MARKETS

How a mystery model surprised the AI industry

An anonymous AI model has taken the developer community by storm. And it's reportedly giving frontier labs a run for their money.

On Thursday, a "stealth model" called Ox Alpha was released on OpenRouter by a third-party provider who decided to remain anonymous during its preview. Its creator said that the reasoning model was designed for coding, sustained agentic work and production workloads.

The model is available for free in preview with "near unlimited usage" for a week, according to OpenCode, with an eye-popping capacity for 100 trillion tokens per day. The model features a 1.05-million token context window and 131,000-token output, or the limit that a model can generate in a single response. 

• The creator noted that the model is suited for long-horizon software development tasks and workflows that "combine text with visual context." 

• Additionally, though prompts and outputs are retained by the model provider, they aren't currently used for training. 

• Though early reports suggest the mystery model is beating frontier models like Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol in coding evaluations like Deep SWE, the model's independent validation isn't yet available on formal public leaderboards. 

Patrick Collison, CEO of Stripe, which is acquiring OpenRouter, described Ox Alpha in a post on X as "very impressive." 

As the model goes viral, many have started to speculate about its origins. While some have suggested that Ox Alpha is the next generation of Google's Gemini models or Microsoft's MAI models, others have speculated that the model comes from a Chinese lab, with some analysis pointing to Z.ai's GLM family. However, none of these guesses have been confirmed.

No matter what the origins of this model end up being, Ox Alpha's overnight stardom shows that the AI industry is easily distracted, and often excited by every shiny new toy that hits the market. However, early testing and industry fervor are one thing, and actual, long-running sustainability and results are another. Without knowing who has created this model, users should be wary about its safety and security in real-world applications. Additionally, because the model doesn't currently use user data for training, that policy could change. In short: While the model is drumming up a lot of excitement right now, it's yet to be seen whether this is a flash in the pan, or if the industry will lose interest once the next big model gets released in the days (or even hours) ahead. Still, this supports the bigger trend of frontier intelligence commoditizing.

TOGETHER WITH GRANOLA

Be The Person Who Remembers Everything

Ever notice those people who walk into meetings and somehow remember every conversation, decision, and next step from the last time? They’re not some kind of savant – they just use Granola. 

Granola’s AI notepad captures your meetings automatically. No bots, no awkward recording prompts, just a quietly powerful log of everything that matters. Ask Granola questions via chat and you’ll instantly get the answer you’re looking for, including who it came from, when it was solved, and what notes were taken on it. Best of all, the more you use it, the better Granola gets – over time, it builds a searchable memory of every discussion and decision made. 

If you’re ready to become the person in the room who remembers everything, it’s time to try Granola for free right here.

STARTUPS

Where small models are challenging AI giants

With enterprises finally clamping down on tokenmaxxing, small models have emerged as a counterweight to expensive proprietary APIs. 

On Monday, marketing and behavioral insights firm NextLM unveiled research detailing the results of its new system, Savant 3.5, built on lightweight, open-source NVIDIA Nemotron models and purpose-made for one specific task: helping salespeople find the best customers. In its niche task, the model outperformed major models from OpenAI, Google, Anthropic, SpaceXAI, and Moonshot, and at a much lower price. 

"I think it validates the idea that small models can not only compete with the frontier, but actually, in some cases, win," Chris Anzalone, founder and CEO of NextLM, told The Deep View. 

Savant is built on top of one of Nvidia's most compact models, sitting at roughly 30 billion parameters with 3 billion active for each input, to find customers. Then it uses Nemotron

3.5 Lightning as a "supporting signal" to sharpen the rankings. 

Here are the results: 

• The most important result was how many eventual customers appeared in the top 10% of a model's customer recommendations. NextLM's Savant model put 24.6% of buyers in that 10%. The closest competitor, GPT-5.6 Sol at 22.8%. 

• Grok 4.5 scored third-highest at 22.3%. Anthropic's models, meanwhile, sat near the bottom of the group, with Fable 5 sitting at 18.1% and Opus 5 ranking dead last at 14.6%. 

