"Gemini 4 Flash" reportedly in A/B testing on AI Studio and Arena; leaker's posts offer no evidence, Google has not announced it

Reported testing on AI Studio and Arena, with no evidence offered

A report that "Gemini 4 Flash," described as an unannounced Google model, was being A/B tested in the developer environment "Google AI Studio" and on "Arena" (formerly LMArena), a site that compares AI answers anonymously, was posted on X on Nov. 5 (7:21 p.m. Nov. 5 Japan time). According to the leaker @Haleeeemahh, a model referred to by the code name "Hydrogen" corresponds to Gemini 4 Flash. A/B testing refers to a trial that shows a different version to only some users and compares their responses. The post did not include screenshots of the testing or any supporting evidence. ORIENS has not been able to confirm the name, its link to Hydrogen or whether testing is taking place, and Google has not made an announcement.

Image in the first post was a joke

The same claim had been posted about 15 and a half hours earlier, on Nov. 4 (3:48 a.m. Nov. 5 Japan time), by a different account, @lyraxana. The attached image had nothing to do with the model's screen; it was a joke image about a hydrogen bomb.

@chetaslua quoted this post the same day (5:27 a.m. Nov. 5 Japan time), posting that there were five new Hydrogen checkpoints (versions of a model saved partway through training) and that it was a Flash model producing about 300 to 500 tokens per second. The attached screenshot shows five images lined up by subject, each labeled "hydrogen," but the text that follows is blurred. It is a screenshot of files opened on the poster's own machine, and it is impossible to tell whether the images were drawn by a Google model.

This speed is also unsupported. In measurements as of Nov. 5 by U.S. evaluation company Artificial Analysis, the current "Gemini 3.8 Flash" ran at 117 tokens per second and "Gemini 4 Argon" at 74 tokens per second. A blog post by OrcaRouter noted that there were reports in September of more than 300 tokens per second for 3.8 Flash, and said the figure varies widely depending on how it is measured. According to the blog, no one has shown a model ID, pricing or benchmarks.

Caution needed on the poster's reliability

@Haleeeemahh is an account that posts information about unannounced models. On Oct. 31, it posted that a version called "Carbon-H" was appearing on Arena, which U.S. tech outlet TestingCatalog picked up, but Google has not confirmed it. It has also dismissed a rumor about the release of "GPT-6.5," saying the screenshot behind it came from a fake site. On the other hand, it is difficult to judge whether the testing is real based on this post alone.

The name "Hydrogen" also originates from individual posts. @White1637402, an account with about 1,500 followers, posted on Oct. 27 (10:52 a.m. Oct. 27 Japan time) that Gemini 4 Flash had surfaced internally under this code name. There was no image or source. @Haleeeemahh posted similar content the same day. There are multiple posts, but it is not known whether each is based on a separate source. Because Arena does not reveal which version gave an answer, it is also difficult to pin down names from user reports.

Name absent from official model list

Google's model list for developers did not include Gemini 4 Flash or the name Hydrogen as of Nov. 5. The only Gemini 4 model listed is Argon, and the latest Flash is Gemini 3.8 Flash. The changelog, which runs through that day, also has no mention. The list covers released models, so it cannot show whether pre-release testing is taking place.

"Gemini 4 Carbon," the higher-end model mentioned in the second half of @Haleeeemahh's post, was announced by Google on Oct. 28 (Oct. 29 Japan time). Early access is for organizations on the defending side against cyberattacks, and access for developers and general users was promised as soon as possible, with no date given. Carbon is not in the model list either, so this part is consistent with official documentation.

There is no official information indicating when the model will be released. As of Nov. 5, participants on the prediction market Manifold put the probability of a Gemini 4-generation Flash being released before Nov. 16 at 20%, and before Dec. 1 at 40%. This reflects participants' views and does not confirm a release.

Focus on pricing and performance between current Flash and Argon

If a model by that name is released, it will be compared with the current Flash and Argon. API pricing for developers per 1 million tokens is $0.75 for input and $3.75 for output for Gemini 3.8 Flash. Gemini 4 Argon costs $2 for input and $10 for output, rising to $4 for input and $20 for output from Dec. 1. 3.8 Flash will also rise to $1.50 for input and $7.50 for output from Jan. 1, 2027. "GPT-6.1 Terra," which OpenAI positions for everyday tasks at a lower price than its higher-end "GPT-6.1 Sol," costs $1 for input and $5 for output.

On Artificial Analysis' overall index (out of 100), Gemini 3.8 Flash scored 31 and Argon 50 as of Nov. 5. @LuminaBench posted on Oct. 27 that if Hydrogen performed close to Argon at a much lower price, Google could temporarily take second place from OpenAI. This is an individual's view, and no documentation of Hydrogen's performance or pricing has been released.

Rumors since late October

On lightweight models, a rumor also emerged on Oct. 25 that the name "gemini-3.8-flash-router" appeared in internal information in a developer tool. The relationship between this name and Hydrogen is unknown.