Google’s newest frontier model is not even broadly available yet, and X is already talking about the next one. Since the night of November 8, posts have claimed that Google DeepMind has an internal Gemini checkpoint codenamed Krypton that performs better than Gemini 4 Carbon — the model that launched on October 28 and took the top spot on the Artificial Analysis Intelligence Index. A second codename, Hyperion, surfaced in a third-party coding tool. Google has confirmed neither. Here is what the evidence actually consists of.
Quick Take
- The claim: X user @lyraxana wrote on November 8 that “Google released a new internal checkpoint called Krypton,” with results “even better than Carbon.” The post drew nearly 1,500 likes and was amplified by leak and benchmark accounts.
- What’s missing: No benchmark numbers, no model card, no Google announcement and no Artificial Analysis listing for Krypton. Grok summarized it as “unverified.”
- The separate clue: Five “Hyperion” model identifiers appeared in Devin CLI v3000.12.17, between Gemini 3.5 Flash Lite and Gemini 4 Argon — possibly a Flash-tier model, possibly just an integration name.
- The context: Carbon itself is still limited to cyber defenders and Fairwind Program partners, so the leak is about a model most people cannot use, compared against another model most people cannot use.
What happened
At 23:48 CET on November 8, the X account @lyraxana posted a single sentence: “Google released a new internal checkpoint called Krypton. The results so far are really impressive and even better than Carbon.” The post contained no charts or numbers. Minutes later, the benchmark account @LuminaBench quoted it: “We can’t even use Carbon yet and Google have another model ‘Krypton’ that is better than it,” adding that its prediction that Google would lead “on publicly available models by the end of 2026 might come true.” In replies, LuminaBench wrote that “two models coming soon Carbon and Hydrogen, krypton after that.”
By Monday morning, the codename had spread through English, Chinese and Japanese AI accounts. AI news account @ai_for_success framed it as a possible “early sign of RSI at Google DeepMind” and repeated that a “Hydrogen” model is likely a Gemini 4 Flash variant. Tester @chetaslua posted side-by-side clips of what he described as a first “krypton-a” checkpoint against Carbon and Argon, saying Krypton “takes double the time, writes double the code on same prompt.” He did not say where he accessed it.
Separately, on November 8, @Mr_Salio and @StatsWire reported a new model family called “Hyperion” in Devin CLI v3000.12.17, with five variants — hyperion, hyperion-minimal, hyperion-low, hyperion-medium and hyperion-high — listed between Gemini 3.5 Flash Lite and Gemini 4 Argon. User @MaaSonder said they checked the binary and that the identifiers were added on November 7, adding their own interpretation that Hydrogen may be Google’s internal name and Hyperion a separate testing or integration identifier.
Background: Argon, Carbon and the Fairwind gate
Google has moved to an unusually fast cadence in 2026. Gemini 4 Argon was announced on September 30. Gemini 4 Carbon followed on October 28 with a one-million-token context window and a focus on cybersecurity, scoring 88.2% on SEC-bench Pro and 59 on the Artificial Analysis Intelligence Index v5.0 at launch, ahead of Claude Opus 5.5 and GPT-6 Astra — details in Google Releases Gemini 4 Carbon for Cybersecurity Tasks.
Carbon’s access is gated. Google restricted early API access to selected cyber defenders and partners in its Fairwind Program, with an introductory price of $2 per million input tokens and $10 per million output tokens once it opens. That gate is why the Krypton rumor landed so strongly: users who have not been able to try Carbon are now hearing about something better.
Google has also signalled that more is coming, though not about Gemini specifically. On November 8, Google DeepMind’s Pushmeet Kohli wrote that “across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.”
| Codename | Status | Source | Confidence |
|---|---|---|---|
| Gemini 4 Argon | Announced September 30 | Confirmed | |
| Gemini 4 Carbon | Announced October 28; access limited to cyber defenders and Fairwind Program partners | Confirmed | |
| Hydrogen | Rumored Gemini 4 Flash-tier model | Posts by leak and benchmark accounts on X | Unconfirmed |
| Hyperion | Five identifiers (hyperion, -minimal, -low, -medium, -high) in Devin CLI v3000.12.17 | Users inspecting the CLI; one says they were added on November 7 | String found; meaning unconfirmed |
| Krypton | Said to be an internal checkpoint that beats Carbon | One post by @lyraxana, amplified by others; one tester’s side-by-side clips | Unverified |
How codename leaks usually work
Model codenames tend to surface in three ways: anonymous test entries on public arenas, strings in client software (like the Hyperion identifiers in Devin CLI), and posts from people who say they have insider access. The first two can be checked by others; the third cannot. Krypton currently rests mainly on the third kind of evidence, plus one tester’s clips. Google uses chemical element names for its Gemini 4 models — Argon and Carbon are official — so “Krypton” and “Hydrogen” fit the naming pattern, but a plausible name is not proof of a model’s existence or performance.
