New version of Carbon may have appeared on Arena
A model seen as a new version of "Gemini 4 Carbon," the top model Google announced on Oct. 28 (Oct. 29 Japan time), is being tested on Arena, a site that compares AI answers anonymously. The X account @Haleeeemahh, which shares information on unreleased models, posted on Oct. 31 (7:30 p.m. the same day Japan time) that a new checkpoint (a version of a model saved partway through training) it called "Carbon-H" was appearing on Arena under the name "Argon 4." According to TestingCatalog, the new version is appearing on Arena under the borrowed name of the already released Gemini 4 Argon. What the "H" in the name stands for has not been disclosed. Google has not commented on the matter.
On the same day, another account, @abhinavflac, had posted that the Gemini 4 Argon checkpoint on Arena had been replaced with a new ID as of the evening of Oct. 30 Pacific time (the morning of Oct. 31 Japan time). U.S. tech outlet TestingCatalog reported on Nov. 1, based on these two posts, that Google is testing a new Carbon checkpoint on Arena.
Testers call it faster and stronger at 3D
In tester posts reproduced by TestingCatalog, many said the new version is faster than the earlier one, better at 3D rendering and takes longer to think. Some said it is better than the earlier Carbon at coding, design and 3D alike, and others pointed out that some answers resemble Anthropic's Claude Opus 5.5. In one individual's comparison asking for a pixel-art dinosaur in a single prompt, Carbon scored 9.5 out of 10, Claude Opus 5.5 scored 9.3 and OpenAI's GPT-6 Astra scored 8.8.
On the other hand, there are also reports of remaining weaknesses. A tester who rated it highly for a program that draws 3D also posted that responses sometimes took 8 to 10 minutes and that its polish of website interfaces remained weak. When it first appeared on Arena on Oct. 21, it was also reported to be two to four times slower than Argon and weak at polishing interfaces. Arena does not reveal which version answered, the output examples circulating were chosen by the posters, and the assessments are only individual impressions.
Public version leads third-party evaluation by 1 to 2 points
No scores for the new version have been released yet. For the publicly announced Carbon, which serves as a baseline, third-party evaluation company Artificial Analysis announced on Oct. 28 that it ranked first with 59 points on its overall index (out of 100) that combines major evaluations. Anthropic's Claude Opus 5.5 scored 58 and OpenAI's GPT-6 Astra scored 57, a gap of only 1 to 2 points. Carbon was measured at the "high" setting for deeper reasoning.
Wrong-answer rate rose from Argon's 4% to 11%
The same evaluation also showed weaknesses. The rate of giving incorrect content as plausible answers (the hallucination rate) rose to 11% from 4% for the previous Argon. That is the second-lowest level among the models evaluated after Argon, but whether the rate falls in the new version is a key question.
- Gemini 4 Argon
- Gemini 4 Carbon
- Gemini 44 → 11
Timing of public release still undecided
As of Nov. 3, Carbon was still offered only to cyber defense organizations vetted by Google. Availability to developers, businesses and general users is promised as soon as possible, with no date given. Developer API pricing is $2 for input and $10 for output per million tokens during the introductory period, and $4 for input and $20 for output after that. Google has not explained whether this test is part of preparations for broader availability.
Anonymous testing of Carbon on Arena was reported from mid-October, leading up to the official announcement on Oct. 28.
- The Information reports Carbon arriving this monthReport
- Post says a model believed to be Carbon has been on Arena since the previous dayLeak
- Post says Carbon disappeared from ArenaLeak
- Google makes official announcement, early access for cyber defenseOfficial
- Post says Argon checkpoint on Arena was replacedLeak
- Post says "Carbon-H" is appearing under the name "Argon 4"Leak
