AI Updates for May 2025: Revealing Latest Trends and Innovations

Written By
Published on
May 31, 2025
Share this
May AI updates and trends highlighted on a monthly calendar view.

How has the world of Artificial Intelligence (AI) fared in the month of May, 2025? The last few months have been nothing short of thrilling, as AI takes the reins and redefines businesses and everyday experiences.

AI is not only making incremental changes and seismic shifts in our cognitive landscape but also toppling old paradigms and reshaping how we think and operate. Over the past couple of years, Large Language Models (LLMs) have soared across diverse fields, acting as catalysts for a mental revolution.

This transformation is unfolding with lightning speed, with something new happening every day. So let’s dive into the AI updates for May 2025.

May AI Updates: Top Model Releases

Packed with practical enhancements and cutting-edge capabilities, the May AI updates introduce the top models shaping the future of AI innovation. Let’s dive into the highlights.

Anthropic’s Claude 4 Series AI Models

Anthropic has released its Claude 4 series AI models—Claude Opus 4 and Claude Sonnet 4—with new capabilities, including “Extended Thinking” with tool use. Both of these models are hybrid models with two modes:

1. Near-instant responses
2. Extending Thinking for reasoning

Additionally, Claude Code has officially hit the shelves, and you can explore its beta extensions in popular coding environments, such as VS Code and JetBrains. Developers can also access it on GitHub, making it easier to integrate top-tier AI into their workflow.

Of the two models, Opus 4, Anthropic’s flagship-tier AI model, boasts many winning points:

  • Best coding model in the world
  • State-of-the-art (SOTA) in coding, tool use, and writing
  • 72.5% score on the SWE-Bench
  • 43.2 % score on the Terminal-Bench

Similarly, Claude Sonnet 4 has undergone significant enhancements compared to its predecessor. This AI model has achieved a stellar 72.7% on the SWE-Bench (SOTA) during internal evaluations.

Snapshot of Claude 4 models' performance in the SWE bench.

Subscribers of Enterprise, Team, Claude Pro, and Max currently have access to the Opus 4 and Sonnet 4 models. If you are a developer, you can access them via:

  • Anthropic API
  • Amazon Bedrock
  • Google Cloud’s Vertex AI

And the best part is that Sonnet 4 is free for everyone who wants to give it a try!

Google’s Imagen 4 and Veo 3

Google unveiled the next generation of its image and video generation AI models, each featuring upgrades and new capabilities over its predecessors:

  • Imagen 4
  • Veo 3

Alongside these new multimodal AI models, the tech giant has also introduced a new AI-powered filmmaking app—Flow.

These models, recently unveiled at Google’s I/O 2025 event held in May, are designed to empower filmmakers and storytellers with amazing AI capabilities. Let’s explore what each model is capable of.

Imagen 4

Google has worked on the generation speed and accuracy of Imagen’s latest version—Imagen 4. This model also supports text and image inputs, like the previous generation, but the generated images now include fine details, such as intricate fabrics, shimmering water droplets, and the delicate textures of animal fur.

This powerful model is now available in Whisk, the Gemini app, and across Workspace apps, including Vids, Slides, and Docs. Enterprises can access it via Vertex AI. However, it remains to be seen if Google will extend the model only to paid subscribers or generously allow all Gemini users to get a taste of it.

Collage showcasing Imagen 4's micro-detail capability with images like a textured butterfly wing and a sharply detailed kingfisher.

Veo 3

This groundbreaking model elevates video generation to the next level by integrating native audio generation. This innovative technology allows for the creation of immersive videos incorporating ambient sounds, realistic background noise, and engaging dialogues. Veo 3 also boasts improvements in prompt adherence, spot-on lip syncing, and real-world physics.

If you are a Google AI Ultra subscriber in the US, you can access Veo 3 via the Gemini app and Flow. Enterprises can access it via the Vertex AI platform.

Flow

Flow is a blessing for filmmakers, enabling them to transform their ideas into reality by simply describing a video clip using natural language prompts. It leverages Gemini, Imagen, and Veo models to generate a video that’s eight seconds long, based on this prompt. This app has a high prompt adherence, generating consistent frames of objects, styles, cast, and locations. It is available to the Google AI Pro and Ultra plan subscribers in the US.

To know more about these new AI models from Google, check out their blog.

