The 5 Biggest AI Stories of 2025 Explained

Recapping the biggest shifts that will shape 2026 as well.

Welcome to The Median, DataCamp’s newsletter for December 19, 2025.

In this edition: OpenAI releases GPT Image 1.5, Google introduces Gemini 3 Flash, Meta launches SAM Audio for sound separation, NVIDIA introduces the Nemotron 3 open model family, and Swedish startup Lovable raises $330 million. We also present five pivotal moments that reshaped the AI industry in 2025.

---

Important note: Our newsletter will be taking a two-week break. We’ll be back in your inbox on Friday, January 9.

---

This Week in 60 Seconds

OpenAI Launches GPT Image 1.5 for Precise Editing

OpenAI has officially released GPT Image 1.5, a major upgrade to the image-generation capabilities in ChatGPT and its API. The new model generates visuals up to 4x faster than previous versions. It introduces precise editing that allows users to add, subtract, or blend elements while maintaining consistent lighting, composition, and subject appearance across subsequent edits. Additionally, the model significantly improves text rendering, handling denser and smaller text with higher fidelity. The release of GPT Image 1.5 is supported by a new dedicated “Images” workspace in the ChatGPT sidebar. For developers, image inputs and outputs are now 20% cheaper compared to the previous model.

Google Introduces Gemini 3 Flash for Fast Reasoning

Google has expanded its latest model family with the release of Gemini 3 Flash, designed to provide frontier-level intelligence with significantly lower latency and cost. Gemini 3 Flash combines the advanced reasoning of Gemini 3 Pro with high-speed efficiency, making it Google’s most capable model for agentic workflows and responsive interactive applications. It delivers standout performance on complex benchmarks, rivaling larger models on PhD-level reasoning while outperforming previous iterations like Gemini 2.5 Pro. Gemini 3 Flash is now the default model in the Gemini app and is rolling out globally to AI Mode in Search, Google AI Studio, and Vertex AI.

Meta Releases First-of-Its-Kind Model for Multimodal Sound Separation

Meta has released SAM Audio, a first-of-its-kind model that makes isolating specific sounds from complex audio mixtures as easy as clicking an object or typing a description. Unlike traditional tools designed for single-purpose tasks, this unified model uses multimodal prompts to identify and extract sound. For example, a user can isolate a single guitar from a full band performance just by clicking on the instrument in a video, or filter out loud traffic noise from an outdoor recording by typing a text prompt. The model also introduces span prompting, an industry first that allows creators to fix audio issues (like a barking dog) across an entire podcast recording by simply marking the relevant time segments. SAM Audio is now available for exploration in the Segment Anything Playground.

NVIDIA Launches Nemotron 3 Family of Open Models

NVIDIA has introduced Nemotron 3, a new family of open models. The lineup uses a hybrid mixture-of-experts (MoE) architecture and comprises three sizes: Nano (30B parameters/3B active), Super (100B/10B active), and Ultra (500B/50B active). The Nano model is available now and delivers 4x higher token throughput than the previous generation, while its 1-million-token context window allows agents to maintain coherence across complex, multi-document tasks. Alongside the weights, NVIDIA has open-sourced three trillion tokens of training data and its NeMo Gym reinforcement learning libraries.

Swedish Startup Lovable Raises $330M at a $6.6B Valuation

Stockholm-based Lovable has raised $330 million in a Series B round led by CapitalG and Menlo Ventures, catapulting the company to a $6.6 billion valuation. Lovable provides a vibe coding tool that empowers non-technical users to create functional apps using natural language, helping the company surpass $200M in ARR in its first year. With 100,000+ new projects created daily for customers like Uber and Zendesk, the new capital will fund deeper integrations with tools like Notion and Jira, alongside enhanced enterprise governance and production-grade hosting infrastructure.

---

New Course: AI for Finance

---

A Deeper Look at This Week’s News

2025 in Review: 5 Pivotal Moments That Reshaped AI in 2025

As we close the year, we’re looking back at the five high-impact moves that reorganized the global economy and fundamentally influenced the trajectory of the AI industry.

Think we missed something? Leave a comment and let us know which event you would choose instead!

1. DeepSeek and the efficiency shock

What happened

In January 2025, the Chinese research lab DeepSeek released DeepSeek-R1, an open-weight large language model that significantly challenged existing assumptions regarding AI training costs.

For several years, the industry operated under the belief that frontier-level reasoning required massive capital investment and the most advanced hardware.

DeepSeek-R1, however, demonstrated reasoning capabilities comparable to OpenAI’s o1 and GPT-4 series for a reported cost of under $6 million (a small fraction of the estimated $100 million required for its peers).

Impact

The release of R1 led to a notable reassessment of the “scale is all you need” dogma within the AI sector.

