The Information Machine

Google Rapidly Iterates Gemini with 3.6 Flash Family and Restricted Cybersecurity Variant · history

Version 4

2026-07-25 08:08 UTC · 54 items

What

Google released Gemini 3.6 Flash and 3.5 Flash-Lite on July 21, 2026, replacing 3.5 Flash just two months after its Google I/O debut, while a cybersecurity-specialized variant (Gemini 3.5 Flash Cyber) remains restricted to governments and trusted partners due to its ability to autonomously find and exploit production vulnerabilities [1][4]. Gemini 3.5 Pro has still not shipped; internal reporting from the LA Times now attributes the delay not only to coding performance shortfalls but also to clashing teams and frustrated engineers [9]. InfoWorld argues Google's CEO used talk of Gemini 4 and a commitment to monthly releases to redirect attention from the Pro delay [10]. Google has disclosed it has begun pre-training Gemini 4 with no timeline attached [1].

Why it matters

The delay of Gemini 3.5 Pro is acquiring an organizational dimension — team conflicts alongside technical shortfalls — that makes a resolution timeline harder to read than a simple benchmark gap. The CEO's framing of Gemini 4 pre-training and monthly releases as a forward-looking story sits in direct tension with a flagship model that has been publicly stuck for over a month. The restricted Flash Cyber model faces the same durability question: access controls hold only until a competitor ships an equivalent without them.

Open questions

  • What specific team conflicts and organizational dynamics does the LA Times reporting describe, and do they suggest the 3.5 Pro delay is primarily a management problem or a technical one? [9]

  • Did Google commit to a monthly release cadence, and if so, does that commitment change how the 3.5 Pro delay is managed internally? [10]

  • Does restricting Flash Cyber to governments and trusted partners provide meaningful protection once a competitor ships a comparable model without access controls? [2][4]

  • When will Gemini 3.5 Pro ship, and what specific coding benchmarks must it clear before Google releases it? [7][12][13]

Narrative

Google released three Gemini models in quick succession in July 2026. Gemini 3.6 Flash, released July 21, replaces 3.5 Flash just two months after its Google I/O debut, claiming 17% fewer output tokens and improvements on benchmarks where 3.5 Flash had underperformed, including DeepSWE coding (49% vs. 37%) and OSWorld computer use (83.0% vs. 78.4%) [1]. Gemini 3.5 Flash-Lite, released the same day, targets high-volume latency-sensitive workloads at 350 output tokens per second and is being integrated into Google Search [1][2]. 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens [2]. Ars Technica reported the unusual pace of deprecating a two-month-old flagship and argued that Google's efficiency-first training priorities may have caused 3.5 Flash's coding shortfall that 3.6 Flash was built to address [3].

Four days before the Flash releases, Google announced Gemini 3.5 Flash Cyber, trained specifically for offensive and defensive security work. In internal evaluations it found 55 unique confirmed vulnerabilities in Chrome's V8 JavaScript engine versus 47 by mainline 3.5 Flash and 36 by Claude Opus 4.6 [4]. In a two-hour test on Google's own production systems, it uncovered remote code execution vulnerabilities in public APIs and a memory-corruption flaw in a sensitive service, then generated a functional RCE exploit bypassing both ASLR and W^X mitigations [4]. Google restricted the model to governments and trusted partners via its CodeMender platform. The Verge and CyberNews framed Flash Cyber primarily as a cheaper alternative to large dedicated AI security models such as Mythos [5][6], while The Neuron argued that the gatekeeping logic holds only as long as no competitor ships an equivalent without restrictions [2].

The unresolved delay of Gemini 3.5 Pro has deepened with new internal reporting. Bloomberg first reported in mid-July that the model had been held back because coding performance fell short of internal targets [7], and Alphabet's share price fell on that report [8]. The LA Times subsequently published an account drawing on internal sources describing coding stumbles alongside clashing teams and frustrated engineers — adding an organizational dimension to what Google had characterized as a pure performance problem [9]. InfoWorld argued the CEO's decision to emphasize Gemini 4 pre-training and a monthly release cadence served to redirect public attention from the stalled Pro model [10]. Google has disclosed that Gemini 4 pre-training has begun, describing it as its most ambitious pre-training effort yet, without attaching a timeline or capability claims [1]. Some social media commentary connected that announcement to possible disappointment with 3.5 Pro development [11], a link Google has not confirmed.

Timeline

  • 2026-05: Google I/O: Gemini 3.5 Flash announced as the flagship model [3]
  • 2026-06: Gemini 3.5 Pro expected to launch but does not ship [14][7]
  • 2026-07-16: Bloomberg reports Gemini 3.5 Pro delayed due to coding performance shortfall; Alphabet shares fall [8][7]
  • 2026-07-17: Google announces Gemini 3.5 Flash Cyber, restricted to governments and trusted partners via CodeMender [4]
  • 2026-07-17: LA Times publishes internal account of the Gemini delay, citing coding stumbles, clashing teams, and frustrated engineers [9]
  • 2026-07-21: Google releases Gemini 3.6 Flash and 3.5 Flash-Lite, deprecates 3.5 Flash, and discloses Gemini 4 pre-training has begun [1][3][15]
  • 2026-07-21: The Verge and CyberNews frame Flash Cyber as a cheaper alternative to large dedicated AI security models like Mythos [5][6]
  • 2026-07-22: Multiple outlets publish analysis on why Google delayed Gemini 3.5 Pro [16][12][13][17]
  • 2026-07-23: InfoWorld argues Google's CEO used Gemini 4 talk and a monthly release commitment to distract from the Pro delay [10]
  • 2026-07-23: The Neuron frames Google's three-model lineup as a specialization strategy and questions the durability of Flash Cyber gatekeeping [2]

Perspectives

Google DeepMind

Presents 3.6 Flash as an efficiency and capability improvement, frames Flash Cyber's vulnerability-finding results as evidence for a lightweight multi-invocation security AI approach, and argues CodeMender access controls are the right safeguard for a dual-use model.

