Gemini 4 Carbon is Google's strongest model yet, built for multi-day agentic coding, 3D web dev, and precision knowledge work. Here's what it does.
What is Gemini 4 Carbon? #
Gemini 4 Carbon is Google DeepMind’s newest frontier model, positioned as the company’s strongest release to date. Google built it around four target use cases: multi-day agentic engineering, high-precision knowledge work, frontier discovery tasks (physics, genomics, drug discovery), and full-stack front-end and 3D web development. It ships alongside Gemini 4 Argon, a faster model of the same size, and runs inside Google’s agentic coding environment, Antigravity, where it can browse the web, write and execute code, and operate unattended for extended stretches.
TL;DR #
- Four use cases define Carbon : multi-day agentic engineering, high-precision knowledge work (finance, legal, healthcare), frontier scientific discovery, and 3D/front-end web development.
- It can run unattended for very long stretches , with one documented test (a historic fort rebuild) still going after roughly 20 hours and dozens of self-review passes.
- Benchmark results are strong but not undisputed number one : Carbon topped the Artificial Analysis Intelligence Index at launch, but Anthropic’s Claude Opus 5.5/Fable 5.2 release pushed past it days later.
- Carbon wins heavily on cost-efficiency , running roughly a third of the per-task price of Fable 5.2 on the Artificial Analysis index while landing only a few points behind on raw intelligence.
- It has a strong habit of building its own verification tools , writing frame-by-frame auditors, fresh “reviewer” sub-agents, and test harnesses to check its own output rather than trusting its own memory of the work.
- Pricing launches at an introductory $2 input / $10 output per million tokens through the end of November, rising to a standard $4/$20, matching Opus 5.5’s standard rate.
- It scores notably low on a stricter hallucination measure , bluffing only around 11% of the time when it doesn’t know an answer, a meaningfully better rate than several competing frontier models on that same test.
What are the four use cases Google built Carbon for? #
Google frames Carbon around four domains rather than a single benchmark score. The first is multi-day agentic engineering: long-running coding tasks that involve deep reasoning, precise code edits across large codebases, verification loops, and cybersecurity review work like scanning a codebase for vulnerabilities before shipping. The second is high-precision knowledge work aimed at finance, legal, and healthcare use cases, where exhaustive retrieval and constraint-checking matter more than speed, and where Google claims zero fabricated citations in its own testing. The third is frontier discovery, covering physics, mathematics, quantum computing, genomics, and drug discovery, a category most developers won’t touch directly but that matters for research labs pushing the model’s reasoning into unexplored territory. The fourth, and the one most directly relevant to working developers, is full-stack front-end and 3D web development: realistic physics, interactive game mechanics, and interfaces the model can visually self-check as it builds them.
How does Carbon perform on long, unattended tasks? #
This is where Carbon’s design choices show up most clearly. In testing, the model was given large, open-ended 3D and research projects with instructions to keep improving the output over many passes without stopping to ask questions. In one case, it was tasked with rebuilding a historic fort in Peshawar, Pakistan using only public sources and a 3D engine released about a month earlier. It logged roughly 30 sources during a research phase, including colonial-era records and old aerial photography, reconstructed the building’s history across different rebuilding periods, and flagged undocumented sections as “inferred” rather than guessing silently. When the 3D engine itself couldn’t do something the project needed, the model edited the engine’s source code directly, including two bug fixes, across six total changes.
The review process matters as much as the build process. Every ten passes, a fresh sub-agent scored the output against real reference photos with no memory of the original plan, starting at 3 out of 10 and climbing toward 5 out of 10 (with 8 considered the threshold for photorealism) after roughly 65 passes and 20 hours of unattended work. That pattern, a state file tracking progress and open problems, a scheduled fresh reviewer, and a minimum number of required passes, showed up repeatedly across other long-running demos, including a 3D recreation of a browser game scene and a physics simulation involving colliding galaxies on a small GPU cluster.
How good is it at coding and 3D web development specifically? #
Nothing gets lost.
312 apps across 41 teams · all in one place
Independent testing found Carbon capable of producing fairly complete native applications from short prompts. One test asked for a native macOS Pomodoro timer app with the instruction to make it “the ultimate” version with useful features, and Carbon returned a fully built Swift app complete with a menu bar controller, project tracking, and focus audio, none of which were explicitly requested. A second test asked it to research two existing note-taking apps (Bear and Obsidian) and build something combining their best features, again fully in Swift, without any reverse-engineering tools. Both were compiled and signed, producing runnable Mac apps rather than just source code.
