> ## Content Index > Fetch the complete content index at: https://www.testingcatalog.com/llms.txt > Use this file to discover other available public pages before exploring further. # Google tests a new Gemini 4 Carbon checkpoint on Arena - URL: https://www.testingcatalog.com/google-tests-new-gemini-4-carbon-checkpoint-arena/ - Published: 2026-11-01T08:43:59.000Z - Updated: 2026-11-01T08:43:59.000Z - Description: Days after Google unveiled Gemini 4 Carbon for cyber defenders, an updated Carbon checkpoint has surfaced on Arena. Testers say it's faster, better at 3D, and thinks a lot longer. - Author: Dwaybe Cubbins - Tags: Google, Latest AI News, AI Rumours, Gemini News Google unveiled [Gemini 4 Carbon on October 28](https://www.testingcatalog.com/google-unveils-gemini-4-carbon-for-cyber-defenders/), but only a subset of cyber defenders can use it for now, with developers and consumers due to get it later. That hasn't stopped the model from changing in public. Testers on Arena are now reporting a new Carbon checkpoint, and early reactions seem more positive than they were last time. > I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. > > Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of [@ArtificialAnlys](https://x.com/ArtificialAnlys?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com)… [pic.twitter.com/GN9LqZ7Zcf](https://t.co/GN9LqZ7Zcf?ref=testingcatalog.com) > > — Sundar Pichai (@sundarpichai) [October 28, 2026](https://x.com/sundarpichai/status/2115489219804352921?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) X user Bee was one of the first to put the spotlight on the new checkpoint results, noting that it feels *"noticeably better and faster than the last one."* Their post quotes another tester who says a build called Carbon-H is being served on Arena under an Argon label and is much better than earlier Carbon builds at coding, design, and 3D. That same tester notes some outputs look *"strangely similar to Opus 5.5."* > Gemini 4 Carbon's checkpoint has been updated on arena! > > And it came right when I'm on a trip and can't test it! > > It feels noticeably better and faster than the last one.. in short google is cooking.. > > Anyways I'll be sharing some outputs i find and I'll have grokbot run some… [https://t.co/wq5fHzAQPx](https://t.co/wq5fHzAQPx?ref=testingcatalog.com) [pic.twitter.com/w7azSxNu3H](https://t.co/w7azSxNu3H?ref=testingcatalog.com) > > — Bee (@thtbee\_) [October 31, 2026](https://x.com/thtbee%5F/status/2116384917869457670?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) This isn't the first time we've seen outputs from Gemini 4 Carbon either. A couple of weeks ago, we shared some of the early outputs from the model. However, at the time, testers claimed that while the 3D work it produced was good, it was painfully slow and weak at front-end coding. With the latest tests, it looks like speed is no longer a major issue, and 3D work is still where it shines. User Leo posted a side-by-side of Carbon and Gemini 4 Argon building a 3D cow. Argon still assembles it from blocks, while Carbon uses rounded and pointed shapes, and even diamond pieces, for a cleaner model. > Gemini 4 Carbon Vs Gemini 4 Argon 3D 建模测试直观对比： > > 可以很明显的看到，Argon 的奶牛还是有方块来建模 > 而 Carbon 已经会使用圆角和尖角，甚至菱形的模块来组件一个更美观的奶牛模型 > > 进步还是非常明显的。。。 [pic.twitter.com/BDvs3kanw5](https://t.co/BDvs3kanw5?ref=testingcatalog.com) > > — Leo · Building w/ AI (@leo114119) [November 1, 2026](https://x.com/leo114119/status/2116714521096265749?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) Meanwhile, notable leaker Chetaslua says Carbon is *"genuinely so good"* in Three.js and shows *"no laziness at all."* They add that it now really thinks before answering, sometimes for 8 to 10 minutes. Frontend is still the sore spot, and they hope Google will now *"fix webdev and slop UI."* > 🚨 Gemini Carbon 4 genuinely is so good in threejs > > this is a google model we are praising , no laziness at all > > now it think like really thinks before output sometimes 8-10 min > > This is seriously going to be a great model , i just hope now google fix webdev and slop ui , buy… [pic.twitter.com/77MO3h2HHc](https://t.co/77MO3h2HHc?ref=testingcatalog.com) > > — Chetaslua (@chetaslua) [October 31, 2026](https://x.com/chetaslua/status/2116578168874438910?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) Another user, OltAi, also tried a pixel-style T-Rex animation in Carbon, GPT-6 Astra, Opus 5.5, and GPT-6.1 Terra, giving them scores of 9.5, 8.8, 9.3, and 4.9 out of 10, respectively. Since these are one person's ratings of a single prompt, they're best viewed as a rough guide. Even so, Carbon's version showed the most movement of the four. > Gemini 4 Carbon vs GPT-6 Astra vs Opus 5.5 vs GPT-6.1 Terra. One dinosaur challenge, four very different results. > > And this comparison is getting interesting. > > The task? Bringing a pixel-style T-Rex to life in a desert environment, complete with movement, shadows and surrounding… [https://t.co/mPF4DEOAXh](https://t.co/mPF4DEOAXh?ref=testingcatalog.com) [pic.twitter.com/rYxp34zFlJ](https://t.co/rYxp34zFlJ?ref=testingcatalog.com) > > — OltAi (@UncleKripro) [October 31, 2026](https://x.com/UncleKripro/status/2116671406020710650?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) That said, apart from these tests on Arena, Carbon is also being used for the work Google built it for. Security company Wiz's head, Nagli, says they keep asking it how it worked out complex exploit chains and have *"never had so many novel discoveries while bug hunting."* > I'm really enjoying Gemini 4 Carbon, I always find myself asking it how it figures out its way to complex chains - never had so many novel discoveries while bug hunting > > If you’re still not using AI to protect and patch your own public web apps, you should’ve started yesterday. [pic.twitter.com/N8oavTt1kK](https://t.co/N8oavTt1kK?ref=testingcatalog.com) > > — Nagli (@galnagli) [October 31, 2026](https://x.com/galnagli/status/2116479102442234094?ref%5Fsrc=twsrc%5Etfw&ref=testingcatalog.com) It's worth keeping in mind that Arena doesn't show which checkpoint provided the answer, the outputs are cherry-picked, and Google hasn't commented on an update. Still, swapping in a stronger build so soon after the announcement suggests Google is continuing to fine-tune Carbon ahead of a wider launch. Google has already said Carbon will launch at Argon's introductory price of $2 per million input tokens and $10 per million output tokens. The company is also working on other improvements. DeepMind [previewed Genie 4](https://www.testingcatalog.com/icymi-google-deepmind-previews-genie-4-with-native-audio/) with native audio last week, and Antigravity is [getting ready for cloud execution](https://www.testingcatalog.com/cloud-execution-is-coming-to-antigravity-soon/) as Jules heads for shutdown on December 31\. A faster, stronger Carbon could be a great fit in both places. If Google can close the frontend gap as well, a wider release could be warmly welcomed.