbon exists for the jobs where the model's own intelligence is the bottleneck: long multi-step coding and debugging, open-ended research. Carbon buys that depth by spending more tokens, which is a design decision with a bill attached.
Two weeks before the public launch, Google had handed Carbon to a small set of cybersecurity defenders under what its own announcement calls the Fairwind programme. The public release is that same model, minus the queue.
/swarm: one prompt, a team, and an adversary
The feature Google is pushing hardest is a preview inside Antigravity, its agentic development environment. You give /swarm one prompt. It returns a coordinated set of agents working in parallel on isolated workspaces, with a live trace of who is doing what.
The division of labour is the interesting part. Carbon is assigned the roles that need the deepest reasoning: architecture, coordination, audit. Faster, lighter models take the parallel execution slots. And every proposed change goes past an adversarial validator whose job is to try to break it. That is a small but real design claim: the system is built to distrust its own output rather than to launder it.
It is also where the money moves. In a planner-executor split, the expensive model touches a small fraction of the tokens and the cheap models absorb the volume. Google's own framing implies as much when it says Argon stays the default for ordinary work while Carbon handles the tasks that test a model's ceiling. The architectural bet, in plain terms, is that most of the work in a long task is not thinking, and should not be billed at thinking prices.
What the validator does to the trust model is less obvious and more important. If an agent team checks its own members, the checker becomes the safety boundary, and a weak validator is worse than no validator because it produces confidence without coverage. Google has not published a validator failure rate, which is precisely the number an outside team would need before letting a swarm touch a production branch.
Google's demos run long and cinematic. An office building reconstruction from seven planning documents and 206 drawings, run overnight, producing an interactive 3D model reported at roughly 906,000 triangles. A working model of the 1780s piano mechanism dropped into a concert hall model with live reverberation. An Apollo lunar module rebuilt from period photographs and blueprints. A materials-science paper turned into a fabric simulation tracking stress across 110 fibres.
Treat those as demonstrations, not evidence. A demo proves a capability exists once, on a task the team chose, with people watching. It says nothing about the median job on a Tuesday.
The numbers Google is selling
Three production benchmarks are the core of the pitch: CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench. Google says Carbon takes first place on all three, ahead of Claude Fable 5.2, and that average cost per task comes in more than twice as low as the next-best model.
Pricing is the sharper weapon. Carbon launches at an introductory $2 per million input tokens and $10 per million output tokens, which the company frames as roughly a fifth of GPT-6 Astra and a quarter of Fable 5.2. That introductory window closes at the end of November, when the rates double to $4 and $20.
Two caveats belong in the same paragraph as the price. First, per-token price is not per-task cost: a cheaper model that thinks longer can lose. Second, a model priced at a fifth of a rival is only a bargain if it finishes the job.
Google also cites a 96.2% score on GraphWalks, first place overall in a financial research evaluation run by hedge fund Balyasny across 23 models, and a tie for first with Fable 5.2 on an independent anti-cheating benchmark built by Nim creator Dominik Picheta. Those figures come from Google's own materials as relayed by launch coverage, and they have not been independently reproduced here.
The independent picture is less flattering
Artificial Analysis, which runs its own harness rather than trusting vendor slides, scores Gemini 4 Carbon at 60 on its Intelligence Index v5.0.1. That is fourth of 74 models in its class, not first. Claude Fable 5.2 sits above it at 65.
The same page adds detail Google's post leaves out. At list pricing of $4 input and $20 output, running the index costs $7.76 per task. Carbon is slow: 24.9 minutes per index task against a median near 12.7. It is also extremely verbose, emitting 205 million output tokens across the index against a median around 73 million. Output speed lands at 82 tokens per second.
Put those together and a specific risk appears. Carbon's advantage in Google's benchmarks comes with a token habit that eats into the price advantage, and a latency profile that matters if the job is interactive. Neither fact kills the product. Both belong in the buying decision, and neither appears in the launch post.
The honest reading is that this is a genuinely frontier model with a distinctive cost structure, not a model that has lapped the field. The Verge reached a similar conclusion: on Google's own benchmark claims, Carbon is one of the strongest systems available and still behind Fable 5.2, whose release landed one day earlier.
Where the claims get stress-tested
Three of Google's examples are more falsifiable than the demos, because they describe work with a checkable output.
The first is the AlphaProof Code result, which DeepMind published separately on 10 November. Agents powered by Carbon produced Lean proofs about Signal's real implementation, work DeepMind puts at about four weeks against a historical norm of years for expert teams. DeepMind's own post says the artefact exceeds six million lines of Lean, with humans reviewing roughly one percent of the codebase. Some launch coverage rounds that to 6.5 million; the primary source is the safer figure.
The second is the Wiz engagement. Google says the security firm used Carbon to scan two mainstream SaaS providers and surface real zero-day vulnerabilities in under two hours, work that would ordinarily take weeks. That is a specific, time-bounded claim from a named third party, which makes it more useful than any demo reel, though Wiz's own write-up is the document to wait for.
The third is the science programme. Google is offering up to six months of access to 10,000 researchers across physics, chemistry, biology, materials and climate science, aimed at hypotheses shelved for being too expensive to test. The supporting number is CASP17, the protein-structure blind test, where Google says Carbon raised the retrieval of distant viral-protein homologues by a factor of 40 across seven low-similarity families.
What to watch, and how to price it
The swarm architecture points at something the whole industry is converging on: not one big model doing everything, but a planner that delegates to cheap executors and a critic that tries to break the result. That shape shows up everywhere now, from ticketing classifiers to proof search, and it changes what "best model" means. The right question stops being which model tops a chart and becomes which combination finishes your task at the lowest total cost per task, including the human review you still have to pay for.
That is the frame we use at aibookmarks.net when we index AI tools: what job does it finish, at what cost per finished job, and how much of the output can you verify without redoing it yourself. Carbon is a good test case for that filter, because its price looks excellent per token and merely competitive per independent task once verbosity is priced in.
Three things would settle the argument. Independent per-task cost measurements on real workflows rather than index tasks. A reproduction of the /swarm architecture outside Google's walled garden, ideally on a codebase Google has never seen. And evidence that the adversarial validator catches failures at a rate better than the 50-50 coin flip a demo can hide.
The launch is a pricing story wearing an agent costume
Strip the staging away and Google shipped two things: a frontier model priced aggressively for the next three weeks, and a preview of a multi-agent orchestration layer that will decide whether that model gets used for long jobs.
The model is strong and, by independent measure, fourth rather than first. The swarm is the more consequential release, because it is the part that changes how work gets divided and billed, and because it is the part with the least public evidence behind it.
Buy the model on the price, if the token bill is what constrains you. Wait on the swarm until someone outside Google publishes a failure rate. And keep the doubling at the end of November in the spreadsheet, because a launch price is a marketing decision with an expiry date.
---
Google's Gemini team, with independent benchmark data from Artificial Analysis and reporting from The Verge. Primary source: Google's 10 November 2026 launch post; the AlphaProof Code figures come from DeepMind's separate 10 November announcement.
Published by AI Bookmarks · More AI news →