Google swoops on failed OpenAI deal

PLUS: Moonshot AI takes open-source crown with K2 release

AuthorAuthor

Zach Mink & Rowan Cheung
July 14, 2025

Good morning, AI enthusiasts. The AI coding startup that almost sold to OpenAI for $3B just landed at Google instead, and the fallout reveals just how messy the OpenAI-Microsoft relationship has become.

With the initial Windsurf agreement collapsing over Microsoft's refusal to grant IP exceptions, Google seized a big opportunity to snag talent — and hand OpenAI another setback in what has become an increasingly frustrating summer.

In today’s AI rundown:

  • Google swoops in as OpenAI, Windsurf deal collapses

  • Moonshot AI’s K2 takes open-source crown

  • Auto-switch between models for different tasks in n8n

  • AI coding tools slow down experienced devs

  • 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

OPENAI, GOOGLE, & WINDSURF

Image source: Ideogram / The Rundown

The Rundown: Google is hiring Windsurf CEO Varun Mohan and several researchers in a $2.4B licensing deal, coming in the wake of the coding startup’s $3B acquisition deal with OpenAI collapsing over complications in its partnership with Microsoft.

The details:

  • Mohan and co-founder Douglas Chen will join Google to work on agentic coding for Gemini, with Windsurf remaining independent under an interim CEO.

  • The OpenAI agreement reportedly fell through over Windsurf’s concerns around Microsoft’s access to its tech through existing partnership agreements.

  • The $3B deal had an exclusivity deadline that expired after Microsoft refused to give OAI an exception that would allow Windsurf to avoid sharing its IP.

  • The new agreement includes multi-year comp packages and a non-exclusive tech license, allowing Windsurf to continue serving enterprise customers.

Why it matters: The Windsurf deal was set to be one of OpenAI’s biggest acquisitions to date — now, it marks yet another headache in what has been a rocky summer for the AI giant. The fallout stemming from Microsoft complications also adds fresh strain to an already frigid relationship between the former AI power couple.

TOGETHER WITH SANA

The Rundown: Sana Agents is your all-in-one AI workspace, combining agents, knowledge, and tools in one easy interface. Automate real work, generate finished outputs from any data, and keep your teams moving fast.

Use Sana Agents to:

  • Plan, create, and collaborate with agents and teammates — all in one place

  • Create finished presentations, dashboards, documents, and even custom apps, ready to export to wherever you work

  • Connect to over 100 platforms, including Salesforce, SharePoint, and Google Drive

  • Run multi-step workflows and handle complex tasks using your live business context

MOONSHOT AI

Image source: Moonshot AI

The Rundown: Chinese startup Moonshot AI released Kimi-K2, a massive 1T parameter open-weights model that matches or beats frontier-level models across a range of benchmarks, with particular strength in coding and agentic tasks.

The details:

  • K2 surpasses models like GPT-4.1 and Claude 4 Opus on coding benchmarks, also scoring new highs on math and STEM tests among non-reasoning systems.

  • The model excels at agentic workflows, with examples showcasing complex multi-step tasks like analyzing data and booking travel with extensive tool use.

  • Moonshot created a new tool called MuonClip that enabled stable training with zero crashes, potentially solving a major cost bottleneck in development.

  • K2 doesn’t have multimodal or reasoning capabilities yet, with Moonshot saying they plan to add those functionalities to Kimi in the future.

Why it matters: Moonshot’s release doesn’t have the fanfare of the “DeepSeek moment” that shook the AI world, but it might be worthy of one. K2’s benchmarks are extremely impressive for any model, let alone an open-weight one — and with its training advances, adding reasoning could eventually take Kimi to another level.

AI TRAINING

The Rundown: In this tutorial, you will learn how to use n8n’s new model selector node to set up intelligent AI model routing that automatically selects the best model for each task at hand.

Step-by-step:

  1. Create a new n8n workflow with “Chat Message” trigger and connect an “AI Agent”

  2. Choose "Model Selector" instead of a single model and configure two options (e.g., ChatGPT o3 and Gemini 2.0 Flash)

  3. Set routing rules so Model 1 triggers when the message contains the keyword “code,” for example, and Model 2 as the default for everything else

  4. Test with different messages: “Write a poem” routes to Gemini, “Python code example” routes to ChatGPT

Pro tip: Create specialized routing for different use cases, such as coding to ChatGPT, creative writing to Claude, research to Gemini, and simple questions to faster models for cost optimization.

PRESENTED BY SALESFORCE

The Rundown: Meet the Agentblazers — Salesforce’s community of builders putting AI to work in the real world. From simplifying back-office ops to unlocking new insights with enterprise data, these are the people making AI practical, accessible, and impactful.

In this conversation with Agentblazer Melissa Hill Dees, you’ll learn:

  • How Agentblazers apply AI to solve real problems

  • Why Salesforce’s Trailhead and Data Cloud platforms provide a unique edge

  • What’s driving enterprise AI adoption across sectors

METR

Image source: METR

The Rundown: AI research institute METR just published new research findings that experienced developers take longer to complete real coding tasks when using AI assistants, despite reporting feeling more productive.

The details:

  • Researchers tracked 16 veteran open-source developers completing 246 actual tasks on massive codebases averaging 22k+ stars and 1M+ lines of code.

  • The devs expected AI tools like Cursor Pro to save them 24% of their time, but testing showed they took 19% longer when AI assistance was allowed.

  • Time analysis showed devs spending less time actively coding and more time prompting, reviewing generated code, and waiting for responses from AI tools.

  • After completing the work, developers still believed AI had made them 20% faster despite the results, showing a disconnect between perception and reality.

Why it matters: These results are a bit surprising given the growing percentage of code being written by AI at major companies. But the time factor might be the wrong parameter to measure — teams should look at not whether AI makes developers faster, but whether it makes coding feel easier, even when it may take a bit longer.

QUICK HITS

  • 🔬BioEmu - Microsoft’s open AI for predicting protein states and energies

  • ⚙️ Devstral - Mistral's SOTA open model family for agentic coding

  • 🗣️ Speech in Flow - Bring images to life with speech in Google Flow

  • 💬 Qwen Chat - Alibaba’s powerful AI models, now available via desktop app

OpenAI CEO Sam Altman announced that the company is pushing back the release of its open-weight model to allow for additional safety testing.

Tesla is incorporating xAI’s Grok assistant into its vehicles, with newly purchased cars coming with a built-in integration and support via software updates for older models.

xAI released a post detailing the technical issues that led to Grok-3’s offensive posts last week, linking them to the mistaken incorporation of “deprecated instructions.”

Meta acquired voice AI startup PlayAI, with the entire team reportedly joining the company next week and reporting to former Sesame AI ML Lead Johan Schalkwyk.

Microsoft released Phi-4-mini-flash-reasoning, a 4B open model designed to run efficient advanced reasoning capabilities for on-device use cases.

X users uncovered that Grok 4 consults Elon Musk’s posts during its thinking process, with xAI pushing a system update to stop basing its answers on its creator’s remarks.

COMMUNITY

Check out our last live workshop with Dr. Alvaro Cintas, The Rundown’s AI professor. By the end of the workshop, you’ll confidently be able to install and use Gemini CLI to boost your productivity right from the command line.

Watch it here. Not a member? Join The Rundown University on a 14-day free trial.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team

Reply

or to participate.