Mistral’s Devstral

Another AI latest news is Mistral’s Devstral, an exciting player in the world of coding. This open-source AI coding agent is designed specifically for performing software development tasks. What sets Devstral apart from other open-source software engineering (SWE) agents is its unique ability to solve real-world software engineering problems and write contextualized code in a codebase. Notably, Mistral, in partnership with All Hands AI, has crafted this AI model.

While existing open-source LLM models can perform isolated coding tasks such as writing standalone functions or code completion, they struggle with writing contextual code in a large codebase. However, Mistral believes that Devstral effortlessly overcomes this problem with its ability to contextualize coding tasks with the existing frameworks and database.

In recent internal evaluations, Devstral made a striking impression by scoring 46.8% on the SWE-Bench Verified, placing it at the pinnacle of its category. It also outperformed larger open-source models, such as Qwen3 and DeepSeek-V3, as well as proprietary models, including OpenAI’s GPT-4.1-mini and Anthropic’s Claude 3.5 Haiku.

Devstral AI coding agent leads in performance on the SWE-Bench Verified, compared to other AI models.

Devstral is available on several platforms, such as:

  • Hugging Face
  • Unsloth
  • Kaggle
  • Ollama
  • LM Studio

Academic and commercial users can access this AI model with a permissive Apache 2.0 license.

Also Read: AI Updates for April 2025: The Latest From the Tech World

Mistral’s Medium 3 Multimodal AI Model

Mistral’s Medium 3 is an enterprise-focused model that delivers supercharged performance without breaking the bank. Touted as SOTA in terms of performance in its class, this model is available at a fraction of the cost of its rivals.

Medium 3 is a multimodal LLM and can be deployed in a hybrid setup (cloud + edge) or on-premise and in Virtual Private Cloud (VPC) setups. This flexibility enables enterprises to post-train this model on their internal data to make its responses grounded. Additionally, this model can be integrated into enterprise systems and tools.

Internal tests reveal that Mistral’s Medium 3 stands out not just in theory but in practice, outperforming highly regarded competitors like Llama 4 Maverick, GPT-4o, and Claude 3.7 Sonnet on several benchmarks, such as:

  • HumanEval
  • ArenaHard
  • Math500 Instruct
  • AI2D

The scores are also on par with those of the DeepSeek 3.1 model across multiple benchmarks.

AI benchmark table: Mistral Medium 3 (highlighted) vs others. Scores on coding, knowledge tasks.

For developers and enterprises eager to jump on the Mistral Medium 3 bandwagon, it is accessible as an API on:

  • Mistral La Plateforme
  • Amazon SageMaker

Mistral will also make it available soon on:

  • IBM Watsonx
  • NVIDIA NIM
  • Azure AI Foundry
  • Google’s Vertex AI

Albeit, Medium 3 comes with a twist. It is not an open-source model and will not be accessible to public repositories like GitHub and Hugging Face.

Windsurf’s SWE-1 Series AI Models

Another exciting AI latest update comes from Windsurf. The company has unveiled the SWE-1 series of AI models—SWE-1, SWE-1 Lite, and SWE-1 Mini. These models do more than just write and edit code; they are focused on complex software engineering tasks.

Each model is designed for specific use cases, with Lite and Mini versions available to all Windsurf users. However, the frontier SWE-1 model is only available to subscribers.

Windsurf Wave 9 graphic: SWE-1 engineering models with Lite, Full, and Mini tiers.

Windsurf details its new AI models in a blog post, where the company also reiterates that despite improvements in coding models, their scope hasn’t increased significantly.

The SWE-1 frontier model is said to compete with the likes of Claude 3.5 Sonnet, boasting features like tool-calling and advanced reasoning capabilities. The SWE-1 Lite—a lightweight model—on the other hand, performs everyday coding tasks with remarkable ease and is available for unlimited use. Similarly, the SWE-1 Mini—a low-latency model—is ideal for real-time tasks. Both models are available to all users, including those on the free tier.

Other AI Breakthroughs for May 2025

Let’s explore other AI breakthroughs for May 2025 that highlight the latest advancements in the AI landscape.

DeepSeek’s Prover-V2

DeepSeek has unveiled its latest triumph—Prover-V2, an updated version of its Prover model. This is a highly specialized, open-source, mathematics-focused AI model specifically designed to tackle the intricate world of proving formal mathematical theorems. Prover-V2 can be downloaded from famous repositories such as GitHub and Hugging Face.

JetBrains’ Mellum

JetBrains takes coding to the next level with Mellum, its first “open” AI model tailored for coding. Mellum is specifically designed for code completion—completing code snippets based on the surrounding context. This cutting-edge AI model is honed on a staggering 4 trillion tokens and packs 4 billion parameters. Previously concealed within JetBrains’ software development suites, this code-generating marvel is freely available on Hugging Face.

