Latest / Neural Newscast / Google Ships Three New Gemini Flash Models [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models [00:05] Announcer: shaping our world. [00:11] Nina Park: I'm Nina Park. [00:13] Nina Park: Welcome to Model Behavior, a Neural Newscast editorial segment where we examine how artificial [00:20] Nina Park: intelligence systems are actually built, deployed, and operated in real-world professional environments, [00:27] Nina Park: Today is July 21st, 2026, and we are looking at a particularly busy window for Google, [00:33] Nina Park: which released three new model variants just before its scheduled earnings call. [00:38] Nina Park: We will also analyze a new strategic move from KPMG and OpenAI that aims to fundamentally change how we interact with enterprise software. [00:48] Thatcher Collins: I'm Thatcher Collins. [00:49] Thatcher Collins: Nina, the timing of these releases is certainly noteworthy. [00:53] Thatcher Collins: Today, Google shipped Gemini 3.6 Flash, 3.5 Flash Lite, and a specialized security-focused version called Flash Cyber. [01:03] Thatcher Collins: This blitz comes immediately after several industry reports suggested that their flagship model, Gemini 3.5 Pro, has been delayed. [01:12] Thatcher Collins: It appears Google is attempting to manage the market narrative by flooding the zone with high-efficiency models while that larger Pro update is still in the oven, so to speak. [01:22] Nina Park: The efficiency gains in these new variants are substantial, Thatcher. [01:26] Nina Park: According to reporting from AI Unfiltered, the updated 3.6 flash uses 17% fewer output tokens to perform the same complex work as its predecessor. [01:38] Nina Park: Google has also cut prices for output tokens by roughly 16%. [01:43] Nina Park: In terms of performance, it is currently leading on the OS World Verified benchmark for computer use, with a score of 83%, which actually edges out GPT-5.6 Luna. [01:55] Nina Park: It seems Google is pivoting toward being the primary choice for browser automation and long-context tasks. [02:01] Thatcher Collins: It is a strong lead in the computer use category, but if you look at coding agents specifically, the picture remains a bit more mixed. [02:09] Thatcher Collins: On the SWE Bench Pro metric, Gemini 3.6 Flash is still trailing behind XAI's Grok 4.5. [02:17] Thatcher Collins: Then there is Flash Cyber. [02:19] Thatcher Collins: Google claims it outperforms Claude Opus in vulnerability hunting, but the fine print is revealing. [02:25] Thatcher Collins: Newer Claude models were not included because Anthropics' safety policies prevent them from attempting those types of autonomous hacking tasks. [02:32] Thatcher Collins: It is not exactly a clear-cut victory if the primary competition opted out. [02:37] SPEAKER_03: That nuance is important, Thatcher, especially as these models become more deeply integrated into professional developer workflows. [02:45] Nina Park: We are also seeing a growing trend toward greater transparency in how these models arrive at their specific answers. [02:51] Nina Park: Ace Data Cloud recently added these new Gemini models to their API, and they highlighted a specific feature called visible reasoning. [03:00] Nina Park: The API now provides a dedicated field that exposes the model's step-by-step thinking trail before it commits to a final reply, allowing for much better auditing. [03:09] Thatcher Collins: That reasoning content field is a significant tool for developers in debugging. [03:14] Thatcher Collins: However, beyond just text, the multimodal capabilities of Gemini 3.1 Pro are also expanding. [03:21] Thatcher Collins: The current documentation describes a model that does not just look at video frames in isolation, [03:26] Thatcher Collins: but understands the intricate relationship between sight and sound. [03:30] Thatcher Collins: In one notable example, it watched a video of a cat and noted that the audio track, which featured heavy hoofbeats, was comically mismatched with the visuals. [03:39] Thatcher Collins: That level of concurrent processing is where Google still maintains a distinctive technical edge. [03:45] Nina Park: While Google focuses on model efficiency, OpenAI is leaning heavily into the enterprise implementation layer. [03:52] Nina Park: Today, Fortune reported that OpenAI has named KPMG an elite partner, which is their highest tier of collaboration. [03:59] Nina Park: KPMG actually served as the pilot client Zero for this initiative, building its own internal supply chain platform using OpenAI's tools. [04:07] Nina Park: They are now pitching a vision called Headless Software, which suggests that the era of navigating through clumsy application screens and rigid software modules is finally coming to an end. [04:18] Thatcher Collins: It is an ambitious pitch, Nina. [04:20] Thatcher Collins: The idea is that your CRM or ERP system becomes a silent back-end infrastructure. [04:25] Thatcher Collins: While an AI agent handles the entire user interface through voice or text, [04:31] Thatcher Collins: you wonder about the decision sandwich framework that economists use to describe the shift. [04:36] Thatcher Collins: AI effectively compresses the execute layer, but it often increases the cognitive burden on the decide-and-deliver layers. [04:44] Thatcher Collins: You might spend less time clicking buttons, but you will spend significantly more time verifying that the agent actually executed the correct command. [04:51] Nina Park: KPMG seems to anticipate that specific friction, Thatcher. [04:55] Nina Park: They argue that their primary value is not just the technology itself, but their institutional knowledge of a client's specific culture and internal politics— [05:04] Nina Park: things a frontier model simply does not know. [05:08] Nina Park: As Google deals with the departure of key researchers like Noam Shazir and John Jumper, [05:12] Nina Park: the competition is clearly moving from just building the smartest possible model [05:17] Nina Park: to finding out who can actually make these systems work inside a complex and often messy corporation. [05:23] Thatcher Collins: It suggests that despite the push for total automation, [05:27] Thatcher Collins: the human bottleneck at the beginning and the end of the workflow remains the most critical factor. [05:32] Thatcher Collins: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [05:37] Thatcher Collins: mb.neuralnewscast.com [05:39] Thatcher Collins: Neural Newscast is AI-assisted, human-reviewed. [05:43] Thatcher Collins: View our AI transparency policy at neuralnewscast.com [05:46] Announcer: This has been Model Behavior on Neural Newscast. [05:50] Announcer: Examining the systems behind the story.