On 3D web work, Carbon rebuilt complex real-world scenes, a Singapore skyline centered on Marina Bay Sands, a Bauhaus-era house from original architectural survey drawings, and an abstract scene described only as “a storm in a teacup,” largely from single prompts followed by light iteration. A recurring theme across these demos is that Carbon tends to write small programs to audit its own output (checking an animation frame by frame, for instance) rather than relying on its own judgment from memory.
How does Carbon compare to Opus 5.5 and GPT-6 Astra on benchmarks? #
On Terminal-Bench 2.0, a measure of agentic coding performance in a terminal environment, Carbon scored 65.2% against roughly 59% for both Claude Opus 5.5 and GPT-6 Astra. On a knowledge-work automation benchmark, it scored around 54% versus the low 40s for competing models. At launch, Carbon briefly held the top spot on the Artificial Analysis Intelligence Index at a score of 60, one point ahead of Opus 5.5.
That lead didn’t last. Within days, Anthropic released Claude Fable 5.2 (part of the Opus 5.5 line), which pushed the index to 65, putting Carbon five points behind. On the Remote Labor Index, a benchmark measuring what share of real freelance tasks a model can fully automate, Fable 5.2 reached 41%, compared to GPT-6 Astra’s 21% months earlier. Carbon’s showing on GDPX, an aggregated benchmark built from real third-party industry tests (quant trading, chip design, and similar domains), landed at 10.9%, behind Fable 5.2’s 16.4% but ahead of GPT-6 Astra’s 9.6%, while reportedly costing roughly a fifth as much per task as Fable 5.2.
One benchmark favors Carbon directly: a stricter hallucination measure that tracks what a model does when it genuinely doesn’t know an answer. Carbon bluffed only about 11% of the time under that test, compared to 59% for Fable 5.2 under maximum effort settings, a reversal from Fable 5.2’s lead on a separate, more forgiving hallucination metric.
Is Gemini 4 Carbon worth using over competing models? #
The honest answer depends on what you’re optimizing for. On raw intelligence, Carbon is no longer the top scorer; Fable 5.2 edged ahead within days of Carbon’s debut. But cost tells a different story. Carbon runs at roughly a third of Fable 5.2’s per-task cost on the Artificial Analysis index ($7.80 versus $26.50), while trailing by only five points on the intelligence ranking. For teams running agentic workloads at scale, that cost gap is likely to matter more day to day than a handful of benchmark points.
made for the enterprise.
↑
Approved
Approved
Approved
Approved
Approved
Approved
Approved
Approved
Approved
Approved
Carbon also thinks for longer than most models by default, averaging around 26 minutes per task on the Artificial Analysis benchmark against a roughly 13-minute median, and generating far more tokens per task than typical. That makes it better suited to being launched on a task and left alone than used as a conversational assistant. Pricing starts at an introductory $2 input / $10 output per million tokens through November 30, rising to a standard $4/$20, matching Opus 5.5’s standard rate, with 95% off cached input tokens.
Frequently Asked Questions #
What is Gemini 4 Carbon built for?
Google designed it around four areas: multi-day agentic engineering tasks, high-precision knowledge work in fields like finance and healthcare, frontier scientific discovery, and full-stack 3D and front-end web development.
Is Gemini 4 Carbon better than Claude Opus 5.5 or GPT-6 Astra?
It depends on the benchmark. Carbon briefly led the Artificial Analysis Intelligence Index before Anthropic’s Fable 5.2 release pushed ahead by five points. Carbon beats GPT-6 Astra on several agentic and knowledge-work benchmarks and costs significantly less per task than Fable 5.2.
How long can Gemini 4 Carbon run unattended?
In testing, it ran continuously for roughly 20 hours and 65+ review passes on a single open-ended 3D reconstruction project, using a state file to track progress and periodic fresh reviewer agents to score its own output objectively.
What does Carbon cost to use?
Standard pricing is $4 per million input tokens and $20 per million output tokens, matching Claude Opus 5.5. An introductory rate of $2/$10 runs through November 30. Cached input tokens cost 95% less.
Does Carbon hallucinate less than other models?
On one stricter measure that tests what a model does when it doesn’t actually know an answer, Carbon bluffed about 11% of the time, notably better than Fable 5.2’s 59% rate under the same test, though results varied across other hallucination benchmarks.