• Additionally, Savant managed to do so at a much cheaper rate, sitting between $0.003 and $0.011 per 1,000 customer prospects scored, compared to between $0.26 to $5.11 for the frontier-model APIs tested. 

NextLM's success provides an example of where small models fit best: niche, specialized tasks. "If a model isn't tuned to your business outcomes, the value really isn't yours," Anzalone said. "It's transferable to whoever uses that same model."

Small models offer several benefits over bulky, expensive models accessed through proprietary APIs. As Savant proved, these models can be trained to do hyper-specific tasks and achieve hyper-specific business outcomes at a fraction of the cost. Additionally, because these models can often be stored on local hardware, they're generally more private and secure than handing your business's most critical data over to a model provider. And for enterprises desperately searching for returns while also being on high alert amid increasingly common AI-powered cyberattacks, small models may sound like attractive prospects. However, that doesn't mean that enterprises are bound to stop using frontier models and APIs entirely. These systems have a few significant advantages: They're easy to use and provide a readily available source of general capabilities. As Anzalone notes, "People are still going to use general models because they have general information." The challenge enterprises now face is striking the balance between the two.

TOGETHER WITH QUIQ

Most CX teams aren't struggling because the model is too weak

Most CX teams aren't struggling to build AI agents because the model is too weak. They're struggling because customer inquiries are messy, policies have edge cases, and every answer needs the right mix of speed, judgment, and control. That's usually where a strong demo starts to quietly fall apart.

In Quiq's new guide, "CX Agent Design: How to Build One You Can Read and Trust," it breaks down the two dead ends most teams hit when building AI for customer service, the building blocks (Guides, Skills, Tools, and governed boundaries) that replace them, and why a CX agent needs more than good code to actually hold up in production.

Get your free copy of the guide here.

HARDWARE

Why foldables make more sense in the AI era

For almost a decade, foldable phones have been a product looking for a problem to solve. They may have found their 

In a special episode of The Deep View Conversations, we make sense of Google's and Samsung's latest hardware and the AI announcements that came with them. But mostly, we talk about the new folding phones, the Pixel 11 Pro Fold and the Z Fold 8. 

While folding and flip phones have existed for years, this summer both Google and Samsung upped the ante by launching new experiences that let AI enthusiasts make the most of the added screen real estate for AI workflows. 

Topics covered include:

• The new AI features available on the Pixel 11 phones 

• How Gemini contributes to the AI experience on mobile

• Does Google still have the lead in AI hardware?

• The minimal hardware improvements to the Pixel devices

• The advantages of owning a foldable in the AI era 

• How Samsung's Galaxy Z Fold 8 series compares

• The advantages of the Z Fold 8's "passport" form factor

• How Apple's foldable, rumored to launch in September, will compete 

If you're trying to understand how AI is changing what you can do with a smartphone, and what your next phone purchase should be if you prioritize AI, you won't want to miss this episode. 

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

LINKS

• Apple cuts 200-plus jobs from Vision Pro and Siri teams

• DeepSeek unveils experimental multimodal AI model rivaling Anthropic

• Anthropic hires ex-Google TPU chief Amir Salek for in-house chip push

• South Korea's Kakao Corp to spin off chat platform into new KakaoAI unit

• Anthropic could reportedly raise $100 billion in IPO

• Hugging Face explores M&A deals worth $13 billion 

• Anthropic's Opus 5 overtakes Fable 5 in enterprise spending

• Claude Security: scans now run on Mythos 5, public beta for Enterprise customers

• Antigravity Remote Control: Access active sessions on any modern supported browser or mobile device across iOS and Android

• Gemini Notebook: Unveiled a slew of updates

• GPT-5.6 Sol: OpenAI cuts API pricing by over 20%

• Nvidia: Solutions Architect, AI Models

• Logicbroker: AI Engineer

• Hilbert: AI Engineer - Core

• Deloitte: Google Gemini Manager, Technical Transformation

GAMES

A QUICK POLL BEFORE YOU GO