Latest / Neural Newscast / Google Launches Gemini 3.7 Flash as Pro Model Stalls [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast. This is Model Behavior, [00:02] Nina Park: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: Welcome to Model Behavior, a Neural Newscast production. I'm Nina Park. Today is August 18th, 2026, [00:21] Nina Park: and we are focusing on the deployment strategies [00:23] Nina Park: and infrastructure of modern artificial intelligence. [00:27] Thatcher Collins: We examine how these systems are built, optimized, and operated in professional environments. [00:33] Thatcher Collins: Joining me today to discuss Google's recent and rapid update cycle is Thatcher Collins. [00:39] Nina Park: It is good to be here, Nina. The sheer velocity of these updates is the primary story right now. On [00:45] Nina Park: August 13th, Google launched Gemini 3.7 Flash. [00:50] Nina Park: What is remarkable is that this release follows only 23 days after [00:54] Thatcher Collins: the launch of Gemini 3.6 Flash on July 21st. We are observing a production release cycle [01:00] Thatcher Collins: that is moving significantly faster than anything else in the industry. [01:04] Thatcher Collins: Clearly prioritizing their workforce model over the flagship promises made earlier this year. [01:10] Thatcher Collins: That cadence is certainly unusual. In this industry, we typically anticipate [01:14] Thatcher Collins: several months between significant model iterations. However, seeing [01:19] Thatcher Collins: three distinct flash family releases in under three months points [01:23] Thatcher Collins: toward a very deliberate engineering strategy. Thatcher, if we look at the benchmarks [01:28] Thatcher Collins: from artificial analysis, what exactly is Google achieving with these rapid-fire updates [01:33] Thatcher Collins: to the Flash model? Is the focus primarily on latency, [01:37] Thatcher Collins: or are we seeing a tangible jump in reasoning capability as well? [01:41] Nina Park: It appears to be a dual-track improvement. According to data from artificial analysis, [01:46] Nina Park: Gemini 3.5 Flash originally launched with an intelligence index score of [01:50] Thatcher Collins: 52 and a throughput speed of roughly 172 tokens per second. The 3.6 update maintained [01:57] Thatcher Collins: that intelligence level but optimized the engine to reach 225 tokens per second. Now, [02:03] Thatcher Collins: with the launch of Gemini 3.7 Flash, the intelligence index has climbed to 56, [02:08] Thatcher Collins: while generation speed has hit 340.1 tokens per second. That represents a massive leap [02:14] Thatcher Collins: in total throughput for a model aimed at coding and agentic workflows. [02:18] Thatcher Collins: 340 tokens per second is an exceptionally high bar for production environments, [02:24] Thatcher Collins: particularly at the aggressive price point Google is targeting. However, Thatcher, [02:29] Thatcher Collins: we should address the discrepancy in their roadmap. While the Flash models are receiving [02:33] Thatcher Collins: these rapid-fire upgrades, the model the industry was truly waiting for, Gemini 3.5 Pro, [02:40] Thatcher Collins: seems to be lingering in a holding pattern. [02:43] Thatcher Collins: Wasn't that flagship supposed to be widely available by now? [02:46] Nina Park: That is the critical question, Nina. Back on May 19th, Google officially launched 3.5 Flash [02:53] Nina Park: and explicitly stated that Gemini 3.5 Pro would roll out next month, [02:58] Nina Park: which would have placed the release in June. We are now in mid-August [03:03] Announcer: and well past that internal deadline. [03:06] Nina Park: While we have seen specialized iterations like 3.5 Flash Light and 3.5 [03:10] Announcer: Flash Cyber, and now the 3.7 Flash, the Pro model remains noticeably absent [03:16] Announcer: from the public product line and lacks a finalized price sheet or public scorecard. [03:21] Announcer: The flagship has essentially become a moving target, while the smaller model thrives in production. [03:27] Thatcher Collins: This situation creates a distinct dilemma for developers and enterprise architects. [03:33] Thatcher Collins: If you are currently building a complex agent or a sophisticated coding tool, [03:37] Thatcher Collins: do you wait for the pro capabilities Google originally promised, or do you commit your resources [03:43] Thatcher Collins: to the model that is actually available and improving every few weeks? [03:48] Thatcher Collins: It appears Google is effectively making that choice for the market [03:51] Thatcher Collins: by offering a highly attractive cost-to-performance ratio [03:54] Thatcher Collins: on the hardware they can currently scale today. [03:57] Announcer: That is the most pragmatic takeaway for any engineer. Analysis from Towards AI notes [04:03] Announcer: that at its introductory price, Gemini 3.5 Flash reaches a cost of about 40 cents [04:08] Announcer: per intelligence index task. For developers building agents this week, [04:12] Announcer: that makes it a very serious production contender. The advice here is straightforward, Nina. [04:17] Announcer: Prioritize the model you can actually purchase, benchmark, and scale right now. You cannot build [04:23] Announcer: reliable business logic [04:25] Announcer: or production code around an attractive product name that does not have a broad release date yet. [04:30] Thatcher Collins: It suggests that Google has developed a very efficient pipeline for iterating on the models already [04:35] Thatcher Collins: in the hands of millions of users. While the industry continues to wait for the next major leap [04:41] Thatcher Collins: in flagship pro performance, the Flash family is effectively filling that void by becoming smarter [04:47] Thatcher Collins: and significantly faster at a relentless pace. It marks a clear shift from waiting for [04:53] Thatcher Collins: the large model to perfecting the efficient one. Thatcher, [04:57] Thatcher Collins: what is your final bottom line for developers navigating this roadmap? [05:01] Announcer: The bottom line is to prioritize tools that are currently scalable and verified by benchmarks, [05:05] Announcer: rather than waiting on future promises. Thank you for listening to Model Behavior, a Neural Newscast [05:11] Announcer: editorial segment. Nb.neuralnewscast.com. Neural Newscast is AI-assisted, [05:16] Announcer: human-reviewed. View our AI transparency policy at neuralnewscast.com. [05:20] Nina Park: This has been Model Behavior on Neural Newscast, examining the systems behind the story.