BIG TECH

Google fights for AI ground with a cheaper Gemini

Amid Google's executive shake-up, the company is still trying to keep up with the breakneck pace of AI model development. 

On Thursday, the company announced Gemini 3.7 Flash, which it calls its most intelligent "workhorse" model to date for coding and agents. Google said in its announcement that the model offers substantial improvements in software engineering, knowledge work and development workflows. 

To start, the model will be priced the same as Gemini 3.6 Flash, its cost-efficient model that it released in late July, at $0.75 per million input tokens and $3.75 per million output tokens. It's still undercut by OpenAI's lowest-cost model, GPT-5.6 Luna, at and $0.20 per million input tokens and $1.20 per million output tokens. 

Google noted that the new model is a "direct result" of developer feedback and offers significant improvements over its predecessor in tasks such as debugging and resolving issues.

• The model also offers better first-pass code accuracy and improved generation of production-ready code: Both in the FrontierCode 1.1 Main and the DeepSWE v1.1 coding benchmarks, 3.7 Flash saw improvements in score between 10 and 15 percentage points. 

• For developers, Google noted that the model is better at adapting to roadblocks, clarification and instruction following, as well as puts more effort into multi-step tasks and tool calling. 

• The model also shows improvements in "knowledge-dense fields" like finance, law and biosciences, and outperforms 3.6 Flash on the GDP.pdf benchmark for processing complex documents by 12 percentage points. 

• The model also shows better adherence in user interface development, generating more functional layouts and apps from simple prompts than the previous generation. 

Notably, Google said it shipped Gemini 3.7 Flash with updated safeguards against misuse for chemical, biological, radiological and nuclear use cases, as well as cyber offenses. 

The model is available in Google Antigravity, its agentic software development platform, through the Gemini API via the Google AI Studio and Android Studio, and through its enterprise platforms. Additionally, the model will be integrated into Gemini Spark, Google's personal AI agent, for Google AI Pro and Ultra subscribers. 

The model follows a significant executive shift at Google, in which two of the company's most influential figures, Demis Hassabis and Jeff Dean, announced big moves, with Hassabis shifting from CEO to chairman of DeepMind and chief scientist of Alphabet, and Dean leaving the company entirely to launch a new startup called Discovery Loop. 

Despite making a big splash with the launch of Gemini 3 last fall, the company has since struggled to keep pace with the other frontier labs, both in terms of the latest models as well as agents and coding tools.

Google has all the ingredients to make a big impact in AI. It has the brand equity laying the foundation for the modern internet, access to capital and compute resources, and a wide market reach among both enterprises and consumers. And there are a few things it can do to differentiate itself, as The Deep View's editor in chief Jason Hiner recently pointed out, including shipping a real rival to Claude Code, clarifying its brands, and using its ubiquity as an enterprise strength. Further, leaning into cost efficiency with Gemini Flash 3.7 could offer another advantage as the industry starts to reckon with massive inference bills. But most importantly, Google can't rest on its laurels. Though the company has critical advantages in compute, user trust, and research, those can't be the only things that it relies on to compete in AI, especially with frontier labs and neolabs alike moving at lightning speed.

TOGETHER WITH TABS

Tabs + PwC: A practical framework for usage-based revenue in the AI era

Hybrid and usage-based models are changing how companies price, and how finance teams recognize revenue, forecast, and close the books.

In this on-demand session, Tabs co-founder Rebecca Schwartz and PwC partner Amit Dhir break down the real operational impact, with frameworks and examples you can apply now.

Watch the session

MARKETS

Ramp data: Businesses are done paying for AI overkill

Anthropic is still leading the way among enterprises, but not with its best models. 

On Thursday, payments platform Ramp published its monthly index on AI spend among its enterprise users, and found that Anthropic has widened its lead in business AI adoption. According to the index, 43.5% of its U.S. business clients paid for Anthropic subscriptions or tokens, compared to 39.7% of businesses paying for OpenAI services. 

However, most of that spend wasn't on its ultrapowerful new models, with only 6% of token spend being attributed to Fable 5, Anthropic's most capable (and expensive) model on the market. 

• Despite Anthropic's lead, Ramp noted that Fable 5 was less popular with businesses than rival OpenAI's GPT-5.6 Sol, which comprised 25% of tokens used by OpenAI Ramp customers. 

• "So with Fable 5, we’ve found a new upper bound for how much businesses are willing to spend on AI. Here, more performance is not worth the price tag," Ara Kharazian, lead economist at Ramp, said in the company's blog post. 

Notably, other AI contenders saw significant gains: The share of models using serving platforms, like Hugging Face, such as those that provide open-source and Chinese models, saw gains in the past month, with 6.1% of Ramp customers that use AI buying into those platforms. Additionally, xAI saw its fastest growth month since July of last year, rising 0.94 percentage points month-over-month and now making up 4% of AI business spend.   

