Gemini 4 Carbon goes wide: Google's strongest model opens up

Google opened Gemini 4 Carbon to paying developers, enterprises, and AI Ultra subscribers on November 10, 2026 — two weeks after the model debuted on October 28 behind a defenders-only gate. Google calls Carbon its strongest model, built for the work that tests a model's limits: complex multi-step coding, long-horizon agentic workflows, and open-ended scientific research.

When Carbon was announced on October 28, independent evaluator Artificial Analysis ranked it first on Intelligence Index v5.0 at 59 points — ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, and Gemini 4 Argon at 50. On the live board as of November 10, Carbon (high) scores 60 on Index v5.0.1 and ranks behind Claude Fable 5.2 (max, 65; high, 62). In Google's own comparison table, Carbon wins 15 of 17 benchmark rows, including Terminal-Bench 4.0 at 65.2% against Astra's 59.1% and Opus 5.5's 59.6%. Every figure in that table is Google's or an evaluator's own; no independent replication has been published.

Google leads with three cybersecurity numbers that cover the defender's workflow. On SEC-bench Pro, which tests finding real vulnerabilities in real software, Carbon scores 88.2%. On ExploitGym, which tests whether a found flaw can actually be exploited, it scores 54.8%. On Collinear's CWE-bench, which tests turning a validated finding into a working root-cause fix, it reaches 71.0%. Google ran these three benchmarks without guardrails.

Google says early testers are already seeing results in practice. In Project Fortify, Google's autonomous scanning, triage, and remediation system, Carbon hit 100% recall on scanning and 95% precision on triage on Google's internal benchmark, and improved on Argon's already strong recall by more than 25% on open-source code. Wiz's Scan for Good team used Carbon to find 0-day vulnerabilities in two major SaaS providers in under two hours — work Google says would normally take weeks. Google's own vulnerability research team used it to prove two bugs in popular open-source software were exploitable, comparable in severity to the recent PHP2Shell exploit, getting them re-classified as critical and patched.

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than twice lower cost per task than the next best external model. On the new Multimodal Knowledge Agent benchmark Google built with Artificial Analysis, Carbon ranks second while costing more than six times less per task on average than the top model.

The price itself is the same as Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. That rate runs through November 30, then doubles to $4 and $20. Argon stays Google's recommended choice for most software engineering and knowledge work — Carbon is the model for the harder jobs, and it uses more tokens to explore, reason, and verify.

On its 375-task agentic honesty benchmark, Carbon passed 92% of tasks against 70% for Argon — declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time. It blocked 99% of attack trajectories in simulated cyberattack testing, cut the per-attempt breach rate to 0.8% from Argon's 3.1%, and reached a 0% prompt-injection attack success rate on Gray Swan's indirect prompt injection benchmark, the first time that milestone has been reached. Google also says prompt-injection robustness remains an area for continued improvement, and that its red teaming has shown models could still be improved against sophisticated attacks in some distributions.

The record is not a clean sweep. On FrontierSWE 2, GPT-6 Astra leads at 65.5% against Carbon's 61.2%, and on the OSWorld 2.1 offline subset Claude Opus 5.5 leads at 87.2% against Carbon's 81.5% — both from Google's own comparison table. Carbon led the composite index at announcement, though not every individual test; as of November 10, Artificial Analysis ranks Claude Fable 5.2 ahead of Carbon on the live index.

Carbon is available now through the Gemini API in Google AI Studio and Gemini for business for all paying developers and enterprises, and to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. Google is also previewing /swarm in Antigravity, which turns one prompt into a coordinated team of agents with roles like architectural design, coordination, and auditing, tracked in real time.

Cybersecurity defenders can apply to the Fairwind Program for Carbon for cyber. And scientists can apply for free access to Argon and Carbon in Antigravity for six months through the program Google opened four weeks ago for 10,000 scientists — thousands are already onboarded, with weekly rolling admission subject to eligibility, capacity, and selection.