This sudden realization that high-level intelligence could be achieved with significantly less compute power contributed to the “DeepSeek crash” on January 27, 2025. During that single trading session, Nvidia lost nearly $600 billion in market capitalization as investors re-evaluated the projected demand for next-generation hardware.

Geopolitically, the event highlighted the limitations of export controls, showing that algorithmic efficiency can help developers navigate hardware bottlenecks. DeepSeek-R1 achieved its performance using older, export-compliant Nvidia H800 chips by using architectural innovations like Mixture-of-Experts (MoE) and a reinforcement learning pipeline focused on chain-of-thought distillation.

Looking ahead

By releasing R1 under an open-source MIT license, DeepSeek shifted the industry toward the commoditization of reasoning weights. This move has forced a reality check for closed-source providers, encouraging a pivot from brute-force hardware scaling toward software optimization and architectural efficiency.

As we look toward 2026, the focus has moved from simply building larger models to delivering more hardware-efficient and accessible intelligence for specialized, real-world tasks.

2. ​​The Rise of Autonomous Agents

What happened

While chatbots defined the previous two years, 2025 was the year of AI agents: systems that use language models to achieve user-defined goals by reasoning, planning, and executing actions through external tools.

Source: AI Agents Cheat Sheet

The industry saw the emergence of several distinct agent categories, including one-prompt agents for simple tasks, coding agents for software development, workflow-based agents for business processes, and complex agentic frameworks.

This growth was supported by the adoption of standardized protocols, such as Anthropic’s Model Context Protocol (MCP), which standardizes how agents access external data, and Google’s Agent2Agent (A2A) protocol, which enables different agents to communicate and collaborate on complex tasks.

You can learn more about this topic through our AI Agents Fundamentals track.

Impact

Because agents are uniquely capable of handling tasks where the path to a solution isn’t a straight line, they have reshaped automation workflows across multiple sectors.

A large-scale field study released by Perplexity and Harvard University this year highlighted several key impacts of this shift:

• Targeted adoption: Adoption is currently highest in knowledge-intensive fields such as digital technology, academia, finance, marketing, and entrepreneurship.

• Sector stickiness: While tech professionals drive the most volume, those in marketing, sales, and management show the highest “stickiness,” meaning their intensity of use grows rapidly once they adopt an agent.

• User evolution: The study tracked a clear evolution in user behavior, where individuals typically start with low-stakes queries, like travel planning, before migrating toward complex, multi-step productivity tasks that significantly improve retention.

• Personal utility: Interestingly, personal use still accounts for over half of all agent queries, even as professional and educational applications continue to scale.

Looking ahead

As we enter 2026, we expect to see the emergence of even more complex agentic behaviors, moving beyond digital tasks into deeper scientific and industrial applications.

We also anticipate an aggressive development of the agent economy, a concept we explored in a previous issue, where agent shops and machine-to-machine commerce become standard. As the friction for agent-to-agent interaction disappears, the global economy will likely move toward a model of automated commerce, where agents autonomously negotiate and execute transactions on behalf of their human users.

3. The maturity of generative media

What happened

The transition of generative media from a creative novelty to a production-grade tool accelerated on May 20, with the introduction of Google Veo 3, which marked the first time AI video clips could feature native, synchronized audio. This trend was followed by OpenAI’s Sora 2 and Kling 2.6, both of which introduced their own native audio capabilities.

In the image generation space, the field was initially shaken by GPT-Image-1, which went viral for its Ghibli-style aesthetic. However, the release of Google’s Nano Banana later in the year shifted the focus toward precise, prompt-based editing.

By the end of 2025, the arrival of Nano Banana Pro and GPT-Image-1.5 appeared to fully resolve the industry’s long-standing text-rendering challenge, enabling models to generate dense, small text with high accuracy.

Impact

This technological maturity has allowed small businesses to produce short ads independently, while industry giants like Coca-Cola and McDonald’s have also adopted these tools for high-profile campaigns.

High-quality image generation has also reshaped marketing workflows, becoming a standard tool for social media content, large-scale campaigns, and content marketing.

As we noted in our coverage of an earlier MIT report this year, while the narrative of widespread industrial disruption is still largely a myth, the tech and media sectors are proving to be the notable exceptions, experiencing significant shifts in how content is produced and scaled.

Looking ahead

As we enter 2026, the industry is pivoting from pure generation toward video-to-video transformations. The December release of Kling O1, the first unified multimodal video model, allows creators to restyle existing footage using text prompts while maintaining the original motion structure.

In the coming year, we expect to see even greater realism, longer clip lengths, and shorter generation times. As these tools become faster and more cost-effective, we anticipate more businesses will integrate AI media directly into their creative pipelines.

4. The EU AI Act

What happened

On February 2, 2025, the legal landscape for artificial intelligence shifted as the “Prohibitions” chapter of the EU AI Act officially entered into force. This milestone moved the Act from a period of guidance into active law enforcement, specifically targeting AI practices deemed to carry an “unacceptable risk” to fundamental rights.

