Google's Gemini 4 Carbon is now on the paid Gemini API and for Google AI Ultra subscribers. Prices, the 1M-token context, who gets it, and how it compares with Fable 5.2.
Google made Gemini 4 Carbon generally available on November 10, 2026. "Today, we're making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers," Koray Kavukcuoglu wrote on Google's blog ↗. The Verge ↗ confirmed the public rollout the same day.
Carbon isn't new to the people following it. Google announced it on October 28 ↗, but only gave it to cybersecurity defenders in its Fairwind Program. That was two weeks after Gemini 4 Argon ↗ launched on October 13. Now anyone on the paid API can call it.
Google pitches the two models for different jobs. Argon "remains the ideal choice for the majority of software engineering and knowledge work." Carbon is for "complex, multi-step coding and debugging and open-ended scientific research." It uses "more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." That last part matters for your bill, as you'll see below.
| Where | Who gets Carbon |
|---|---|
| Gemini API (AI Studio, Android Studio) | Paid tier only. Not on the free tier |
| Gemini for business | Paying enterprises, via API or the model menu |
| Gemini app and AI Mode in Search | Google AI Ultra subscribers |
| Google Antigravity | Ultra subscribers and enterprises |
| Science program | Free for 6 months for accepted researchers |
The consumer rollout is the catch. On Hacker News ↗, one Plus subscriber saw Carbon in the model list marked "Unavailable - Gemini 4 Carbon is currently available to Ultra plan subscribers." Another wrote that their Gemini app "still only" had 3.1 Pro, 3.6 Flash and 3.6 Thinking. The Verge says it's "not clear when — or if" lower tiers or free users will get it.
Per Google's model page ↗, the API model ID is gemini-4-carbon. It has a 1,048,576-token context window shared between input and output. Default output is 65,536 tokens, or up to the full 1M with long decode. It takes text, images, video, audio and PDFs, and returns text. Function calling, code execution, computer use, search grounding and structured outputs are all supported. Explicit context caching is not; caching is implicit only.
Carbon launches at Argon's introductory price. It doubles on December 1, per the Gemini API pricing page ↗:
| Tier (per 1M tokens) | Input to Nov 30 | Output to Nov 30 | From Dec 1 (in / out) |
|---|---|---|---|
| Standard | $2.00 | $10.00 | $4.00 / $20.00 |
| Batch and Flex | $1.00 | $5.00 | $2.00 / $10.00 |
| Priority | $3.60 | $18.00 | $7.20 / $36.00 |
Output pricing includes thinking tokens. Google's note says "every token the model processes is counted and billed," including internal tool instructions when grounding is on.
For scale: Claude Fable 5.2 lists at $8 / $40 and GPT-6 Astra at $10 / $50 ↗ for prompts up to 272K tokens. The Verge puts Carbon's launch price at "around a fifth the cost" of Astra and "a quarter the cost" of Fable 5.2.
Google's main numbers, all self-reported:
#1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, "at more than 2X lower cost per task compared to the next best external model."
#2 on the new Multimodal Knowledge Agent benchmark it built with Artificial Analysis, at "more than 6X lower" average cost per task than the leader.
Cyber results (via Fairwind): 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, from the October announcement.
Honesty: 92% on Google's 375-task agentic honesty benchmark, up from 70% for Argon.
Google is also previewing /swarm in Antigravity. It turns one prompt into "a coordinated team of agents." Carbon takes the planning and auditing roles, and lighter models do the parallel work. Our lead image is from Google's own Pelican Quest demo ↗, a 3D campus game that Google says Carbon built this way.
As The Verge notes, "companies can and do cherry-pick benchmark results." None of these numbers are ours.
Cost per task is already disputed. "everyone says Carbon is cheaper. not on this one," posted @_adishj ↗, who shared a run in Motion where Fable 5.2 cost $60.87 in 1h 30m and Carbon cost $67.53 in 2h 44m.
Early users are mostly positive. "Pretty Good tbh," wrote protosphinx ↗ after adding it to Proto. On HN, one user said an unfinished paper draft got "interesting points all the other AI's missed" from Carbon, while calling it "a sample of one."
Legal AI company Harvey said ↗ Carbon was "especially strong on deal management and case law research" in its early testing.
The skeptics have a fair point. "It uses more tokens than Argon? I already get very little out of my Pro subscription (using Antigravity) with Argon," one HN commenter asked ↗.
Reddit wasn't reachable for this story, so these reactions come from X and Hacker News.
Developers should test Carbon this month, because the price doubles on December 1. Until November 30, it's the cheapest frontier-class model per token. But Google itself says it thinks longer than Argon. Judge it on your bill per finished task, not the price per million tokens. The Motion comparison above shows how quickly that can go the other way.
Don't switch your default model yet. Google says Argon is still the right pick for most coding and knowledge work. Carbon is for the hard jobs: a bug you've been chasing for days, a migration across many packages, or a security review. Use it there and leave routine prompts on Argon or a cheaper model.
Consumers on Google AI Pro aren't missing out yet. Carbon is Ultra-only in the Gemini app. We wouldn't upgrade to Ultra just for this. If you're choosing an assistant, our ChatGPT vs Claude vs Gemini comparison covers the plans. For coding tools, see our best AI for coding guide and our AI coding assistant benchmark.