Google DeepMind's Gemini 4 Carbon, a frontier Gemini 4 language model with a staged rollout to paid users and API.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Gemini 4 Carbon (also referred to as Gemini Carbon, Gemini 4 Argon or Carbon High in creator coverage) is a frontier model from Google DeepMind in the Gemini 4 family. The DeepMind site lists "Gemini 4 Carbon is now available" among its latest news items.
Creators report it ranks first on the Artificial Analysis Intelligence Index v5 (score 59) and leads on a range of coding, agentic, science and cyber benchmarks, with strong results on Vals Index, Vending Bench and medical reasoning. Output of up to 1M tokens is reported, as is lower token use per task than Opus. One reviewer found it cheaper than a competing frontier model but noticeably behind it in quality on demo tasks.
Reports on access differ. Early coverage described a rollout starting with cyber defenders and trusted partners. Later coverage says it is available to Pro/Ultra subscribers and paid API users. Reported introductory API pricing is $2/$10 per million input/output tokens, rising to $4/$20. These figures come from creators, not official pages, and should be checked against Google's pricing documentation.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Announced Oct 28 and publicly released this week, second on the leaderboard behind Fable 5.2. Available in the API and to AI Ultra subscribers; $2/$10 per million tokens, rising to $4/$20. In the Megabonk test it looked better than Fable's, costing 46 cents and 11 minutes.
Google's new frontier model, now available. The creator demos an Excel clone, a Fall Guys clone and a Rube Goldberg machine. He finds it good, cheaper than Fable 5.2, but well behind Fable in quality.
Google's announced but not publicly released model. It beats Fable 5.2 by about 2 points on business workflows and was ranked just behind Fable 5.2 on Artificial Analysis.
Google's frontier model, ranked number one on Artificial Analysis, with top accuracy, lowest hallucination, and Vending Bench leadership. Currently limited to an initial cohort of cyber defenders.
Step up from Gemini 4 Argon, not yet publicly available, only for some early and cybersecurity testers. It tops Artificial Analysis above Opus 5.5 and GPT-6 Astra.
Earlier Gemini 4 model released two weeks before Carbon, used as the comparison baseline. Carbon shows big gains over it, e.g. Terminal Bench 65 vs 57.
Google's new frontier model, announced but not yet publicly released; rollout starts with cyber defenders and trusted partners. It ranks #1 on the Artificial Analysis Intelligence Index v5 (59) and leads on many coding, agentic, science and cyber benchmarks. Introductory pricing is $2/$10 per million tokens, rising to $4/$20, and it uses fewer tokens per task than Opus.
Google's latest flagship model, released to Pro/Ultra and paid API users. Strong benchmark results (top on Vals Index, Blueprint Bench, medical reasoning, near-saturating ARC AGI 3 with a harness) and up to 1M output tokens.
Google's new model, roughly caught up with GPT-6 Astra and Claude Opus 5.5 on most benchmarks. API price is $2/$10 per million tokens until end of November. The creator tested it on Beauc Bench, Mega Bonk and a Gary Busey 3D-printable STL bust.
Google's new frontier model, tested by the creator for about a week. It built a cyberpunk city, Rube Goldberg machine, Fall Guys clone, Unreal Engine ocean scene and an Excel clone, typically spending about 30 minutes per task. Benchmarks: Deep Sweep 77.9, Vibe Code Bench 91.9, Terminal Bench 57, OS World 69. Weaker at 3D than Astra and Opus.
Google's new frontier model, now available to AI Pro/Ultra subscribers. It is state of the art on several benchmarks (Vals Index, Automation Bench, legal, agentic coding), offers a 1M-token output limit, and costs $2/$10 per million tokens at introductory pricing. Creator tested SVGs, a Mega Bonk clone, a physics playground, scheduling, research, and 3D-printable STL generation, and concludes Google has caught up but it isn't a huge leap.
Google's new frontier model that on average beats GPT6 Astra and top Claude models on benchmarks and ranks first on several. It can output 1 million tokens and has the lowest hallucination rate (15%) on the Omniscience benchmark. Access is currently limited to trusted cyber defenders, with paid API and Ultra rollout next.
New frontier model claiming state-of-the-art benchmarks above Opus 5.5 and GPT-6 Astra, priced at $2/$10 with a 1M-token output limit. Only rolling out to trusted cyber defenders via the Fair Wind program. Artificial Analysis scored it tied with GPT-6 Astra.
Newly launched model. The creator had Perplexity Computer gather its benchmark data and chart it against Qwen 3.6, MiniMax M2.5 and Kimi K2.5.
Closest neighbours by meaning, not just by category.
Model · Google
Smaller, faster, lower-cost Gemini 4 model from Google DeepMind with a 1M-token context window.
Model · Google
Google's faster, lower-cost Flash model in the Gemini 3.5 family, aimed at agentic and coding workloads.
Model · Google
Google's fast, low-cost Gemini Flash model with strong agentic coding results, plus a Cyber variant for defensive security.
Model · Google
Google's fast, lower-cost multimodal Flash-tier Gemini model for the Gemini app, AI Studio and coding tools.
Model · Google
Google's Gemini 3.1 Pro multimodal large language model, available in the Gemini app and via the Gemini API.
Model · Google DeepMind
Google DeepMind's lightweight, low-cost Gemini 3.5 model aimed at fast, high-volume tasks.
