FULL STORY

Gemini 4 Carbon: From Leak to Launch

After leaks of internal Gemini 4 checkpoints, Google officially launched Gemini 4 Carbon on Nov 11, touting long-horizon tasks. Early benchmarks showed strong coding gains as it rolled out across AI Studio and Search's AI Mode.

2026-11-09 ~ 2026-11-11 · 8 episodes · 118 posts

Episode 1 · Leaked Google Gemini 4 Krypton checkpoint reportedly beats unreleased Carbon (2026-11-09, 9 posts)

Starting November 9, multiple X users and Reddit posters claimed Google is internally testing a Gemini 4 checkpoint codenamed Krypton, reportedly outperforming the unreleased Carbon and, by some accounts, all public models. The next day, screenshots from early testers in Google AI Studio further exposed the checkpoint. None of this has been officially confirmed.

Confirmed

  • @lyraxana and @YashasGunderia both reported an internal checkpoint named Krypton with impressive results beating Carbon, without providing benchmarks or screenshots.
  • @raskingballs relayed leaker Lyra's claim that Google has generated a new internal Gemini 4 Krypton checkpoint, a naming pattern typically seen in pre-release internal testing.
  • @koltregaskes claimed to have tested the first checkpoint (krypton-a), saying it took twice as long and produced twice as much code on the same prompt, calling it Google's "hardest-working" model.
  • @gaganghotra said Carbon has not shipped publicly in AI Studio or the Gemini App, yet internal staff are reportedly already using its next iteration, Krypton.
  • @imjustnewatai traced Google's catch-up narrative: from the mediocre early Bard, to Argon about a month ago (reportedly below "6 Astra" and "Opus 5.5"), then Carbon a week later, and now Krypton allegedly surpassing all public models.
  • Reddit users (relayed by @ThePrimeNumbers) said Gemini 4 Krypton is in progress, linking @andrewcurran's claim that several Google engineers confirmed a full Gemini 4 model family is being built.
  • @kimmonismus relayed blogger chatter that Krypton's output quality is "stunning," speculating Google may be approaching or already at recursive self-improvement (RSI); the author could not verify this.
  • On November 10, @APPSO reported that early testers' AI Studio screenshots showed Krypton clearly beating the two-week-old Carbon in code generation, vision, and 3D scenes, with user @Chetaslua demonstrating tests via a single prompt.

Unconfirmed

  • Krypton's existence, version details, and claims of beating Carbon or all public models have no official confirmation; the AI Studio screenshots are third-hand, and Google has not responded. The "engineers confirmed" and "near-RSI" claims remain anonymous leaks or personal speculation.

Why it matters

  • If true, Google's internal iteration has already moved a generation ahead while Carbon remains unreleased, possibly building a full Gemini 4 family. The AI Studio screenshots add credibility, suggesting Google may be rapidly closing in on or overtaking frontier models from OpenAI and Anthropic—worth continued tracking.

Episode 2 · Gemini Reportedly Accelerating to Monthly Releases with Major UI Gains (2026-11-09, 2 posts)

Google's Gemini appears to be shipping major checkpoints roughly monthly, from Krypton to Carbon and Argon, with leaks claiming Gemini 4's Carbon/Krypton versions show front-end and UI capabilities far beyond Argon.

Episode 3 · Gemini 4 Carbon Leaks Across Google Products: Three Tiers, Pricing and Ultra Paywall Surface (2026-11-10, 27 posts)

Starting November 10, Google's next-generation model "Gemini 4 Carbon" has been leaked in multiple places, but Google has issued no confirmation. The leaks began with testingcatalog, which found three Gemini 4 Carbon variants — Low, Medium, High — in the latest Google Antigravity build, along with an entry resembling "gemini-4-carbonccpa".

Confirmed

  • testingcatalog discovered Low, Medium, and High variants of Gemini 4 Carbon plus related entries in an Antigravity build — the closest thing to a primary source in this cluster
  • Multiple Reddit users (ThenFlounder6906, Clear-Complaint-322, StrangeWitness2449, DrCataclysm) posted highly consistent screenshots of the Gemini app's model picker: Quick answers maps to 3.5 Flash-Lite, the general assistant to 3.8 Flash, and a new "4 Carbon" entry appears above the Code & analysis tier — some with a lock icon, marked unavailable or labeled New
  • Some screenshots also showed unreleased models like Gemini 4 Argon (Medium/High, New) and 3.8 Flash (High), and even a Claude Opus 5.5 entry
  • A screenshot relayed by SRRSMITH from a Gemini Pro user showed a new locked "Most complex work: 4 Carbon" entry in the mode list

Not Yet Confirmed

  • markk declared "Gemini 4 Carbon is here" citing what was claimed to be a live demo and tagged @GoogleDeepMind, but received no official response
  • CodeByPoonam claimed to have generated a portrait of Sundar Pichai using "Gemini 4 Carbon" with an attached image; the post had no further detail and its authenticity is unverified
  • A GCI (Generic Capability Index) tracking chart posted by scaling01 showed Gemini 4 Carbon scoring around 134, above Gemini 4 Argon (around 130+); the leaderboard's source and credibility are unclear
  • A Reddit leak relayed by muzaffer22 claimed Gemini 4 Carbon would launch that same day, backed only by a screenshot of a Google engineer Logan Kilpatrick tweet containing the single word "Gemini" — no official source, and no launch followed

Why It Matters

If the leaks hold, Carbon would be Google's more powerful next-gen model following Gemini 4 Argon, offered in Low/Medium/High tiers — hinting Google may shift toward multi-tier, task-complexity-based model selection. Multiple independent screenshots corroborate each other, boosting the leaks' credibility, but until an official release, all model names, capabilities, and timing should be treated as rumor.

