Gemini 4 Carbon Launches: Beats Fable 5.2 at Half Cost, Finds Critical Vulns

Google has released the Gemini 4 Carbon model via API, making it available to Ultra subscribers ($100/month). The launch follows the recent release of Anthropic's Fable 5.2 and positions Carbon as a high-reasoning model targeting complex multi-step programming, debugging, and open-ended research.

Benchmark Results and Cost Efficiency

In three benchmarks — CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench — Carbon ranks first. Crucially, each task costs over 50% less than the second-place model (Fable 5.2). On AutomationBench, Carbon's high-reasoning tier scores 53.88% versus Fable 5.2 Max's 51.45% (with default fallback) and Argon's 51.29%. At list prices, the average cost per task is $2.12 for Carbon , $4.64 for Fable 5.2 Max, and $1.70 for Argon.

Real-World Demonstrations

Apollo lunar module reconstruction: Carbon analyzed historical images and technical drawings to rebuild a 3D model and generate a video.

Bird sound simulator: Starting from research on avian vocal structures, Carbon implemented an interactive 3D simulation that synthesizes six bird species in real time.

Marble physics game: A user test (X @kimmonismus) produced a marble game with realistic physics; lighting and initialization still need refinement.

Video-to-3js reconstruction: Carbon reconstructed real video segments using Three.js (X @emollic).

Product photography: Carbon created promotional images, leveraging the "nano banana" tool.

Vulnerability Discovery

Google disclosed that Carbon recently uncovered two critical vulnerabilities in a popular open-source project, with severity comparable to the historic PHP vulnerabilities. Patches have been released; full advisories will follow coordinated disclosure.

Agent Swarm Orchestration (/swarm)

A new feature in Google Antigravity, /swarm, lets a single prompt spawn a team of agents. Carbon acts as architect — designing, coordinating, and reviewing deep-reasoning tasks — while faster lightweight models execute in parallel workspaces and verifier agents check each change. In a demo, seven municipal planning PDFs were fed to the swarm; after 10 overnight iterations the agents delivered an interactive Three.js model of Google's office building comprising ~906,000 triangles .

Pricing and Token Consumption

During the introductory period (until November 30), Carbon is priced at $2 per million input tokens and $10 per million output tokens (50% off). Access is via Gemini API, Gemini for Business, Antigravity, CodeMender, and the Gemini App (Ultra only). Google explicitly warns that Carbon consumes more inference tokens than Argon because it explores multiple reasoning paths and extends verification loops for complex tasks. For daily work, Google recommends using Argon as the primary model.

Availability and Reception

Carbon is currently restricted to Ultra subscribers, a limitation that has drawn criticism on social platforms. The model's benchmark dominance and cost advantage are offset by higher token usage and limited access.

References: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/, https://x.com/koraykv/status/2120186714358702187, Quantum Bit reporting.