A widely shared post claims Google is about to release Gemini 4 Krypton, the top of a four-tier lineup “built along the periodic table”: Hydrogen, Carbon, Argon and Krypton. We checked the claims against Google’s official announcements, media reports and posts by early testers. Krypton does exist as a test checkpoint, but most of the lineup story does not match the facts.

The text has been circulating in AI-focused Telegram channels, including cgevent. It says Gemini 4 models are ranked by the “weight” of the element: Hydrogen is the fastest and cheapest, Carbon a balanced all-rounder, Argon a deep-reasoning model, and Krypton a heavyweight flagship built to take on Anthropic’s Claude Fable 5.2 and OpenAI’s GPT-6.2 Sol. All four are supposedly tied together by a dynamic router that sends each request to the right model.

Confirmed: Argon and Carbon are already out

Gemini 4 Argon is not a mid-tier model. It is the first and so far the only generally available Gemini 4 model. Google announced it on 30 September, initially for cyber defenders in its Fairwind programme. Since 13 October developers can use it through the Gemini API at $2 per million input tokens and $10 per million output tokens, The New Stack reports. Google calls it a flagship model for software engineering, maths, science and agent workflows.

Gemini 4 Carbon is not a “balanced all-rounder below Argon” either. Google introduced it on 28 October as its strongest model for cybersecurity and long-running tasks, built on Argon. Access is still limited to selected cyber defenders. The independent Artificial Analysis index ranks Carbon first with 59 points, ahead of Argon.

What we know about Krypton and Hydrogen

Google has not announced Krypton. On 9 November a tester known as chetaslua posted on X a side-by-side run of Krypton, Carbon and Argon on the same prompt and called it “the first checkpoint of krypton-a”. He said Krypton took about twice as long and wrote about twice as much code. Another model tracker, Kol Tregaskes, wrote that Krypton “is real” and sounds like a big model. It later surfaced in blind-test arenas.

Opinions differ: some testers call Krypton the strongest of the three, others say the gain over Carbon is small. No benchmark numbers have been published. Release dates in circulation (25 November and 8 December) come from Reddit posts, and prices are users’ guesses, according to an overview by Kie.ai.

Hydrogen is the reported codename of a fast, low-cost Flash-tier model. X users first mentioned it in late October, citing several test checkpoints. Google has not confirmed it, and it is not listed in the Gemini API documentation.

What the viral lineup gets wrong

  • “The heavier the element, the more powerful the model” does not hold. Carbon (atomic number 6) is lighter than argon (18), yet Carbon came later and, by Google’s own figures, outperforms Argon.
  • The family is not a closed set of four. Internal documents seen by Business Insider also list a model called Barium; its Barium-B version is what shipped as Argon. Google’s internal codenames usually differ from public names.
  • The model roles and the “inference-compute scaling principle” have no source. Google has announced nothing of the kind, and the roles given to Argon and Carbon contradict the official ones.
  • A dynamic router across all four models is not confirmed anywhere.
  • One of the “direct rivals” does not exist yet. OpenAI has not released GPT-6.2 Sol; it is known only from rumours, and OpenAI’s documentation still lists GPT-6.1 Sol as the latest Sol model.

How to read leaks like this

Stories about new AI models often grow out of a few posts on X: an arena screenshot, one tester’s impressions, a codename. Someone then assembles them into a tidy “company plan” with roles, prices and dates — and the tidiness is what makes it convincing. Check for an official announcement, a listing in the API documentation and independent tests. Until those appear, Gemini 4 Krypton is a promising test checkpoint, not a flagship with known specifications.

Photo: Google headquarters in Mountain View, 2009. Håkan Dahlström / Wikimedia Commons, CC BY 2.0

Автор: Олексій Сердюков

Поділіться своєю думкою

Your email address will not be published. Required fields are marked *