Mistral Forge and Google’s New Gemini Tiers [Model Behavior]

[00:04] Nina Park: I'm Nina Park. Welcome to Model Behavior. On this program, we examined the construction and operational deployment of AI systems within professional environments. [00:15] Nina Park: Today is March 18, 2026, and we are tracking several major shifts in how the industry's most prominent players are positioning their latest frontier models. [00:26] Thatcher Collins: And I'm Thatcher Collins. [00:28] Thatcher Collins: Yesterday at the NVIDIA GTC conference, we observed a significant strategic pivot in how European AI is courting the enterprise sector, Nina. [00:39] Thatcher Collins: Specifically, we are seeing a renewed emphasis on data sovereignty and localized infrastructure as startups move to compete with the dominant American cloud providers, Nina. [00:50] Nina Park: The primary focus of that shift is the French developer Mistral. [00:54] Nina Park: They recently announced Mistral Forge, a platform designed to allow corporations to build bespoke models [01:00] Nina Park: trained specifically on their own proprietary documents and internal workflows. [01:06] Nina Park: This represents a departure from Standard Retrieval Augmented Generation, or RAG, which simply pulls data at the moment of the query. [01:14] Nina Park: Mistral claims Forge allows companies to actually train models from the ground up using their updated Open Weight Library, which now includes the flagship Mistral Small 4. [01:26] Nina Park: By moving the training process in-house, they are promising a level of performance and security [01:32] Nina Park: that multi-tenant cloud systems often struggle to guarantee. [01:36] Thatcher Collins: I want to look at the friction inherent in that approach, Nina. [01:40] Thatcher Collins: Traditionally, enterprises have favored fine-tuning because full-scale training from scratch [01:45] Thatcher Collins: is notoriously resource-intensive and requires high-level technical expertise that many firms simply don't have. [01:53] Thatcher Collins: Mistral is attempting to lower that barrier by embedding forward-deployed engineers directly within client teams. [02:01] Thatcher Collins: It is a labor-intensive strategy we've seen successfully utilized by firms like IBM and Palantir. [02:08] Thatcher Collins: with a target of $1 billion in annual recurring revenue. [02:12] Thatcher Collins: This calendar year, they are betting that industrial giants, such as the semiconductor leader ASML, [02:18] Thatcher Collins: are seeking a degree of control and privacy that the current OpenAI ecosystem does not provide. [02:25] Nina Park: It certainly signals a move away from the general consumer spotlight currently occupied by Anthropic and OpenAI. [02:32] Nina Park: Transitioning to subscription models, Google also moved to reorganize its consumer ecosystem yesterday. [02:38] Nina Park: They have rebranded Google One AI Premium to Google AI Pro [02:43] Nina Park: and introduced a new higher-tier offering called Google AI Ultra. [02:47] Nina Park: This appears to be a consolidation effort aimed at simplifying their messaging [02:51] Nina Park: as their AI services expand across different hardware platforms. [02:55] Thatcher Collins: The nomenclature is becoming quite dense, Nina. [02:59] Thatcher Collins: The AI Pro tier remains $20 monthly, but it now features a 1 million token context window [03:06] Thatcher Collins: and expanded usage limits for their high-reasoning thinking models. [03:11] Thatcher Collins: What is particularly notable is the hardware integration. [03:15] Thatcher Collins: Google Home Premium subscribers are now receiving Gemini Live on Nest devices. [03:21] Thatcher Collins: This facilitates much more natural language automation and allows the system to maintain household-specific memory, [03:28] Thatcher Collins: which could change how users interact with their environments on a daily basis. [03:33] Nina Park: It appears Google is pursuing total vertical integration. [03:38] Nina Park: From the multimodal reasoning of Project Mariner and Genie to the 30 terabytes of storage offered in the Ultra Tier, [03:47] Nina Park: they are positioning themselves as an all-encompassing service provider. [03:51] Nina Park: While Mistral builds specialized sovereign tools for entities like the European Space Agency, [03:58] Nina Park: Google is attempting to make Gemini the foundational invisible operating system [04:04] Nina Park: for both the private home and the professional inbox. [04:08] Thatcher Collins: Two very different paths for scaling model utility. [04:13] Thatcher Collins: Thank you for listening to Model Behavior. [04:16] Thatcher Collins: You can find more detail on these updates at mb.neuralnewscast.com. [04:22] Thatcher Collins: Neural Newscast is AI-assisted, human-reviewed. [04:26] Thatcher Collins: View our AI Transparency Policy at neuralnewscast.com.