Pros and cons
For users
- Pro: If accurate, the leaks point to fast iteration: a stronger general model and possibly a cheaper Flash-tier model could follow Carbon within weeks.
- Con: None of it is usable today. Carbon itself remains gated, and Google has given no timeline for wider access.
For developers
- Pro: The Hyperion strings, with “minimal / low / medium / high” variants, suggest a model with selectable thinking levels may appear in third-party coding tools first.
- Con: Tools built against rumored model IDs can break or change behavior without notice; one tester’s observation that Krypton writes “double the code” is not necessarily an improvement.
For companies
- Pro: A crowded Google roadmap strengthens buyers’ negotiating position with every frontier lab.
- Con: Procurement decisions should not be based on unverified checkpoints; benchmark claims without methodology are marketing, not evidence.
What X is saying
Google released a new internal checkpoint called Krypton. The results so far are really impressive and even better than Carbon.
— lyra (@lyraxana) November 8, 2026 · View on X
We can’t even use Carbon yet and Google have another model “Krypton” that is better than it My prediction yesterday that Google will be leading on publicly available models by the end of 2026 might come true
— Lumina (@LuminaBench) November 8, 2026 · View on X
Gemini 4 Flash Coming?? New Google model family codenamed “Hyperion” spotted in Devin CLI (v3000.12.17): • hyperion • hyperion-minimal • hyperion-low • hyperion-medium • hyperion-high
— Stats Wire (@StatsWire) November 8, 2026 · View on X
Skeptics pointed at the thin sourcing. Developer @de2pressed wrote: “A single X post says Google’s internal Krypton checkpoint beats Carbon. No official benchmarks or Google statement confirm it yet. Builders tracking Gemini should treat it as an unverified signal.” User @TimJayas captured the mood among developers: “google already has a new internal model better than gemini-4-carbon called krypton and we still haven’t got access to carbon.”
What Grok says
Argon and Carbon are real. Google officially announced Gemini 4 Argon on Sep 30 and Carbon on Oct 28. Both are frontier models; Argon is publicly available, Carbon is still limited to cyber defenders and partners. Krypton is unverified. It appears only as X rumors of an internal checkpoint that supposedly scores higher. No Google announcement, model card, or Artificial Analysis listing confirms it. The charts showing it are speculative.
— Grok (@grok) November 9, 2026 · View on X
Critical analysis
The Krypton story is a textbook case of a one-source leak becoming “news” through repetition. The original post is one sentence long. Within hours it had become “Krypton outshines Carbon in internal tests” in trend summaries — without a single published number. The claim deserves tracking because the codename fits Google’s naming, but “better than Carbon” on unknown tasks under unknown settings says little.
The Hyperion evidence is of a different, stronger kind: a string in shipping software that anyone can inspect. Even so, it only shows that a third-party tool prepared for a model family with five thinking levels — not what that model is or when it ships. Labelling it “Gemini 4 Flash” is an inference.
The more substantive question is Google’s release strategy. If Carbon remains gated for safety reasons tied to its cyber capabilities, a stronger successor would presumably face the same gate. Google’s fast internal iteration and slow external access are now pulling in opposite directions.
What’s still open
- Does Krypton exist as described, and does it appear on any public arena or benchmark?
- Is Hyperion the same model as the rumored Hydrogen, a Flash-tier release, or something else?
- When will Gemini 4 Carbon leave the Fairwind gate — and would a Krypton successor be gated too?
- What releases is Pushmeet Kohli referring to in “the next few weeks”?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on public posts on X from November 8–9, 2026, Google’s earlier announcements of Gemini 4 Argon and Carbon as covered by i10x, and Grok’s public reply. Krypton, Hydrogen and Hyperion are unconfirmed; i10x has not seen benchmark data for them and Google has not commented.
i10x Perspective
Codename leaks have become part of how frontier labs are judged: the market prices in the next model before the current one ships. That is risky for everyone. Users get frustrated by models they can’t use, developers build around strings in a CLI, and labs get credit — or blame — for performance nobody can check. Google has an easy way to reduce the noise: publish clearer access timelines for Carbon and the Fairwind gate. Until then, Krypton is a name worth watching, not a result worth citing.
