ModelsArchitectures & capability
Google announces Gemini 4 Carbon for restricted testing, makes Gemini 4 Argon generally available and previews Genie 4
Google announced Gemini 4 Carbon, which Artificial Analysis ranked first on its Intelligence Index v5, but limited it to the U.S. government and selected cyber defenders. Gemini 4 Argon became generally available, and Genie 4 reached trusted testers.
From the Google AI article of November 2, 2026 and cited sources: Google announced Gemini 4 Carbon on 28 October. [1] [3] [4]
SOURCES[1] The latest AI news we announced in October 2026 blog.google[3] Gemini 4 Carbon: our next frontier model (Google) blog.google[4] Gemini 4 Carbon is the new #1 (Artificial Analysis) artificialanalysis.ai
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Google AI article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 2, 2026 article published
- Covered
- Nov 3 edition
- Evidence
- 10 sources across 7 sites
- 1The latest AI news we announced in October 2026blog.google
- 2Linked video: The latest AI news we announced in October…youtube.com
- 3Gemini 4 Carbon: our next frontier model (Google)blog.google
- 4Gemini 4 Carbon is the new #1 (Artificial Analysis)artificialanalysis.ai
- 5Google releases Gemini 4 Argon (TechCrunch)techcrunch.com
- 6Gemini 4 Argon: Built to solve your hardest problems…blog.google
StrongHow it’s measured
YouTube video · 596.6K views · 11.8K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Google announced Gemini 4 Carbon on 28 October. For now it is in pre-release testing through the Fairwind Program, which includes the U.S. government and selected cyber defenders. The same day, Artificial Analysis placed it first on its Intelligence Index v5 with a score of 59. [1] [3] [4]
Carbon does not lead everywhere. On Artificial Analysis's measures, it hallucinates more often than Argon (11% vs 4%). It also trails Claude Opus 5.5 on coding agents (59 vs 62) and Kimi K3.1 Pro on knowledge work (1609 vs 1678 Elo). [4]
Gemini 4 Argon became generally available on 13 October through the paid Gemini API, AI Studio, Antigravity, the Gemini app and AI Mode in Search. It has a 1M-token combined input/output limit. Introductory pricing of $2/$10 per million input/output tokens rises to $4/$20 after 30 November. [1] [6]
Carbon tops one independent composite index but hallucinates more than Argon and trails rivals in some areas. Implication: for now, teams can build only on Argon, and should budget for its price doubling after 30 November.
Executive brief
Google is holding back its strongest model. On 28 October it announced Gemini 4 Carbon, and the same day the independent evaluator Artificial Analysis ranked it #1 on its Intelligence Index v5. Even so, only the U.S. government and selected cyber defenders can use Carbon for now. Its predecessor, Gemini 4 Argon, became widely available on 13 October. Google's October roundup also lists Genie 4, a real-time world model that generates its own audio, plus updates to EmbeddingGemma, SynthID, quantum computing and health.
What changed and event timeline
Argon first goes to cyber partners
TechCrunch reported that Argon launched first to cyber partners in Google's Fairwind Program and was framed as a defensive cyber model. Google's later roundup gives Carbon that role instead.
Argon becomes generally available
Available through the paid Gemini API, AI Studio, Antigravity, the Gemini app and AI Mode in Search, at introductory pricing.
Genie 4 announced
A real-time interactive world model with native audio, controlled by text prompts. Only trusted testers have it so far.
Carbon announced and ranked #1
Carbon is in pre-release testing through Fairwind. Artificial Analysis published an evaluation the same day placing it first on Intelligence Index v5.
Google publishes its October roundup
It describes the Argon-to-Carbon gap as "two weeks" and says Carbon testing includes the U.S. government.
Capabilities and access
- Gemini 4 Argon: a 1M-token combined input/output limit. Fairwind partners get a version without cyber guardrails (Google).
- Gemini 4 Carbon: same pricing as Argon; Artificial Analysis says it is "not yet publicly available" (AA).
- Genie 4: coming to AI Ultra subscribers through Project Genie and to early-access partners in the API and AI Studio (Google).
Read the full section
- Gemini 4 Argon: a 1M-token combined input/output limit; introductory pricing of $2/$10 per million input/output tokens, rising to $4/$20 after 30 November; Fairwind partners get a version without cyber guardrails (Google). A robotics variant also exists (roundup).
- Gemini 4 Carbon: same pricing as Argon; Artificial Analysis says it is "not yet publicly available" (AA).
