Skip to content

Gemini 4 Carbon launches at Argon's token price

Google DeepMind opened Gemini 4 Carbon to paying developers on Nov 10, 2026, pitching it at complex agentic work rather than everyday coding.

Draftship Editorial3 min read

Gemini 4 Carbon, which Google DeepMind describes as its strongest model, is now available to developers, enterprises, scientists and consumers as of Nov 10, 2026. The company had given Carbon to a set of cybersecurity defenders through its Fairwind Program two weeks earlier. Google positions Carbon as a step beyond Gemini 4 Argon rather than a replacement: the announcement says Argon remains the right choice for most software engineering and knowledge work.

What Google actually shipped

According to the DeepMind announcement, Carbon is available through the API in Google AI Studio and Gemini for business for all paying developers and enterprises. It is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, wrote the post.

How it works and what it costs

Pricing is the same introductory rate as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. The source does not say how long "introductory" lasts.

The split between the two models matters when picking one. Google says Carbon is built for tasks that test the limits of a model's intelligence, such as complex multi-step coding and debugging and open-ended scientific research, and that it uses more tokens than Argon because it explores thoroughly, reasons deeply and runs extended verification loops. That extra token spend is deliberate in Google's framing, and the benchmark claims are stated as cost per task rather than cost per token.

Benchmarks and a multi-agent preview

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task compared with the next best external model. On the new Multimodal Knowledge Agent benchmark, built with Artificial Analysis to test agents reasoning over dense files such as charts, images and schematics, Carbon ranks second while being more than 6X lower in average cost per task than the top performer. Google also reports top placement out of 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark.

Alongside the model, Google is previewing /swarm in Google Antigravity, which turns a single prompt into a coordinated team of agents that can be tracked in real time. Carbon can take roles that need the deepest reasoning, such as architectural design, coordination and auditing, while faster models run in parallel isolated workspaces with adversarial verifiers testing each change. One example given: seven City Council planning PDFs and a single prompt produced a 906K-triangle Three.js reconstruction of Google's Platform 37 office across 10 overnight loops.

What it changes for app builders

The build-and-verify loop is the part most relevant to people shipping web apps. Google says Carbon runs build-and-verify loops, recovers from errors, and checks both its code and the rendered UI to produce complete working applications, and that it handles multi-package migrations and cross-subsystem root-cause investigations. If that holds up outside Google's own demos, the bottleneck moves from producing a first draft to reviewing what the model produces.

The cost-per-task framing is the practical takeaway. A model that burns more tokens can still be cheaper per finished job, so comparing sticker token prices across models tells you less than comparing the cost of a completed task. That is inference on my part: the source gives no dollar figures per task and no token counts for its examples.

Limits and open questions

The announcement lists no context window, no rate limits, no latency figures, no regional availability and no general availability date for /swarm. The cybersecurity results, including 100% recall on scanning and 95% precision on triage in Google's Project Fortify, come from Google's internal benchmark per the source. The benchmark wins are Google's own summary, and the post text does not include methodology for them.

For people building with AI

Model choice is becoming a real decision rather than a default. Draftship includes a choice of AI model on paid plans, described in using the AI chat and choosing a model, so the same cost-per-task question applies to whatever you select there. Draftship is not affiliated with Google DeepMind, and nothing in this announcement concerns Draftship itself.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.