Apple’s Matrix3D

Matrix3D, Apple’s new unified AI model, is the fruit of a collaborative effort between the company’s ML team and the prestigious Nanjing University and the Hong Kong University of Science and Technology (HKUST). This LLM can generate 3D views from multiple 2D images. It is now available for everyone to explore and experiment with by downloading it via Apple’s listing on GitHub.

Stability AI and Arm’s Stable Audio Open Small

The collaboration between Stability AI and Arm has resulted in a lightweight text-to-audio generation AI model—Stable Audio Open Small. According to their announcement, this open-source audio model is capable of generating short audio samples using text prompts. Additionally, it is said to have a fast generation time, enabling efficient handling of bulk use cases. It is available for download from GitHub and Hugging Face.

Xiaomi’s MiMo AI Models

Xiaomi enters the AI arena with a bang, unveiling MiMo, its first open-source reasoning-focused AI model. It is a text-based model and doesn’t have multimodal capabilities. With its impressive reasoning capabilities in a relatively smaller parameter size, MiMo is on par with reasoning models such as DeepSeekR1, Alibaba’s Qwen QwQ-32B, OpenAI’s o1, and Google’s Gemini 2.0 Flash Thinking. For users eager to explore the world of reasoning with Mimo, it can be downloaded from Xiaomi’s listing on GitHub and Hugging Face.

Microsoft’s Phi-4-Reasoning-Plus

Microsoft Research has announced the latest release—Phi-4-Reasoning-Plus—an open-weight language model built for tasks that require deep, structured reasoning. Based on the architecture of the previously released Phi-4, this new model integrates supervised fine-tuning and reinforcement learning. It is engineered to tackle complex challenges across mathematics, science, coding, and logic-based tasks with exceptional performance.

May AI Updates: Notable Feature Updates You Cannot Miss

Now, let’s explore the standout features and innovations introduced this May that are driving AI forward.

Gemini 2.5 AI Models: Deep Think Mode and Native Audio Output

Google leveraged the I/O 2025 to showcase two brand-new features for the Gemini 2.5 family:

  • Deep Think: An enhanced reasoning mode powered by the Gemini 2.5 Pro model
  • Native Audio Output: A new, natural, and human-like speech available via the Live API

OpenAI’s GPT-4.1 AI Models: Turbocharged Coding Capabilities

OpenAI has announced the expansion of its recently released GPT-4.1 series of AI models to ChatGPT. The company is making these advanced models now available within the chatbot, which were so far exclusive to the company’s API. This expansion is expected to significantly enhance the chatbot’s capabilities and user experience.

Gemini 2.5 Pro: Improved Coding Capabilities

Google has rolled out an impressive update to its Gemini 2.5 Pro AI model, rebranding it as the Gemini 2.5 Pro Preview (I/O edition). This upgraded version boasts enhanced coding capabilities and is capable of creating interactive web applications from scratch. It can also perform other complex coding tasks such as code transformation, code editing, and developing complex agentic workflows. In a blog post, the tech giant detailed the updated LLM.

Google Gemini: Multiple File Uploads With a Single Prompt

In another effort to enhance Gemini, Google is also introducing a high-quality-of-life enhancement to the AI model. The company rolled out an update that now allows users to upload up to 10 images or files in a single go. This multi-file uploading capability is supported by all the new general-purpose Gemini AI models. The new feature is available on the Gemini web client, as well as Android and iOS apps.

OpenAI ChatGPT Search: Shopping Feature and Support for Multiple Citations

OpenAI has announced a suite of new features for ChatGPT’s Search functionality, introducing an experimental shopping experience showcasing products with a more engaging and visually appealing browsing. With direct purchase links now available, users can easily transition from searching to shopping. Furthermore, OpenAI has added utility features such as:

  • Multiple citations for the same information
  • Search capability for the WhatsApp bot
  • Trending searches and autocomplete suggestions

ChatGPT Deep Research Tool: GitHub Connector

OpenAI is launching an innovative “connector” for ChatGPT deep research, a powerful tool designed to search across the web and other sources to compile and generate thorough research reports. In its beta phase, this connector can now link to GitHub, enabling developers to ask queries about a codebase and engineering documents. This feature will be accessible to ChatGPT Plus, Pro, and Team users.