7 more related posts →

Episode 4 · Google Launches Gemini 4 Carbon, Its Strongest Model for Long-Horizon Tasks (2026-11-11, 57 posts)

On November 11, Google CEO Sundar Pichai and the official Google AI account simultaneously announced the release of Gemini 4 Carbon, a new frontier model rolling out the same day to Google AI Ultra subscribers worldwide across desktop, mobile, and the web-based Gemini App. Google positions it as the most capable Gemini model to date, purpose-built for sustained long-horizon tasks, with deep reasoning across complex knowledge work, software engineering and web development, cybersecurity, and scientific reasoning, and frontier-level performance in agentic workflows. Current status: the model is officially available, but initially only to Ultra subscribers and API developers; Pro-tier subscribers do not yet have access.

Confirmed

  • Announcement channels: Sundar Pichai, GoogleAI, and GeminiApp official accounts all posted the announcement with links to the official blog, with consistent content (m1, m3, m4, m16).
  • Availability: accessible to AI Ultra subscribers in the Gemini App; developers can use it via the Gemini API (AI Studio) and Antigravity (m2, m5, m7, m8, m12).
  • Performance and capabilities: developer relations engineer patloeber called it a major leap over the previous-generation Argon, with identical API features and unchanged pricing (m7); Philipp Schmid reported bug recall improved by more than 25% over Argon (m8, m12); koraykv (official Google post) said the model tops multiple independent benchmarks, including Cursor-related ones, and internal teams have already used it to assist in mathematically proving Signal's core cryptographic protocols (m9).
  • Industry reaction: Hugging Face's Omar Sanseviero (now on the Google team) assessed it as clearly stronger than Argon and a significant upgrade (m10, m11).

Unconfirmed

  • Capability details mentioned by testingcatalog and full benchmark data await the official blog and subsequent reviews (m6).
  • No official timeline for when Pro-tier subscribers will get access (m6).

Why it matters

  • Gemini 4 Carbon arrived only weeks after the previous-generation Argon went live (m13), a markedly accelerated release cadence reflecting Google's aggressive strategy in the frontier-model race.
  • Its focus on long-horizon agentic workflows and cybersecurity directly targets high-value scenarios like complex software engineering and scientific discovery.
  • API pricing is on par with Argon while performance has improved substantially (25%+ better bug recall), offering developers significantly better value.

37 more related posts →

Episode 5 · Google launches Gemini 4 Carbon; early tests show big coding gains over Argon (2026-11-11, 16 posts)

On November 11, Google released the new model Gemini 4 Carbon, and several bloggers with early access along with the Every team promptly published day-one tests. The overall verdict: coding ability substantially surpasses the previous Gemini 4 Argon, but writing and reliability still lag.

Confirmed

  • Every team (danshipper) day-one test: Carbon scored 44% higher than Argon on their Senior Engineer benchmark; a major coding upgrade—more solid code, more thorough tests and documentation, better 3D object generation, and the ability to independently push large projects for hours at a time with little supervision.
  • Blogger kimmonismus tested 13 coding briefs (3D, websites, apps) in Antigravity and called the results stunning. Highlights: the TIDEFORGE brief demanded 1 million particles and it actually ran 1,048,576, with WebGPU compute driving motion calculations, switching flow modes altering the same real-time simulation, and zero follow-up needed. He also noted Carbon self-tests and fixes its output before submitting, with tight integration—but cautioned its green-light (self-test pass) results can't be fully trusted.
  • In another Every test with 9 real tasks under a 10-minute limit: low thinking mode averaged 81, mid 62, high 54—high mode hit the time limit 5 times and scored 0 three times; with the time limit removed, the gap nearly vanished (81/76/80), suggesting low thinking mode is more practical under real deadline pressure.
  • YouTube channel Prompt Engineering (author disclosed recently joining Google with DeepMind early access) reported an Artificial Analysis intelligence index score of 60 plus a million-token output test.
  • Rohit Krishnan (relayed via zacharynado) considers Carbon a clearly frontier-grade model, notably better than Argon at writing and ideation, but still prone to occasional hallucinations, overconfident answers, and verbose output—human oversight required.

Unconfirmed

  • m9 is a repost, and the original poster expressed reservations (flagged in the post) about whether the model name is actually Carbon, so its experience report is for reference only.

Why it matters

  • Carbon marks a key upgrade to Google's coding model lineup, with multiple independent day-one tests consistently pointing to a coding leap that directly affects comparisons against competitors; findings like "low thinking mode beating high mode" and "don't fully trust the green-light self-tests" also signal that users still need human oversight on real timelines and acceptance checks.

Episode 6 · Gemini 4 Carbon Crafts 3D Pop-Up Book Site and Fictional Instrument Page (2026-11-11, 3 posts)

Midway through the series, Gemini 4 Carbon built a true 3D pop-up book website in three follow-ups, generated a fictional optical instrument product page with 10 detachable, reassemblable parts, and fixed a portfolio site's sculpture and typography with a single prompt.

Episode 7 · Carbon Test Series Wraps: Model Builds and Self-Fixes FPS Game and Poster Editor (2026-11-11, 2 posts)

In the final parts of his Gemini 4 Carbon series, kimmonismus had the model build a tactical FPS with AI defenders and original 3D weapons, plus a layer-based poster editor. Carbon tested and fixed its own work, with one follow-up prompt repairing the editor's Undo and export issues.

Episode 8 · Google brings Gemini 4 Carbon to Search AI Mode for Ultra subscribers (2026-11-11, 2 posts)

Google has started rolling out Gemini 4 Carbon in Search's AI Mode for AI Ultra subscribers, touting deeper reasoning, better instruction following, and more in-depth back-and-forth conversations.