Google's Gemini model family has progressed through several generations: the original Gemini 1.x series, the Gemini 2.x line, the Gemini 2.5 update, and most recently the Gemini 3.x series. The Gemini 3.5 Pro was teased at Google I/O in May 2026 as "coming next month" but never materialized; by July 2026, Bloomberg reported it was months behind schedule [5]. In parallel, Google announced that pre-training for a successor — Gemini 4 — had begun [1].
DeepMind underwent leadership changes during this period. Demis Hassabis stepped down, Noam Shazeer left for OpenAI, and Sergey Brin reportedly returned to the office to work on coding models [6]. Koray Kavukcuoglu, the new DeepMind chief, publicly addressed Gemini 4's progress in September 2026 [2].
The research landscape for this report is shaped by two categories of information: official Google communications and community-sourced leaks. Neither category provides a complete picture, and the report treats each accordingly.
2. Official Google Confirmations
Google has made the following statements about Gemini 4, all of which are treated as confirmed:
- July 21, 2026: The Google DeepMind blog stated that pre-training for Gemini 4 had started, describing it as "our most ambitious pre-training run yet" [1].
- July 22, 2026: Sundar Pichai reiterated this on the Alphabet Q2 earnings call [1].
- September 23, 2026: Koray Kavukcuoglu spoke at The Information's AI Agenda Live Summit, confirming that Gemini 4 had entered "early post-training" and expressing a desire for it to be released "much earlier" than the end of 2026 [2].
- September 24, 2026: 9to5Google reported that Kavukcuoglu added that internal testing was underway and that Gemini 4 was already powering a project called Antigravity [1].
What Google has not confirmed includes: the release date, model variants, model ID, API endpoint, documentation, benchmark scores, pricing, context window size, and the status of Gemini 3.5 Pro [1].
3. Leaked Information
3.1 The Argon Checkpoint (September 14, 2026)
Tech researcher Lentils (@Lentils80) posted on X that Gemini 4 Pro checkpoints had "finally started appearing internally." The checkpoint was codenamed "Argon." Lentils reported a 256k output token limit, 2.4 minutes of compute on "High thinking effort," and a 2M token context window — though the context window was described as "still not decided" [3]. These details are treated as plausible but unverified; they come from a single source with no corroborating documentation.
3.2 The "gemini-3.8-flash" Anomaly (September 17, 2026)
On September 17, 2026, a model labeled "gemini-3.8-flash" appeared on LMSYS Chatbot Arena and produced outputs far beyond what Flash-tier models typically deliver. Community members inferred this was a Gemini 4 Pro checkpoint. Notable demonstrations included a PS5 SVG render (~10 minutes generation), a BMW M4 model, a voxel pagoda, and a pelican riding a bicycle. The 256k output token ceiling was reportedly hit [3]. This is one of the more credible community observations because it is based on direct interaction with a live model on a public platform, though the model's identity has not been officially confirmed.
3.3 The Qwinah Claims (September 17, 2026)
Developer Qwinah (@MaaSonder) claimed to have "ghost routed" into the model's backend, reporting the internal name G4P-ARGON / VIA_3.8_FLASH, a 10M+ token context window, 256k output, permanent cross-session memory, offline web access, script compilation and execution, and physical robot control [3]. These claims are treated with skepticism by the research community. The 10M context window contradicts the 2M figure reported by Lentils, and the more extraordinary claims (robot control, script execution) have no independent verification. The credibility of these claims is rated very low for the more speculative items and low for the 10M context figure.
3.4 The Barium-b Checkpoint (September 26, 2026)
A new checkpoint labeled "barium-b" was reported on LMSYS Chatbot Arena on September 26, 2026 [4]. No detailed capability descriptions or specifications were published alongside this report.
3.5 The Floatplane Demo (September 26, 2026)
Viral X clips showed a 3D floatplane takeoff demo comparing Gemini 4 Pro against Claude Opus 5.5. The Gemini 4 Pro output reportedly showed superior acceleration, water reflections, and flight stability. The clips accumulated approximately 170K–220K views [4]. These are community demos, not controlled benchmark evaluations, and the comparison conditions (prompt, settings, generation time) are not documented.
