The Rundown AI / Articles / AI / DeepSeek strikes again

DeepSeek strikes again

PLUS: Patients control AI and robotics with thought

Zach Mink

December 2, 2025

Good morning, AI enthusiasts. Chinese startup DeepSeek shook the AI world with its R1 release earlier this year, rattling markets and triggering talks of tighter U.S. chip controls.

Now the Whale is back with a big encore, dropping two new models that rival GPT-5 and Gemini 3 Pro, open-sourcing them, and pricing them at a fraction of the cost. Suddenly, near-frontier AI has never looked cheaper.

In today’s AI rundown:

DeepSeek’s new models rivaling GPT-5, Gemini-3 Pro
Runway tops video leaderboard with 4.5 release
Create on-brand marketing campaigns with Pomelli
Kling’s all-in-one video model for generation, editing
4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

DEEPSEEK

🐳 DeepSeek’s new models rivaling GPT-5, Gemini-3 Pro

Image source: DeepSeek

The Rundown: Chinese AI startup DeepSeek just released V3.2 and V3.2-Speciale, two reasoning models that perform on par with SOTA models like GPT-5 and Gemini 3 Pro — while cutting costs and staying accessible under an open-source license.

The details:

V3.2 matches or nears GPT-5, 4.5 Sonnet, and Gemini 3 Pro on math, tool use, and coding tests, with the heavier Speciale surpassing them in several areas.
The Speciale variant hit gold-medal scores at the 2025 International Math Olympiad and Informatics Olympiad, also placing No. 10 overall at IOI.
V3.2 pricing comes in at $0.28 input / $0.42 output per 1M tokens, a fraction of Gemini 3 Pro ($2 / $12), GPT-5.1 ($1.25 / $10), and Sonnet 4.5 ($3 / $15).
Both 685B parameter models ship under an MIT license, with weights available on Hugging Face for anyone to download.

Why it matters: DeepSeek’s R1 release rattled markets and sparked U.S. chip export control talk, and the V3.2 follow-up shows the Chinese lab isn’t a one-hit wonder — open-sourcing a model with frontier performance at a massive price cut. For U.S. labs charging premium API fees, the pressure to justify that gap just got a lot more intense.

TOGETHER WITH SONAR

💻 Coding personalities that redefine reliability & risk

The Rundown Sonar just published The Coding Personalities of Leading LLMs, analyzing 4,400 Java tasks completed by GPT-5, Claude Sonnet 4, Llama 3.2, and other leading LLMs. The key finding? Models that excel at benchmarks often fail at writing secure, maintainable code in real-world applications.

Here’s what stood out:

LLMs excel at syntax but struggle with security
Generated code often lacks maintainability and clarity
Newer models come with harder-to-detect, more nuanced flaws

Download the full report — no form-fill required.

RUNWAY

🎥 Runway tops video leaderboard with new 4.5 release

Image source: Runway

The Rundown: Runway just released Gen-4.5, a new AI video model that claims to usher in a “new frontier for video generation”, topping benchmarks and showing strong performance across realism, motion, and creative control.

The details:

4.5 moves to the top spot of Artificial Analysis’ Text-to-Video leaderboard, after gaining hype under the codename “Whisper Thunder” in testing.
Runway says Gen-4.5 handles physics, fluid dynamics, and human movement more naturally, with details like hair and fabric staying consistent across frames.
4.5 can handle a range of styles, but excels in cinematic/realism visuals — with Runway saying outputs are “indistinguishable from real-world footage”.
The model was also codenamed ‘David’, with co-founder Cristobal Valenzuela comparing the small company’s ranking to a ‘David vs. Goliath’ victory.

Why it matters: Runway has already pushed AI into professional creative workflows, and 4.5 feels the closest yet to the cinematic capabilities needed to be widely adopted across Hollywood. While the next frontier is longer generations and even better audio/speaking sync, the year-over-year improvement in AI video is mind-blowing.

AI TRAINING

🔥 Create on-brand marketing campaigns with Pomelli

The Rundown: In this tutorial, you will learn how to use Pomelli (by Google Labs and DeepMind) to automatically create your business identity and generate on-brand marketing campaigns with ready-to-use creatives.

