Artificial Intelligence Insights

Research — May 22, 2025 Generative AI Digest: Robust funding and a focus on rich media generation A significant new funding round from Safe Superintelligence has set the stage for a robust April in generative AI funding. From April 1–25, AI startups secured over $3.7 billion in funding across 127 transactions. This report, along with its accompanying funding table, highlights some of the standout rounds. It also covers a range of notable new product releases, research updates and regulatory developments from the past few weeks. OpenAI LLC has introduced a series of unusually named model releases, with version 4.1 following 4o, but announced alongside the phaseout of GPT 4.5. Meanwhile, Meta Platforms Inc. has faced accusations of manipulating benchmark scores. These announcements might tempt us to continue last month's focus on nomenclature and benchmark performance, but there is another thematic overlap in these releases — with Meta's multimodal direction and OpenAI's introduction of native image generation for the 4o model — rich media generation. These announcements are far from the only updates in image or video generation over the past few weeks. A major emphasis has been on not just higher-quality outputs, but also equipping users with stronger toolsets to tailor generated outputs. Contributions include Canva Inc.'s announcements at Canva Create, ByteDance's SeedEdit and a recent array of AI capabilities from Adobe Inc. A clear focus in the space is the growing importance of AI agents and, relatedly, multi-step AI-enabled workflows. Expectations are shifting from merely generating generic images or videos to more tailorable outputs, achieved not just by better prompt adherence but by leveraging enterprise data, or AI editing tools, for example. Product releases and updates Google LLC announced the first of its Gemini 2.5 models, an experimental version of 2.5 Pro, which it followed up with the lower-cost and faster 2.5 Flash. As expected with the "Pro" designator, the model is pitched as being valuable for complex tasks, with Google presenting an array of benchmark-leading performances across science, math and reasoning tasks. Google's Cloud Next 2025 event contained a wide array of AI announcements, but an interesting research project the company announced in April — Agent2Agent Protocol — stood out. The new open protocol works alongside Anthropic PBC's Model Context Protocol, to allow agents using MCP to communicate, improving the interoperability of agent ecosystems. ChatGPT got an image-generation upgrade, with the ability to natively create and modify photos. This announcement came ahead of new reasoning models OpenAI o3 and o4-mini. The mini models are designed for faster, more cost-effective reasoning tasks, and OpenAI suggests that o4-mini outperforms o3-mini in areas like data science and tool-use. GPT 4.1 was also announced as an upgrade on the non-reasoning GPT-4o series, with different size options in GPT‑4.1, GPT‑4.1 mini and GPT‑4.1 nano. The 4.1 model series has come with a 1 million token context window, with OpenAI addressing criticism of its historically limited context windows, which reduced the amount of information that could be provided to its models. Microsoft Corp. announced a research project — BitNet b1.58 2B4T — an open-source compressed language model that can be run locally using chips with significantly less memory than models of comparable benchmark performance. Microsoft claims to be able to demonstrate how it can reduce memory footprints with its model-quantization approach without degrading performance. This may hold significant opportunity for deploying models on-device. Meta announced the long-awaited Llama 4 to mixed reception. A number of multimodal mixture-of-experts models were announced over three parameter sizes. The largest, Llama 4 Behemoth, weighs in at 2 trillion total parameters, with 16 experts, although only a portion of these parameters (288 billion) is in active use at any time, depending on the task. Llama 4 Maverick (400 billion total, 17 billion active parameters) and Llama 4 Scout (109 billion total, 17 billion active parameters) are now available. However, the reception of the model was damaged by a controversy around a user preference table known as LMArena, where the model made available for users to test was a chat-optimized customized model rather than the publicly available Llama 4 Maverick. The vanilla variant scores poorly in the same rankings. Krisp Technologies Inc. launched AI Accent Conversion, supplementing its existing tools for noise cancellation, transcription and summarization of live meetings. The service works with popular conferencing platforms and preserves the speaker's original voice but converts the phonemes to match American English pronunciation. Initially supporting Indian English accents, the business suggests it will expand to other input accents. Canva AI, a conversational AI assistant, was announced at Canva Create with the intent to allow users to use prompts to create and iterate around concepts. Canva Code relatedly allows a user to build interactive features into a design, instructing the model to build a calculator into a website design, or a timer, or even a game, for example. These announcements came alongside upgrades to AI-enabled charting tools and photo editors. Runway AI Inc.'s Gen-4 video-generation model was announced, with an emphasis on greater realism and consistency of output. The ability to generate consistent assets across frames of a clip, or to ensure that scenes adequately reflect desired lighting, camera angles and even physics, appears to be the main driving force for the release. Beijing Byte Dance Telecommunications Co. Ltd., the Chinese internet behemoth best known in the West for its TikTok application, has delivered a wide array of AI capabilities in recent months. Announcing a new reasoning language model in April — Seed-Thinking-v1.5 — with 200 billion total and 20 billion activated parameters, ByteDance also showcased image tools Seedream 3.0 and editing tool SeedEdit. Seedream 3.0 is supposedly able to output images at a 2K resolution. SeedEdit can be used to "transform" images over multiple steps using text prompts, engaging in tasks like filter generation, style transfer, background removal and color harmonization. The last few weeks have also seen several AI-rich media and creative tools announce sizable rounds, including Runway, Krea and Capsule.Video Inc. Anthropic joined the "research game" with