Evolution: Consistent promotional framing; 3.6 Flash's release implicitly acknowledges 3.5 Flash underdelivered on coding without stating it directly.

LA Times

Internal reporting describes the Gemini 3.5 Pro delay as driven by coding stumbles compounded by organizational problems — clashing teams and frustrated engineers — not solely a benchmark shortfall.

Evolution: New in this pass; introduces an organizational dimension absent from Bloomberg's earlier technical framing.

InfoWorld

Argues Google's CEO framed Gemini 4 pre-training and a monthly release cadence as a forward-looking story to redirect attention from the 3.5 Pro delay.

Evolution: New in this pass; the most pointed editorial critique of Google's communications strategy in this thread.

Ars Technica (Ryan Whitwam)

Reports the releases factually while arguing Google's efficiency-first training priorities may have caused 3.5 Flash's coding shortfall that 3.6 Flash is now correcting, and notes the awkward optics of deprecating a two-month-old flagship.

Evolution: Consistent; no update from original report.

Bloomberg

Reported that Gemini 3.5 Pro's delay stems from coding performance falling short of internal targets, a claim that moved Alphabet's share price.

Evolution: Consistent; the LA Times adds organizational detail on top of Bloomberg's technical framing.

The Verge / CyberNews

Frame Flash Cyber primarily through a market-positioning lens — as a cheaper, smaller alternative to large dedicated AI security models like Mythos — rather than through the access-control or dual-use framing Google emphasizes.

Evolution: Consistent with their initial framing; no new reporting.

The Neuron (Eric Gerard Ruiz)

Analytically favorable toward Google's specialization approach, framing the three-model lineup as a mature admission that purpose-built models outperform generalists for agents; raises a pointed question about whether access-control gatekeeping for Flash Cyber holds once a competitor ships an equivalent publicly.

Evolution: Consistent with introduction in prior pass; no new reporting.

Tensions

  • Bloomberg attributes the 3.5 Pro delay to a coding performance shortfall; LA Times internal reporting adds clashing teams and frustrated engineers, suggesting organizational dysfunction is also a factor. [7][9]
  • Google presents 3.6 Flash's gains as straightforward improvement; Ars Technica argues Google's cost-reduction focus may have produced the coding underperformance in 3.5 Flash that 3.6 Flash is now correcting. [3][1]
  • Google frames Flash Cyber's restricted access as the appropriate safeguard for a model that generates working exploits; The Verge and CyberNews treat the same model primarily as a cheaper market entrant, implicitly downplaying the dual-use framing. [4][5][6]
  • Google argues access controls are sufficient to manage Flash Cyber's dual-use risk; The Neuron argues that logic holds only until a competitor ships a comparable model without restrictions. [4][2]
  • Google publicly frames Gemini 4 pre-training as its most ambitious effort yet; InfoWorld argues the announcement functioned as a distraction from the unresolved 3.5 Pro delay. [1][10]

Sources

  1. [1] Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — DeepMind Blog (2026-07-21)
  2. [2] 😺 You Need 3 Geminis Now — The Neuron (2026-07-23)
  3. [3] Google announces Gemini 3.6 Flash and cybersecurity AI, teases 3.5 Pro and Gemini 4 — Ars Technica AI (2026-07-21)
  4. [4] Introducing Gemini 3.5 Flash Cyber — DeepMind Blog (2026-07-17)
  5. [5] Google launches a cheaper alternative to large AI security models like Mythos — reactive:google-gemini-36-launch
  6. [6] Google debuts smaller, cheaper Gemini Flash AI cybersecurity model — reactive:google-gemini-36-launch
  7. [7] Gemini 3.5 Pro Delayed Over Coding, Bloomberg Reports — reactive:google-gemini-36-launch
  8. [8] Alphabet shares fall on report its AI model Gemini 3.5 Pro is delayed — reactive:google-gemini-36-launch (2026-07-21)
  9. [9] Inside Google's Gemini delay: Coding stumbles, clashing ... — reactive:google-gemini-36-launch
  10. [10] Google CEO distracts from Gemini 3.5 Pro delay with talk ... — reactive:google-gemini-36-launch
  11. [11] Chubby♨️ on X: "Google has begun pre-training Gemini 4, marking a completely new foundation model. This is really exciting! The announcement blog for 3.6 Flash states that Gemini 4 is being completely revamped. Presumably, the recent developments for 3.5 Pro were disappointing, so they're starting with a new pre-training program. This could be interesting!" / X — reactive:google-gemini-36-launch
  12. [12] Google Gemini 3.5 Pro faces "delays" over coding performance misses - Neowin — reactive:google-gemini-36-launch
  13. [13] Gemini 3.5 Pro Delay: Why Google Postponed Its AI Again - Geeky Gadgets — reactive:google-gemini-36-launch
  14. [14] Gemini 3.5 Pro delays due to coding performance — reactive:google-gemini-36-launch
  15. [15] Gemini 3.6 Flash released on AIStudio — reactive:google-gemini-36-launch (2026-07-21)
  16. [16] Google Delays Gemini 3.5 Pro Due to Coding Goals | Learnijoy NewsCenter — reactive:google-gemini-36-launch
  17. [17] Google Delayed the Gemini 3.5 Pro AI Model. Here's a Theory on Why. - Business Insider — reactive:google-gemini-36-launch