Gemini 4 Carbon is now available
AI-assisted digest, checked against the original source before publishing. How we work
Google DeepMind's strongest model opens to everyone, with a new multi-agent preview in Antigravity.
DeepMind frames the two as a division of labor rather than a swap. Argon remains the recommended pick for most software engineering and knowledge work, while Carbon is aimed at tasks that stretch a model's intelligence: complex multi-step coding and debugging, plus open-ended scientific research and discovery. Carbon deliberately consumes more tokens than Argon in order to explore thoroughly, reason deeply and run extended verification loops.
On independent benchmarks built to mirror real workflows, DeepMind says Carbon takes the top spot on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench while costing more than twice as little per task as the next best external model. Carbon is available through the API in Google AI Studio and Gemini for business at Argon's introductory price, and it is rolling out to Google AI Ultra subscribers across Google Antigravity, the Gemini app and AI Mode in Search.
For legal, finance and enterprise operations, the pitch combines frontier visual understanding, reasoning and exhaustive retrieval to synthesize grounded insights from large stacks of dense documents and data, while following strict schemas to produce polished spreadsheets, presentations and reports. DeepMind also argues that existing benchmarks poorly capture how agents reason natively over charts, images, blueprints, schematics and tables, so it partnered with Artificial Analysis on a new Multimodal Knowledge Agent benchmark.
The biggest new surface is a /swarm preview in Google Antigravity, which turns a single prompt into a coordinated team of agents tracked in real time. Carbon is assigned roles demanding the deepest reasoning, such as architectural design, coordination and auditing, while faster, lighter models run in parallel isolated workspaces alongside adversarial verifiers that test every change. Showcase runs included an interactive Three.js exhibition of the Eiffel Tower's construction built from historical photos and engineering plates, and a multiscale textile visualizer implementing a virtual-fiber finite element framework from an academic paper with minimal prompting.
What to watch: whether /swarm graduates from preview and how broadly it is exposed, whether Carbon's token-hungry verification loops retain their claimed cost edge at production volume, and whether the new MKA benchmark gains the credibility needed to make multimodal agentic performance a genuine buying criterion rather than another vendor-designed score.
Key facts
- • Priced at $2 per million input and $10 per million output tokens, matching Argon.
- • Ranks #2 on Artificial Analysis's new Multimodal Knowledge Agent benchmark.
- • A /swarm team produced a 906K-triangle Three.js reconstruction of Google's Platform 37 office.
- • That office build drew on seven City Council planning PDFs across 10 overnight loops.
- • The Eiffel Tower exhibition covers the landmark's 1887–1889 construction.
Why it matters
DeepMind is splitting its frontier lineup into a default model and a premium reasoning tier, giving builders a concrete routing rule: send verification-heavy or open-ended work to Carbon and keep routine engineering on Argon. The /swarm preview signals that the competitive frontier is shifting from single-model benchmarks toward orchestrating mixed-model agent teams with adversarial checks, so exp
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime
