Google Opens Gemini 4 Carbon: An Agentic Model 2–6x Cheaper Than Fable 5.2

Anyone who has tried handing an agent a big job, such as a migration across dozens of files, knows how it ends. The model loses the thread, forgets what it promised three steps ago, and confidently reports a success that never happened. Gemini 4 Carbon, opened by Google on November 10, targets exactly this pain. The question is how far the numbers from the release apply to your project.

What Carbon is and how it differs from Argon

Carbon is the flagship model of the fourth Gemini generation. It is built on Gemini 4 Argon, which came out a month earlier. Google positions them this way: Argon is the fast, economical choice for most engineering tasks, while Carbon is for cases that "test the model at the limit of its intelligence." These include multi-step debugging, architectural design, and open-ended research.

The price of depth is tokens. Carbon thinks longer and runs self-verification loops. By enthusiasts' estimates, it spends about 205 million thinking tokens on a run of the composite index, versus 113 million for Argon.

Technical details: a single context of 1,048,576 tokens (input, output, reasoning, and tool calls are counted together), LOW/MEDIUM/HIGH thinking levels with MEDIUM as the default, and direct ingestion of Office and OpenDocument files. One important detail for those working through the API: the model ignores the temperature, top_p, and top_k parameters. If your pipeline depends on them, you will have to rework it.

The numbers: where Carbon wins and where it doesn't

On the Artificial Analysis index, Carbon scored 59 points in v5.0 and 60 in v5.0.1. That is second place behind Claude Fable 5.2. However, the cost per task is about $7.76 versus $26.54 for the leader. On the new visual agentic benchmark MKA (blueprints, clinical images, financial tables), Carbon scored 54% versus 57% for Fable 5.2, while costing roughly six times less.

In applied agentic tests the picture is better: first place on CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. According to Google, a task there costs more than half as much as with the closest competitor. On GraphWalks, which tests the ability not to lose the thread across long trajectories, Carbon scored 96.2% versus 84.2% for Argon.

In pure coding the results are more modest. On FrontierSWE 2 it scores 61.2% versus 65.5% for GPT-6 Astra. On OSWorld-2.1 it reached 81.5% versus 87.2% for Claude Opus 5.5. On SWE-sweep the result is only 9.6%. Most of the figures are published by Google itself, so independent tests on your own tasks are a good idea.

/swarm and cybersecurity

Antigravity has gained a /swarm mode (still in preview). A single prompt spins up a team of agents. A coordinator hands out work to workers in isolated git worktrees, so edits do not conflict. Verifiers run in parallel, and their only job is to break and challenge diffs. Agents communicate through a shared message board, and idle ones consume nothing. Carbon acts as architect and auditor, while routine work goes to lighter models. Inside Google, the mode was run for five weeks on multi-package migrations and root-cause searches.

The main story of the release is security. SEC-bench Pro (finding real vulnerabilities): 88.2%. ExploitGym (verifying exploitability): 54.8%. CWE-bench (patching): 71.0%. Wiz found zero-days at two large SaaS providers in under two hours, and the Fortify system showed 100% recall and 95% precision. It was precisely because of these capabilities that the model was initially given only to participants in the Fairwind program for defenders.

"Honesty" deserves a separate mention. On an internal set of 375 tasks, Carbon passed 92% versus 70% for Argon. It refuses to fill in missing experimental data in 98% of cases and points out methodological gaps in the assignment on its own in 68% of cases.

What to do with this

Access is open through the paid Gemini API (AI Studio, Android Studio), Gemini for business, and Antigravity. The Gemini app and AI Mode in search are available to Google AI Ultra subscribers (from $100 per month). Scientists get six months of free access to Argon and Carbon in Antigravity.

The practical takeaway is simple. Don't move everything to Carbon: for routine work, Google itself recommends Argon. Use Carbon where the agent runs for a long time and an error is costly. That means migrations, security audits, and research pipelines. Budget in advance for higher token spend on reasoning and run the model on your own set of tasks. If your main criterion is cost per solved task rather than the absolute maximum quality, Carbon currently looks like one of the most interesting options.