Latest / Neural Newscast / SpaceXAI Launches Grok 4.5 with Cursor Integration [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior, [00:03] Nina Park: AI-focused news and analysis on the models shaping our world. [00:11] Nina Park: I'm Nina Park. Welcome to Model Behavior. [00:15] Nina Park: Model Behavior is our dedicated space where we examine the complex intersection [00:19] Nina Park: of artificial intelligence and the real world. [00:22] Nina Park: We focus specifically on how these powerful systems are architected, how they are being deployed in professional environments, and the tangible operational impacts they have on industry and governance alike. [00:33] Thatcher Collins: I'm Thatcher Collins. Today is July 9th, 2026. [00:37] Thatcher Collins: We are tracking a series of massive shifts in the AI market landscape this morning. [00:42] Thatcher Collins: From high-stakes corporate acquisitions that are consolidating the developer tool market [00:47] Thatcher Collins: to new benchmarks and model efficiency that challenge the current pricing models of the major labs, [00:53] Thatcher Collins: the industry is entering a more mature and cost-conscious phase. [00:57] Nina Park: We begin with yesterday's announcement from SpaceX AI, the entity formerly known as XAI. [01:04] Nina Park: The company has launched Grok 4.5, which they are characterizing as their first model built specifically for coding, autonomous agents, and professional knowledge work. [01:14] Nina Park: This release is particularly significant as it represents the first major product launch following SpaceX's $60 billion acquisition of the AI coding startup Cursor just weeks ago. [01:27] Nina Park: According to SpaceX AI, Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs, using data sets specifically curated for science, engineering, and mathematics. [01:39] Thatcher Collins: Nina, the hardware is impressive, but what stands out here is the economic strategy. [01:45] Thatcher Collins: Musk is making a pricing argument rather than a pure intelligence claim. [01:49] Thatcher Collins: Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens. [01:56] Thatcher Collins: To put that in perspective, OpenAI's GPT 5.6 Sol variant costs $5 for input and $30 for output. [02:04] Thatcher Collins: SpaceX is claiming that Grok 4.5 is roughly comparable to Anthropik's Claude Opus 4.7 in terms of capability, [02:11] Thatcher Collins: but functions at a much lower cost and higher speed. [02:15] Nina Park: That is exactly the point VentureBeat and other outlets are highlighting. [02:19] Nina Park: For enterprise-level agentic workloads, where a model might run for hours across a massive codebase, [02:25] Nina Park: a 90% reduction in cost per task is a massive competitive advantage. [02:30] Nina Park: Independent evaluations from artificial analysis already rank Grok 4.5 fourth on their GDP valet index. [02:37] Nina Park: It sits just behind the latest Claude releases, but offers a Pareto frontier for performance versus cost. [02:43] Nina Park: Thatcher, how does the cursor integration actually manifest for developers? [02:48] Nina Park: Well, it's now the default model for GrokBuild, their terminal-based coding agent, but more importantly, the partnership allowed SpaceX AI to use high-quality interaction data from the cursor editor, basically learning from how expert engineers edit and debug code in real time. [03:03] Nina Park: However, we have to acknowledge the baggage here. [03:06] Nina Park: Grok has had a turbulent year, including scandals involving generated anti-Semitic content [03:10] Thatcher Collins: and deepfake investigations in the UK and EU. [03:14] Thatcher Collins: While the engineering looked solid, the regulatory environment in the European Union means this [03:18] Nina Park: model won't even be available there until mid-July at the earliest. [03:22] Nina Park: It's a rebuilding from the foundations moment, as Musk described it. [03:27] Nina Park: While SpaceX AI chases the developer market, OpenAI is focused on the interface. [03:34] Nina Park: This morning, OpenAI began the global rollout of its new GPT-Live voice models. [03:40] Nina Park: There are two versions, the full GPT-Live 1 for paid users and a mini version for the free tier. [03:47] Nina Park: The core innovation here is full duplex conversation, meaning the model can listen and speak simultaneously without the awkward interruptions or silence filling that plagued previous versions. [03:59] Thatcher Collins: I was looking at the technical briefing for this, Nina. [04:02] Thatcher Collins: One of the most interesting features is how it handles difficult queries. [04:07] Thatcher Collins: Instead of trying to solve everything within the voice model itself, [04:10] Thatcher Collins: GPT Live 1 now delegates complex questions to the GPT 5.5 text model in the background. [04:17] Thatcher Collins: It will actually say something like, let me check that for you, [04:21] Thatcher Collins: to maintain a natural conversational flow while it waits for the more powerful model to process the data. [04:27] Thatcher Collins: it feels less like a chatbot and more like a human assistant performing a lookup. [04:31] Nina Park: They're also emphasizing simultaneous translation. [04:35] Nina Park: OpenAI demonstrated an agent receiving a live feed in their own language while a speaker was talking, [04:42] Nina Park: similar to what you'd see at the United Nations. [04:44] Nina Park: This is a clear play to make voice the primary interface for computing. [04:49] Nina Park: The company is even allowing users to call the AI by a name. [04:53] Nina Park: In demos, they used chat to signal when the