3.6 The Benchmark Leak (August 12, 2026)
On August 12, 2026, a post on X titled "Gemini 4 Pro Benchmark Leak" claimed wins over GPT-5.6 Sol and Claude Fable 5, with a 1.5M token context window. No evaluation files, methodology, or source data were provided. Secondary outlets repeated the claims on August 19, 2026 [4]. The absence of source data makes these claims unverifiable and they should be treated as unsubstantiated.
3.7 The "Predicted" Benchmark Sheet (September 4, 2026)
A "predicted" benchmark sheet circulated on September 4, 2026. The Gemini 4 column was labeled "(PREDICTED)" while other columns were copied verbatim from OpenAI's Astra launch page [4]. This is a clear case of misinformation — the sheet presented speculative Gemini 4 figures alongside real OpenAI data in a format designed to appear authoritative.
3.8 The Desktop App Artifact (August 24, 2026)
A Gemini desktop app update on August 24, 2026, contained internal references to Google "preparing internally for Gemini 4." This is the only hard artifact of a planned launch found in the research so far [4]. It does not provide any specification details but confirms that internal preparation was underway.
4. Credibility Assessment
Each major claim is assigned a credibility rating based on the number of independent sources, the presence of corroborating evidence, and the plausibility of the claim.
| Claim | Credibility | Basis |
|---|---|---|
| Pre-training started July 21, 2026 | Confirmed | Official Google DeepMind blog [1] |
| Post-training as of Sept 23, 2026 | Confirmed | Kavukcuoglu at The Information summit [2] |
| Argon checkpoint exists | Probable | Single researcher report, corroborated by Arena observations [3] |
| 256k output token limit | Plausible | Consistent across multiple independent observations |
| 10M context window | Low | Single claim (Qwinah), contradicted by the 2M figure from Lentils [3] |
| Beats GPT-6 Astra on benchmarks | Unverified | No source data, no evaluation files [4] |
| October 2026 release | Speculative | Polymarket pricing (68% by Oct 31), analyst estimates [1] |
| Gemini 3.5 Pro canceled | Unconfirmed | Multiple reports say canceled/shelved; Google says "testing with partners" |
| Physical robot control | Very low | Single unverified claim from Qwinah [3] |
| Floatplane demo authenticity | Moderate | Viral clips with high view counts, but no controlled comparison [4] |
| "gemini-3.8-flash" is Gemini 4 Pro | Probable | Community inference from capability mismatch, no official confirmation [3] |
5. Rumors and Speculation
The following items circulate in community discussions but lack confirmation from any official or high-credibility source:
- Release window: October 2026 is the most commonly cited window, with Polymarket markets assigning a 68% chance of release by October 31, 2026. Analyst estimates range from November to December 2026 [1].
- Model specifications (all unverified): 1.5M–10M token context window, 256k output token limit, approximately $2.25 per 1M input tokens and $11.25 per 1M output tokens, TPU v6 architecture, and a "High-effort" reasoning mode [1][3].
- Gemini 3.5 Pro status: Multiple reports indicate the model has been canceled or shelved, but Google's last public statement describes it as "testing with partners," leaving its status ambiguous [1][5].
- DeepMind leadership: The departure of Hassabis and Shazeer, combined with Brin's reported return, has fueled speculation about the direction of Gemini 4's development, particularly around coding models [6].
6. Timeline of Events
| Date | Event | Type |
|---|---|---|
| May 19, 2026 | Google I/O: Gemini 3.5 Pro announced as "coming next month" | Official |
| July 1, 2026 | Anonymous Gemini Flash checkpoint appears on LMArena | Community |
| July 16, 2026 | Bloomberg reports Gemini 3.5 Pro is months behind schedule | Reported |
| July 21, 2026 | Google DeepMind blog confirms Gemini 4 pre-training has started | Official |
| July 22, 2026 | Sundar Pichai: "most ambitious pre-training run yet" on Alphabet earnings call | Official |
| August 12, 2026 | "Gemini 4 Pro Benchmark Leak" post on X | Unverified leak |
| August 24, 2026 | Gemini desktop app update reveals internal Gemini 4 preparation | Artifact |
| September 2, 2026 | Gemini 3.8 Flash ships; no Gemini 4 mention | Official |
| September 4, 2026 | "Predicted" benchmark sheet circulates (misinformation) | Misinformation |
| September 14, 2026 | First "Argon" checkpoint leaked by Lentils | Leak |
| September 17, 2026 | "gemini-3.8-flash" on LMArena inferred as Gemini 4 Pro; Qwinah ghost-routing claims | Community inference / Unverified claim |
| September 23, 2026 | Kavukcuoglu confirms Gemini 4 in early post-training at The Information summit | Official (reported) |
| September 24, 2026 | 9to5Google reports Gemini 4 powering Antigravity in internal testing | Official (reported) |
| September 26, 2026 | Floatplane demo, "barium-b" checkpoint, viral X clips | Leak / Community |
7. Hugging Face, GitHub, and Model Registry Status
No Gemini 4 model weights, model cards, or repositories were found on Hugging Face, GitHub, or any public model registry. This is consistent with the model still being in the training and post-training phase and not yet released for public download or API access [4].
