Google Rolls Out Gemini 4 Carbon to AI Ultra, Keeps Argon

Carbon ranks first on the Multimodal Knowledge Agents (MKA) benchmark Google launched with Artificial Analysis, using three times fewer tokens than the next-best model across 200 tasks that test multimodal performance in agentic workflows. Following a preview at the end of last month, Google's Gemini 4 Carbon now reaches AI Ultra subscribers in the Gemini app, AI Mode and Google Antigravity.

The claim comes from Google, and the benchmark it points to was built with a Google partner. Every performance figure in the announcement originates with the vendor selling the model.

Which tiers actually get Carbon?

• AI Ultra subscribers: Gemini app, AI Mode and Google Antigravity.
• AI Pro and above: Argon remains available alongside Carbon; Google has not retired the previous model.
• Developers: Gemini API paid tier in AI Studio and Android Studio; AI Ultra subscribers can try Carbon in Google Antigravity and AI Studio.
• Enterprises: Gemini for business via the API or model drop down, Google Antigravity, and specialized agents including CodeMender and AlphaEvolve.
• Scientists and researchers: application-based access to both Argon and Carbon through Google Antigravity's highest usage tier, free of charge for six months, onboarded on a rolling basis and subject to eligibility, capacity and selection.
• Cybersecurity defenders: application to join the Fairwind Program for access to Carbon for cyber, also on a rolling basis.

In plain terms: the top consumer tier gets the new model first, and every other constituency has to apply or route through the API.

What Google says Carbon does better

Google draws a deliberate line between its two models. "Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model's intelligence." The company cites "complex, multi-step coding and debugging and open-ended scientific research and discovery" as the target work.

For legal, finance and enterprise use, Google says Carbon brings "higher-precision knowledge work, artifact generation, and data analysis," including the ability to follow "strict schemas and formatting to generate polished and precise spreadsheets, presentations, and reports."

What Google does not publish: task-level accuracy rates, latency or throughput figures, or a head-to-head comparison of Argon against Carbon on identical prompts. The lineup split is a vendor judgment call dressed as guidance.

The benchmark, and what it leaves out

Google's stated reason for building a new test is that existing benchmarks "don't adequately capture how agents reason natively over dense files with charts, images, blueprints, schematics, and tables." So it partnered with Artificial Analysis on the Multimodal Knowledge Agents benchmark, where Carbon ranks #1 while using 3X fewer tokens than the next best model across 200 tasks.

Absent from the announcement: the identity of the second-place model, how the 200 tasks were selected and scored, whether Artificial Analysis independently verified the token counts, and what a three-fold token reduction means for wall-clock time or total spend on a real project. One vendor-co-authored benchmark is a starting point, not a track record.

Pricing doubles after November 30

• Introductory rate until November 30: $2 per million input tokens, $10 per million output tokens.
• Standard rate afterward: $4 per million input tokens, $20 per million output tokens.

That is a 100% increase on both sides of the ledger. Google disclosed no volume threshold, commitment term or promotional condition attached to the introductory rate, and November 30 is the only deadline named. The company has also published no separate per-seat pricing for Carbon inside the Gemini app, AI Mode or Antigravity, so API list prices are the only public numbers so far.

/swarm and the agent model

Google Antigravity is previewing /swarm, which turns a "single prompt into a coordinated team of agents you can track in real time." Developers can assign Carbon to roles that demand the deepest reasoning — architectural design, coordination and auditing — while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change.

Google also flags three domain claims:

• Multi-agent engineering: multi-package migrations, cross-subsystem root-cause investigations and complex architectural redesigns.
• Web application development: build-and-verify loops, error recovery, and checks of both the code and the rendered UI to produce complete, working applications.
• 3D and creative coding: stronger visual design, spatial reasoning and attention to environmental detail.

/swarm remains a preview with no announced availability date, and the domain claims ship without published evaluation data. The claims are directional; the access is staged.

The test buyers face in December

The next hard date is set by Google itself: on December 1, Carbon's list price doubles and early adopters meet the model at full rate. Google has not said what independent evidence it will put forward before then, and the rolling application windows for researchers and cybersecurity defenders give no promise of a firm seat.