Kharazian noted that both OpenAI and Anthropic, however, have seen slower adoption in recent months. "That’s not because new AI spenders are switching to open-source/Chinese models … it means more of their growth will have to come from existing businesses spending on AI, particularly the advanced spenders, and those businesses are increasingly spending on open-source," he wrote. 

It's important to note that Ramp's data comes with a caveat: It only covers Ramp customers, which are largely startups and small businesses, though it has a growing list of enterprise clients as well. Ramp is also a relatively new company, only starting in 2019, and many large enterprises rely on legacy expensing and billing systems that Ramp doesn't account for. 

Still, the data points to a growing trend that we're seeing among businesses: We are in a post-tokenmaxxing era, and these companies are trying to put a leash on their spending without quashing their AI ambitions entirely.

Ramp's data may represent a reality check for the AI industry. Having the bulkiest, most expensive and most capable model may be the thing that helps these labs get an upper hand on benchmarks, but it doesn't mean much to the end user or enterprise buyer if it makes their wallets sting. Efficiency is quickly becoming the new paragon of AI, with companies like Microsoft, Meta and Google all shifting their attention to lightweight models. It's also why open-source models are growing in popularity, and why companies like Pathway are making strides with models as small as 150 million parameters. That's because the vast majority of enterprise tasks don't need multi-trillion parameter models. For things like writing emails, scheduling meetings or making documents, that kind of power is often overkill. And while the shift towards efficiency may sound like good news for easing the crunched compute market, it calls into question the blind devotion to scaling laws that companies like OpenAI and Anthropic have staked their missions on. 

TOGETHER WITH AWS RE:INVENT

AWS re:Invent 2026 is designed for hands-on problem-solving

Over 2,200 sessions, 70% are interactive. re:Invent is returning to Las Vegas, November 30 - December 4.

You’ll have the opportunity to: 

• Run a service against a realistic workload before it hits production

• Ask a service engineer about the constraint that's actually blocking you

• Access AWS certification discounts, with exam prep built into the schedule

• Talk through implementations with 400+ partners on the expo floor

Learn more and sign up by August 25th to save $1,200.

PRODUCTS

Android's big leap is from apps to agents

The smartphone has been built around apps and taps for nearly two decades. Google thinks AI will fundamentally change that.

In this episode of The Deep View Conversations, we talked with Sameer Samat, president of Android ecosystem at Google, about what the company means when it says it's transforming Android from an operating system into an intelligence system.

Samat explains why the next generation of computing could shift us from micromanaging our devices to simply telling them what we want to accomplish. We dig into how AI agents could navigate apps and complete multistep tasks and why those agents need to follow us across phones, computers, cars, watches and glasses. And what happens to the app-centric model that has defined smartphones for the past 15 years?

We also get into some of the practical ways this is already taking shape. Samat discusses Google’s app automations and Rambler, the new Google Keyboard experience that can turn your voice brain-dumps into polished text. He also explains how Google is thinking about permissions, sandboxing and human oversight as AI agents gain the ability to take action on our behalf.

The conversation goes well beyond the phone. We talk about why smart glasses and cars could be especially powerful interfaces for AI agents, what Google learned from the original Google Glass, and why the best AI features may be the ones consumers barely think of as AI.

Other topics covered include:
• How AI is already changing work inside Google
• Why product managers can now build functional prototypes themselves
• Samat's favorite overlooked AI tool
• His "calendar cleanse" strategy for getting time back

If you’re trying to understand where mobile computing goes next, what AI agents will actually look like on phones, and how Google plans to weave intelligence across devices, this conversation offers insights into what the company is building and why.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm

LINKS

• Databricks closes $5 billion funding round at $190 billion valuation

• AMD reportedly plans to raise $4.75 billion in debt sale 

• Microsoft begins merging consumer Copilot app with 356

• Anthropic investors reportedly in talks for $2 trillion valuation in IPO

• Taiwan says it was targeted by AI-driven "overseas" cyber attack

• OpenAI hires Wiz's Dali Rajic as revenue head, replacing Denise Dresser

• X: The social media platform is expanding its open source codebase, letting users see if they’ve been "shadowbanned." 

• Runway: The video platform now connects Figma, Dropbox and Notion to its agent. 

• GPT-5.6 Sol: OpenAI has debuted an "ultrafast" version of its most powerful model, now running up to 14 times faster. 

• Perplexity Search as Code: The AI firm released optimizations to its deep research tool to improve performance while cutting cost per task by nearly 10%.

• Broccoli AI: AI Operations Lead

• Distyl: Applied AI Researcher, System Discovery

• Ent: AI Transformation

• AGI Inc.: AI Researcher

GAMES

A QUICK POLL BEFORE YOU GO