agent wants a response, [04:57] Nina Park: which... [04:58] Thatcher Collins: which solves that persistent problem of the AI jumping in whenever there's a pause in the room. [05:02] Thatcher Collins: That's useful, but the friction point remains data privacy. Axios reported that audio is stored for [05:09] Thatcher Collins: 30 days for context and memory. While OpenAI says training on voice queries is turned off by default, [05:15] Thatcher Collins: the storage of that audio data is still going to be a sticking point for enterprise users who [05:20] Thatcher Collins: handle sensitive information. Nina, does this move toward voice suggest OpenAI is [05:25] Thatcher Collins: finally preparing for their own hardware device? [05:28] Nina Park: OpenAI was non-committal during the press briefing, stating they had no hardware news to share. [05:34] Nina Park: But the industry consensus is that voice-first models of this caliber are the prerequisite for any dedicated AI device. [05:41] Thatcher Collins: Moving from the consumer space to enterprise orchestration, Salesforce has also made a significant move this week. [05:48] Thatcher Collins: Five years after their $27 billion acquisition of Slack, [05:51] Thatcher Collins: the two products are finally functioning as a unified system. [05:55] Thatcher Collins: Slackbot can now pull CRM data, generate Tableau charts, [05:59] Thatcher Collins: and even send DocuSigns directly through conversational prompts. [06:02] Thatcher Collins: It's about time, honestly. For years, Slack and Salesforce felt like two separate islands. [06:09] Thatcher Collins: This integration allows the personal AI agent in Slack to access the entire Salesforce platform, [06:14] Thatcher Collins: including Data360 customer profiles. It's a direct response to the agentic trend we're [06:19] Thatcher Collins: seeing with Krondé Claude. If you can tell Slack to summarize my top five leads and draft an email, [06:24] Thatcher Collins: and it actually has the CRM context to do it, that's a massive productivity gain for sales teams. [06:29] Nina Park: exactly and speaking of claude we should mention anthropic's latest updates they're bringing claude [06:36] Nina Park: co-work to mobile and web platforms this allows tasks to start on a desktop and continue autonomously [06:42] Nina Park: in the background with the agent checking the progress on their phone later but the more [06:46] Nina Park: academic news from anthropic involves their j lens study their researchers claim to have [06:51] Nina Park: discovered a j space inside claude's neural network [06:54] Thatcher Collins: This study is fascinating because it moves past the black box problem. [06:58] Thatcher Collins: They used a new mathematical technique to find a specific zone where the model holds concepts [07:02] Thatcher Collins: it can actually reason with and report on, surrounded by a larger area of automatic processing. [07:08] Thatcher Collins: It almost mirrors human theories of consciousness, where we have a global workspace for active thought. [07:13] Thatcher Collins: Nina, does this give Anthropic a leg up in the safety and transparency debate? [07:17] Nina Park: It certainly helps their positioning as the safety-first lap. [07:20] Nina Park: If they can mathematically prove how a model arrives at a specific verbalized conclusion, [07:25] Nina Park: it makes the system more auditable. [07:27] Nina Park: However, the market is moving fast. [07:29] Nina Park: Anthropik's Claude 4.7 is currently the top-ranked model for agentic work, [07:34] Nina Park: but with SpaceX AI undercutting them on price and OpenAI improving the interface with GPT Live, [07:39] Nina Park: the pressure is on Anthropik to maintain that quality lead. [07:43] Thatcher Collins: We are seeing a shift in the intelligence scoreboard. [07:47] Thatcher Collins: For the last three years, the industry was obsessed with who had the smartest model, [07:52] Thatcher Collins: Now, as we see with Grok 4.5 and the OpenAI voice updates, the focus is shifting to who [07:58] Thatcher Collins: has the most affordable and usable model. [08:01] Thatcher Collins: Musk isn't trying to outthink Claude. [08:04] Thatcher Collins: He's trying to make it too expensive for companies to choose anything other than Grok [08:08] Thatcher Collins: for their heavy lifting. [08:10] Nina Park: It's a strategy we've seen him use in spaceflight and electric vehicles, driving [08:15] Nina Park: down the cost of entry until the incumbents have to pivot. [08:19] Nina Park: Whether that works in the software layer, where quality can be more subjective than rocket thrust, is the multi-billion dollar question. [08:28] Nina Park: We'll be watching the Enterprise pilots of Grok 4.5 closely over the next few weeks to see if the value proposition holds up. [08:35] Thatcher Collins: And I'll be testing that GPT Live translation on the go. [08:39] Thatcher Collins: If it's as seamless as the product demos suggest, the barrier between humans and AI agents just got significantly thinner. [08:46] Thatcher Collins: It will be interesting to see if these voice interfaces can finally overcome the latency [08:50] Thatcher Collins: issues that have held back real-world adoption in the past. [08:53] Nina Park: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [08:58] Nina Park: You can find our technical show notes and further analysis online at mb.neuralnewscast.com. [09:04] Nina Park: Neural Newscast is AI-assisted, human-reviewed. [09:08] Nina Park: View our AI transparency policy at neuralnewscast.com. [09:12] Announcer: This has been Model Behavior on Neural Newscast. [09:15] Announcer: Examining the Systems Behind the Story