- Genie 4: coming to AI Ultra subscribers through Project Genie and to early-access partners in the API and AI Studio (Google).
- EmbeddingGemma 2: an open, 740M-parameter multimodal embedding model that runs on-device (roundup).
Technical analysis for researchers and developers
- Genie 4 is built from several Gemini models. One secondary source says a Gemini 3.8 Flash agent watches the scene as it generates (AI Weekly).
- How Carbon was evaluated: Artificial Analysis tested it at high reasoning effort across 9 evaluations.
- Genie 4 has no published benchmarks, so its results cannot be reproduced (DataCamp).
Read the full section
- Genie 4 is built from several Gemini models. Gemini Omni grounds how things look and behave, and other Gemini models handle reasoning to keep the simulation consistent (Google). One secondary source says a Gemini 3.8 Flash agent watches the scene as it generates (AI Weekly). Google's promo video shows turning an image into a world with ambient sound, entering a world described in text, and robots practising tasks.
- How Carbon was evaluated: Artificial Analysis tested it at high reasoning effort across 9 evaluations. Each task cost $7.76 and used about 123k output tokens, against $10.04 for Claude Opus 5.5 (AA).
- Genie 4 has no published benchmarks, so its results cannot be reproduced (DataCamp).
Claims and evidence
Neither the Lancet paper nor the Science paper was checked directly.
Read the full section
| Claim | Evidence |
| Carbon is #1 on AA Intelligence Index v5 (score 59) | Independent: AA |
| Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench, and 0% attack success on Gray Swan prompt injection | Vendor-reported: Google |
| Argon is #1 on the Vals Index and scores 87% on ERQA | Vendor-reported: Google |
| AMIE helped clinicians prepare in 75% of visits and matched their final diagnoses 90% of the time (published in The Lancet) | Roundup summary only: roundup |
| Magic state cultivation cut errors about 40-fold on Willow (published in Science) | Roundup summary only: roundup |
Neither the Lancet paper nor the Science paper was checked directly.
Context and prior work
Genie 4 holds a consistent scene for up to two minutes, twice as long as Genie 3, according to Orca Router. Genie 3 had no audio (DataCamp). Argon arrived in a crowded field: TechCrunch describes recent competing launches from OpenAI (GPT-6 Astra) and Anthropic (Fable) and says the Gemini app reached a billion monthly users in August (TechCrunch).
Limitations, safety and contested findings
- Carbon hallucinates more than Argon: 11% versus 4% on AA's measure (AA).
- Carbon is not first everywhere.
- Genie 4 has known failure modes, including object permanence, text legibility, complex physics and audio accuracy.
Read the full section
- Carbon hallucinates more than Argon: 11% versus 4% on AA's measure (AA).
- Carbon is not first everywhere. It trails Claude Opus 5.5 on coding agents (59 vs 62) and Kimi K3.1 Pro on knowledge work (1609 vs 1678 Elo) (AA).
- Genie 4 has known failure modes, including object permanence, text legibility, complex physics and audio accuracy. DataCamp judges it suited to prototyping, not production (DataCamp).
- Accounts of Argon's launch conflict. TechCrunch says Argon launched to cyber partners first, while Google's roundup says it rolled out globally after "early safety testing."
Business and practitioner implications
- Pricing: Argon's price doubles after 30 November, so budget for $4/$20 per million tokens.
- Carbon access: outside Fairwind, there is no date for general availability.
- Hallucination trade-off: where errors are costly, Carbon's higher hallucination rate may favour Argon until outside testing settles the question.
Read the full section
- Pricing: Argon's price doubles after 30 November, so budget for $4/$20 per million tokens. Heavy caching (95% discount on cached inputs) reduces the cost (Google).
- Carbon access: outside Fairwind, there is no date for general availability.
- Hallucination trade-off: where errors are costly, Carbon's higher hallucination rate may favour Argon until outside testing settles the question.
- On-device and verification tools: EmbeddingGemma 2 (up to 6x less local storage) and the global SynthID Detector rollout support private, offline search and checks for AI-generated content (roundup).
The source trail.
Sources (10)
The latest AI news we announced in October 2026
Article text retrieved; extracted text may omit tables or interactive elements. A transcript of the video linked from this page is supplied (linked video manual captions; https://www.youtube.com/watch?v=sFfPkUJd_Jk); automatic text may contain errors.
blog.googleLinked video: The latest AI news we announced in October 2026
linked video manual captions
www.youtube.com