These regulations introduced strict bans on social scoring systems, real-time remote biometric identification in public spaces by law enforcement, and the use of AI for subliminal manipulation or the untargeted scraping of facial images from the internet.

You can learn more about this topic in our course on Understanding the EU AI Act.

Impact

The enforcement of these rules required a significant adjustment for global technology providers, leading to what many call a “Splinternet” of AI experiences. This term describes a fracture in the global digital experience where the features of a single AI platform vary depending on the user’s region and local laws.

For example, an AI-powered hiring platform operating outside the EU might analyze a candidate’s video interview to infer “confidence” or “enthusiasm” from their facial expressions and tone of voice. In the EU, however, this specific use of biometric-based emotion recognition is prohibited in a recruitment context to protect candidates from unproven or intrusive inferences.

While this creates a complex compliance environment for developers, it has also established Europe as a global leader in ethical AI. The EU’s efforts are already serving as a template for other nations, such as Japan and Canada, which are looking to balance the drive for innovation with better protection for their citizens.

Looking ahead

As we look toward 2026, we expect to see an even larger industry-wide focus on the delicate balance between fostering high-speed innovation and implementing the restrictions necessary to protect the public.

The ongoing challenge for policymakers and tech leaders alike will be to ensure that safety-first models do not stifle creativity, but rather provide a stable environment where trust can drive long-term adoption.

5. The Stargate Project

What happened

Announced on January 21, 2025, the Stargate Project is an unprecedented $500 billion joint venture aimed at establishing a massive AI infrastructure network within the United States. The initiative is led by a consortium including OpenAI, SoftBank, Oracle, and the UAE-backed MGX.

The project centers on building a distributed supercomputing network, including a 5-gigawatt data center campus—a facility designed to consume more power than the city of Seattle.

Impact

The project signaled a shift in AI strategy from private enterprise toward state-backed industrial policy, effectively designating compute power as a strategic national resource. The project also reflects a broader trend of AI Nationalism, where infrastructure is viewed as a theater of geopolitical conflict and essential for global hegemony.

As we explained in a previous edition of The Median, anchoring compute capacity within U.S. territory allows the country to achieve resilience of both location and supply.

By localizing these assets, the U.S. reduces its strategic exposure to Taiwan, where the majority of Nvidia’s advanced chips are manufactured, granting resilience of location.

However, the project’s long-term success also depends on diversifying the hardware market to reach resilience of supply. Currently, the ecosystem remains overwhelmingly dependent on a single provider; if Nvidia stumbles (whether through supply chain disruptions or manufacturing issues), the entire U.S. infrastructure feels it immediately.

Looking ahead

As we head into 2026, the focus is on the physical execution of this multi-year infrastructure plan and its integration with the national power grid. While the industry continues to innovate in software efficiency, the Stargate infrastructure is intended to provide the brute force volume required for the next generation of frontier models.

---

Industry Use Cases

AI-Driven Adtech To Play a Big Role in the 2026 World Cup

Sportradar has developed AI-driven adtech to play a key role in the 2026 World Cup. While major tournaments have historically been the exclusive domain of top-tier partners with massive budgets, challenger brands often struggle to make a striking impact. This new platform addresses the gap by using real-time sports data and dynamic creative optimization to automatically tailor advertising to live scores and on-pitch milestones. By delivering contextually relevant messages across all digital channels within 48 hours of key events, the system enables smaller brands to connect with fans at critical moments.

AI-Driven Fleet Management Secures Underwater Infrastructure

Skana Robotics has developed a breakthrough for its SeaSphere software, allowing groups of unmanned vessels to communicate underwater using AI. Underwater submersibles have historically struggled to communicate over large distances without surfacing, a move that creates significant exposure risks. New fleet management software addresses this by using AI to facilitate long-distance data sharing while vessels remain submerged. By processing peer data, units can autonomously adapt their maneuvers to meet a collective mission without rising to the surface. This coordination is currently being applied to protect critical underwater infrastructure and global supply chains.

Robot Bartender ADAM Serves Vegas Sports Fans

Fans at Las Vegas’s T-Mobile Arena are now being served by ADAM, a robotic bartender trained in high-fidelity simulations to handle the chaotic lighting and reflections of a crowded stadium. Using real-time edge processing, ADAM can detect misplaced cups and measure liquid levels with sub-40-millisecond latency. This allows it to correct pours mid-motion, ensuring delivery remains precise during the peak rush of a game. This deployment mirrors a broader shift toward industrial dexterity, where autonomous systems move from factory floors into dynamic public spaces.

Tokens of Wisdom

Within three to five years, world models will be the dominant model for AI architectures, and nobody in their right mind would use LLMs of the type that we have today.

—Yann LeCun, AI Researcher