8. Contradictions and Open Questions
Several contradictions emerge from the available evidence:
- Naming inconsistency: The Gemini family jumped from 1.x to 2.x to 2.5 to 3.x without a "Gemini 4" generation until now. The appearance of "gemini-3.8-flash" on LMArena further complicates the naming scheme — it is unclear whether the 3.8 designation is a product name, an internal version, or a community label [4].
- Checkpoint naming vs. product naming: "Argon" and "barium-b" are checkpoint codenames, not official model names. It is unknown whether the final product will use these names, a variant of them, or an entirely different designation [3][4].
- Conflicting timeline signals: Kavukcuoglu stated a desire for release "much earlier" than end of 2026, while analyst estimates point to November–December 2026. The Polymarket market at 68% for October 31 suggests a roughly even chance of an October release [1].
- Gemini 3.5 Pro status: The model was teased at Google I/O, reported as delayed by Bloomberg, and then described as "testing with partners" by Google. Whether it will ship as a standalone product, be absorbed into Gemini 4, or be canceled entirely remains unknown [1][5].
- Context window discrepancy: The 2M token figure from Lentils and the 10M+ token figure from Qwinah cannot both be correct. The 2M figure is more consistent with the broader pattern of reported specs and is treated as more plausible [3].
9. Missing Evidence
- No benchmark scores of any kind (no published, no verified leaked)
- No API model ID or endpoint documentation
- No pricing information
- No release date
- No context window specification (conflicting reports exist)
- No parameter count or architecture details beyond "TPU v6" in rumors
- No multimodal capability confirmation from official sources
- No official model card or technical report
- No evaluation files or methodology for the claimed benchmark wins
- No confirmation of the Gemini 3.5 Pro cancellation
10. Conclusion
Gemini 4 is a confirmed Google DeepMind project that has progressed from pre-training (July 2026) to early post-training (September 2026). Internal checkpoints have surfaced on community platforms and show capabilities that exceed the current Gemini 3.x lineup, particularly in areas such as output length, 3D generation, and interactive UI creation. However, the model is not publicly available, no benchmarks have been published, and many of the leaked specifications rest on single, unverified sources. The most reliable signals — the confirmed pre-training and post-training phases, the existence of the Argon checkpoint, and the 256k output limit — are consistent across multiple independent observations. The more extraordinary claims (10M context, robot control, benchmark wins over GPT-5.6) should be treated as speculative until corroborated. The release window remains uncertain, with October 2026 being the most commonly cited but far from confirmed.
Sources
[1] Google DeepMind Blog, "We have started our most ambitious pre-training run yet, for Gemini 4," July 21, 2026; Sundar Pichai, Alphabet Q2 Earnings Call, July 22, 2026; Koray Kavukcuoglu, The Information AI Agenda Live Summit, September 23, 2026; 9to5Google, September 24, 2026.
[2] Koray Kavukcuoglu, The Information's AI Agenda Live Summit, September 23, 2026.
[3] Lentils (@Lentils80), X post, September 14, 2026; Qwinah (@MaaSonder), X post, September 17, 2026; LMSYS Chatbot Arena, "gemini-3.8-flash" model page, September 17, 2026.
[4] Researcher 2 findings: APIMaster.AI, LMSYS LMArena leaderboard (September 2026), TokenCalculator, X posts (August 12, September 4, September 26, 2026), Gemini desktop app update (August 24, 2026).
[5] Bloomberg, "Gemini 3.5 Pro months behind schedule," July 16, 2026.
[6] Multiple sources on DeepMind leadership changes: Demis Hassabis departure, Noam Shazeer departure to OpenAI, Sergey Brin reported return to office.