Step-by-step:

Go to Pomelli, click "Let's Get Started", enter your website URL - Pomelli scans and extracts colors, fonts, taglines, tone, and product cues
Edit your "Business DNA" summary by adjusting colors, values, and copy to match your brand vision - all fields are customizable
Prompt the campaign generator: "Create a scary but kind Halloween campaign", review three variations like "Give them something good" or "Treat yourself, no tricks"
In the creative editor, change headlines, text, fonts, colors, resize for different placements, use "Fix layout" to auto-reflow elements, and add a call-to-action

Pro tip: Be sure to test the conversion rate of Pomelli-generated creatives vs. your previous creatives.

PRESENTED BY IBM

🏆 Start small, then scale for AI success

The Rundown: Smarter AI starts with AI-ready data. When your models run on trusted, high-quality data, they perform more reliably, scale more smoothly, and deliver measurable value across the organization. Start small, validate impact quickly, and expand on a foundation built for long-term success.

Begin with AI that leverages:

Trusted, high-quality data
Both structured and unstructured data
Seamless integration across sources

Watch the webinar: ‘AI Agents Run on Data—Is Yours Ready?’ and learn more.

KLING

📽️ Klings all-in-one video model for generation, editing

Image source: Kling AI

The Rundown: Chinese startup Kuaishou launched Kling O1, a new AI video system that handles both video creation and editing in a single model — letting users generate clips, swap characters, make granular edits, and restyle footage in a single interface.

The details:

O1 accepts up to seven inputs at once, capable of interpreting images, videos, subjects, and text — with outputs of 3-10 seconds.
Users can edit existing footage with text commands like "remove bystanders" or "shift to nighttime" while preserving characters and scenes.
Other features include image, element, action, camera movement, and video references, start and end frames, multi-subject capabilities, and more.
Kling's internal tests show the model winning against Google Veo 3.1 and Runway’s Aleph on video reference and editing tasks.

Why it matters: Between Runway and Kling, December is kicking off with some massive AI video upgrades. O1’s all-in-one and edit-anything capabilities (similar to Runway’s previous Aleph drop) are making granular edits to video possible like never before — a leap much like what Nano Banana brought to images earlier this year.

QUICK HITS

🛠️ Trending AI Tools

🎥 Runway Gen-4.5 - Runway’s new top-rated video model
🐳 DeepSeek V3.2 - DeepSeek’s latest powerful open-source release
🎬 Kling O1 - Video model with multimodal understanding and editing
🧠 DeepSeek V3.2 Speciale - Open-source deep reasoning model

📰 Everything else in AI today

Black Forest Labs announced a new $300M funding round at a $3.25B valuation, coming on the heels of the company’s Flux.2 image model release.

Accenture and OpenAI are partnering to provide ChatGPT Enterprise to tens of thousands of consultants, also launching a program to help clients deploy AI agents.

OpenAI is taking an ownership stake in Thrive Holdings, a firm owned by one of its investors, Thrive Capital, saying the deal will scale impact across enterprise operations.

Nvidia invested $2B in chip design software maker Synopsys, with the multi-year partnership aimed at using AI and computing to speed up product engineering.

Epic CEO Tim Sweeney is lobbying for game marketplaces like Steam to stop using ‘Made with AI’ tags, saying the tech will be “involved in nearly all future production.”

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Taya P. in Vancouver, BC:

"I made a College Compass this weekend on Base44 for my son. It's an AI-powered guide to navigating the college application process, managing university research, applications, tasks, and deadlines all in one place. It auto-populates scholarship fit based on his specific filters, maps his entire 5-year journey by grade, and gets richer over time. It adjusts based on his performance, making suggestions on the area of study specialization and best best-matched universities.

It even has an admission probability calculator and makes suggestions on how to improve his success. Best of all, it empowers my son with a strategic framework that supports HIS needs and allows me as his mother to get out of the way, knowing he's got every support he needs!"

How do you use AI? Tell us here.