Tech • AI • Robotics • Game
Google Just Gives the World Its Most Powerful AI Model Ever: Carbon
9.4/10TL;DR
Google has broadly released Gemini 4 Carbon, a high-end AI model aimed at difficult coding, cybersecurity and research tasks, while rivals including Anthropic, OpenAI and Meta push further into AI-powered workplace software.
KEY POINTS
Google expands Carbon access
Gemini 4 Carbon was previously limited to a small cybersecurity cohort through Google’s Fairwind program, but is now available to developers and businesses through AI Studio and Gemini for Business, with broader access for Google AI Ultra subscribers in the Gemini app, Anti-Gravity tools and AI Mode in Search. Demis Hassabis announced the wider launch as Google positioned Carbon as its most powerful model to date.
Strong benchmark performance
Google’s new model ranked first on CursorBench 4.0, Perplexity’s Wander and Zapier’s Automation Bench, outperforming rivals such as Claude Fable 5.2 and Opus 5.5 at less than half the cost per task of the best outside model. It scored 96.2% on GraphWalks, reached 60 on the Artificial Analysis Intelligence Index, and tied for first on SkillSwitchBench. On a real-world investment research evaluation covering 23 frontier models, Carbon also placed first.
Aggressive pricing strategy
Carbon is priced at $2 per million input tokens and $10 per million output tokens, matching Argon and roughly aligning with GPT 6.1 Soul, with a 95% discount on cached input until November 30. After that, pricing is set to double to $4 and $20. Google is explicitly steering everyday coding and knowledge work toward Argon, while reserving Carbon for harder debugging, open-ended research and tasks where deeper reasoning justifies higher token use.
Knowledge work and agent swarms
Google is emphasizing document-heavy professional workflows in law, finance and operations, where Carbon can process charts, tables and images and return spreadsheets, reports or slide decks. On the new Multimodal Knowledge Agent benchmark, it placed second at 54.3% while costing under one-sixth of the top model’s average task price. In Anti-Gravity’s swarm mode, multiple agents can work overnight from a single prompt, producing outputs such as an interactive 3D model of Google’s Platform 37 building from seven planning PDFs and 206 engineering views after 10 rounds of iteration.
Coding and software engineering gains
Carbon’s strongest coding results appear in multi-agent engineering work, including cross-package migrations, bug tracing across subsystems and major refactors. One internal test had the model build a Rust UI library from scratch, rewrite a desktop app and return a working demo overnight. In another case, HubX reported that Carbon fixed 94% of 16 real bugs in a codebase with 14,000 tests, with zero regressions.
Cybersecurity focus remains central
Security was Carbon’s initial deployment area, and Google describes it as its strongest cybersecurity model so far. Through Fairwind, it led on SecBench Pro, Exploit Gym and CWE v1 for vulnerability finding, exploitation and remediation. In Project Fortify, Google’s internal scan-triage-repair pipeline, the model achieved 100% scanning recall and 95% triage precision, while on open-source code its recall was more than 25% higher than Argon.
Scientific research use cases
Google presented Carbon as a science tool for physics, math and life sciences, especially for combining data, testing hypotheses and solving multistep problems. Axiom Bio said the model topped its translational science tasks and identified previously unseen toxicity mechanisms. In encryption research tied to Signal, Carbon and other models helped generate more than 6 million lines of machine-checked Lean proofs verifying core security goals, with only about 1% of the code base requiring human review.
Broader AI office competition intensifies
While Google pushes high-end models, Anthropic is moving deeper into productivity software with Claude Meetings, Whiteboards, Sheets and Data, aiming to make Claude the front end for workplace tools. OpenAI has launched Space as a shared ChatGPT workspace, and Meta introduced Spaces for Muse Agent. The competitive battleground is shifting from supplying models to software vendors toward building the office suite directly and becoming the first application workers open each day.
CONCLUSION
The wider launch of Gemini 4 Carbon sharpens competition at the top end of AI, where model quality, pricing and integration into daily work are converging into a single commercial fight. The next phase will be decided not just by benchmark wins, but by whether these systems can deliver measurable gains in coding, security, research and office productivity.
Ask a question