Google Gemini: GitHub Integration and Coding Feature Access

Google has just unveiled a game-changing feature for Gemini, which allows the AI chatbot to seamlessly connect with GitHub repositories, transforming it into a powerful coding ally, as announced in a post by the official Gemini app handle. The features include:

  • Generating and modifying functions
  • Clarifying complex codes
  • Answering queries regarding the codebase
  • Debugging codes

However, this advanced capability is currently exclusive to paid subscribers of the platform.

Gemini App: Improved Personalization Features and Native Image Editing

Google is introducing a range of personalization features to enhance the user experience. Gemini is set to boast:

  • Improved memory and the ability to better understand and anticipate users’ needs by drawing information from their Google accounts
  • Native image editing feature powered by AI to enable users to make precise edits to their images directly within the chatbot

May AI Updates: Important Tools and AI Agents Introduced This Month

Here’s a look at the AI latest updates featuring groundbreaking tools and AI agents launched this May, driving innovation.

Google’s Project Mariner

Google’s Project Mariner is the company’s cutting-edge experimental AI agent designed to navigate and use websites with unparalleled agility. This tool competes with its contemporaries, such as OpenAI’s Operator, Amazon’s Nova Act, and Anthropic’s Computer Use. US subscribers to Google’s AI Ultra plan will have access to Project Mariner.

Google’s AI Agents

Google has unleashed its AI mode to all US Search users, allowing them to engage in dynamic conversation with an AI agent for:

  • Web page summarization
  • Easier online navigation
  • Shopping assistance

Google is also ramping up personalization with its Deep Research agent to craft in-depth research reports by tapping into relevant websites, featuring seamless integration with Drive and Gmail. Plus, Project Astra, Google’s cutting-edge real-time multimodal AI, is being more deeply integrated into Search and Gemini to enhance both verbal and visual interactions. The future of search is here!

Hugging Face’s Free AI Agent

Hugging Face has rolled out the Open Computer Agent—a demo of an AI agent that can complete a variety of web-based tasks. This free tool, available on the website, allows users to access websites such as Google Maps, Google Search, and even ticket booking platforms to complete tasks autonomously.

LinkedIn’s AI-Powered Job Search Tool

LinkedIn has enhanced the job search experience with its new AI feature. This feature enables professionals to input their desired job/role as a natural language search query, and it will find and show relevant roles. Launched in the US earlier this month, this AI job search tool is now expanding its access worldwide. However, this innovative feature is exclusive to LinkedIn Premium subscribers.

Figma’s New AI-Powered Tools

Figma launched its latest suite of AI tools at its annual Figma Config 2025 event. The new AI tools include:

  • Figma Sites
  • Figma Make
  • Figma Buzz
  • Figma Draw

These new tools are set to supercharge website building and prototype creation for paid “full seat” subscribers.

Hot AI Innovations to Keep an Eye on Next Month

Here’s a glimpse of the most exciting AI innovations and breakthroughs set to make waves in the tech world next month.

Amazon Web Services is Reportedly Working on an AI-Powered Coding Agent

Amazon Web Services (AWS) is reportedly working on a groundbreaking AI tool; the project is internally called “Kiro.” This tool can:

  • Generate code in real-time
  • Tap into AI agents to complete complex coding tasks

This tool is said to be a comprehensive system capable of:

  • Writing code
  • Designing user interfaces (UIs)
  • Pinpointing glitches and bugs
  • Optimizing existing code

This ambitious project is scheduled to launch in June.

Meta’s “Behemoth” AI Model Faces Unforeseen Hurdles

Meta Platforms has opted to delay the much-anticipated release of its flagship AI model, “Behemoth.” Originally slated for an April rollout, the team encountered challenges in significantly enhancing its capabilities, leading to mounting questions within the company about whether the improvements over previous versions would truly impress. Now, the target for its launch has shifted to June. While delays often sound like setbacks, they can signal a commitment to excellence.

Entering an Era of AI Revolution

The recent artificial intelligence news highlights one key point: May has been remarkable for advancements in AI, showcasing everything from extraordinary multimodal breakthroughs to cutting-edge open-source coding agents.

As we stand on the cusp of a new era, tools like Maayu.ai highlight a significant transformation toward intelligent systems that not only respond to our queries but also proactively anticipate our needs. It is not just about technological progress; it is about creating a future that feels more intuitive, adaptive, and remarkably human-centered. This season promises to be an exhilarating ride through the fast-paced world of AI innovation! Stay tuned for the latest updates; the game is about to change!