a new tool designed to help Claude be better at retrieval and more complex research tasks. The company also announced integration into the Google Workspace suite as an aligned component in the company's stated mission to turn Claude into a collaborative partner for professional and personal work. Adobe announced a major spread of AI updates across its Creative Cloud portfolio at the Adobe Max London Creative Conference. These included upgrades to Firefly — bringing the new Firefly Image Model 4 and the first Firefly Video Model into the app, which can now be augmented by third-party models from Google Cloud and OpenAI, among others — and Firefly Boards, which was released in public beta. Firefly Boards is a rebrand of Project Concept, a collaboration space where users can generate, collate and adjust visual assets. Funding and M&A Security automation vendor Torq Technologies Ltd. picked up Revrod Ltd., an Israel-based AI vendor founded in 2023. Revrod, which was still in stealth mode, offers advanced AI for security operations technology, particularly focused on multi-agent retrieval-augmented generation technology, agentic architectures and workflow models. Safe Superintelligence reportedly secured $2 billion in its latest round of funding, led by new investor Greenoaks Capital Partners. The startup (co-founded by OpenAI's former chief scientist, Ilya Sutskever; former Apple Inc. head of AI Daniel Gross; and investor Daniel Levy) has a single objective — delivering on its eponymous mission — and little is known about the company besides its mission statement. This transaction also saw participation from new investors Lightspeed Ventures, Alphabet Inc. and NVIDIA Corp., as well as returning investors DST Global and Andreessen Horowitz. The funding round has resulted in a post-money valuation of $32 billion for the company. To date, Safe Superintelligence has raised a total of $3 billion in funding — a remarkable amount of funding for a company with no product. Runway, an image and video generator, received $307.4 million in funding in April. The round was led by General Atlantic Service Company, with participation from Fidelity Management & Research Company, Baillie Gifford & Co., SoftBank Vision Fund II-2 L.P., NVIDIA and other investors. The company intends to use the proceeds for AI research, hiring, and expanding its Runway Studios film and animation production offerings. AI video editor Capsule announced a $12 million series A. The round was led by Innovation Endeavors, with participation from HubSpot Ventures, Swift Ventures, Human Ventures, Bloomberg Beta, and a variety of angel investors. Capsule's strategy is to empower businesses to create on-brand videos without having to rely on agencies. Creative AI tool Krea has raised $83 million across several rounds, the latest a series B led by Bain Capital Ventures. The company's investors have included Andreessen Horowitz, AVF Management, Pebblebed and Abstract Ventures. Chatbot Arena, an AI benchmarking platform that relies on users to assess model responses, is forming a new company known as Arena Intelligence. The new business suggests that this transition would make it easier to improve the platform. Korean robotics startup RLWRLD emerged from stealth in April with $15 million of seed funding. An array of investors has been cited, including Hashed, Global Brain Corp., LG Electronics Inc. and SK Telecom Co. Ltd. The physical AI startup appears to be focusing on robotics foundation models, using data from Asian manufacturing businesses. Mindset Ai Ltd. has raised £4.3 million to expand its UK operations, open a Chicago office, develop new AI features and scale its go-to-market efforts. Led by Edge VC and Pembroke VCT, with participation from existing investors and Mountside Ventures, the company seeks to help SaaS companies easily embed AI agents. AI security vendor Straiker emerged from stealth with $21 million in initial funding from Lightspeed Ventures and Bain Capital Ventures. Straiker is presented as safeguarding AI applications and agents, with two modules related to attack simulation and automatic threat blocking, respectively. Content-delivery network and cybersecurity vendor Cloudflare Inc. has acquired Outerbase Inc. The stated intent is to enhance the developer database experience, making it easier to build and manage AI-enabled applications on Cloudflare's global network. For more funding information, please see the funding table below, where we have collected other funding highlights from the month, including major rounds, such as NexGen Cloud's series A. Politics and regulations Protect Elections from Deceptive AI Act was introduced in the US by Senators Amy Klobuchar (D-Minn.), Josh Hawley (R-Mo.), Chris Coons (D-Del.) and Susan Collins (R-Maine). This bipartisan bill aims to ban the use of AI to create deceptive content about federal candidates in political ads. Klobuchar emphasized the need for transparency and rules to prevent misleading AI-generated content. The bill's previous version passed the Rules Committee, suggesting potential broad support. The European Commission has published the AI Continent Action Plan, along with consultations on the Cloud and AI Development Act and the Apply AI strategy. This plan aims to advance Europe's digital sovereignty and suggests political desire to reduce reliance on non-European datacenters. It includes potential sovereignty requirements for public institutions like government, healthcare and education. The plan also proposes measures to boost European AI capacity and adoption, such as funding 13 AI factories and sector-specific "Apply AI" initiatives. A new Regulatory Intelligence Office underpinned by AI was announced in the United Arab Emirates. The goal appears to be to give AI a role in drafting, reviewing, and updating federal and local laws, with the intent to speed up the process of legislative activity. The Irish Writers Union submitted a petition with over 1,500 signatures to the Department of Enterprise in Dublin, protesting copyright violations in AI training. Meta was explicitly targeted in the document, following recent reporting about the training sets employed by the social media giant.. Gain access to our full news & research coverage and the industry-specific data that informs our insights This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global. Economics of Networks is a regular feature from S&P Global Market Intelligence Kagan. Consumer Insights is a regular feature from S&P Global Market Intelligence Kagan