Generative AI Digest: More AI agents and more funding

In recent weeks, both OpenAI LLC and Amazon Web Services Inc. have made significant announcements related to AI agents, while agentic coding startup Cognition AI Inc. acquired what remained of Exafunction Inc., better known as Windsurf. However, July's developments extended beyond agents: X.AI Corp. attracted a mix of praise and criticism, Meta Platforms Inc. acquired AI voice company PlayHT Inc., major Chinese AI developers secured substantial funding rounds, and California courts issued copyright rulings favoring foundation model providers.

In July, xAI drew significant attention with a $10 billion funding round and new model releases. The release of Grok 4 potentially establishes xAI as a contender in the AI sector, with the model and its "heavy" variant achieving notable results on several of the most challenging AI benchmarks. However, some critics point out that Grok 4 falls short of some competing models regarding cost efficiency, response latency and context window size, potentially limiting its practical applications. In addition, the launch was mired in a backlash after reports of antisemitic responses from the Grok chatbot sparked media scrutiny and raised safety concerns. These concerns were likely not aided by the introduction of "Ani," an AI companion in Grok's companion mode that uses explicit language, which again highlights xAI's less stringent approach to response safety compared with other generative AI companies.

Product releases and updates

XAI made headlines with the release of Grok 4 and its resource-heavy sibling, Grok 4 Heavy. The model demonstrated leading performance against several well-known benchmarks, including the challenging crowdsourced specialist questions of Humanity's Last Exam and puzzle benchmark ARC-AGI-2. Grok 4 Heavy, available via a $300-per-month "SuperGrok Heavy" subscription, is built using a multi-agent architecture. The release was marred by antisemitic content from the Grok chatbot on X, which received significant focus online, as well as concerns related to long response times and cost. XAI also introduced a set of AI character personas for Grok, which gained traction but drew further backlash — particularly around the female Ani companion, which some viewed as overly sexualized.

Creative design platform Picsart unveiled several new tools, including Ad Maker, which is an AI-driven end-to-end ad campaign tool, and a conversational AI assistant that can leverage a range of creative AI models.

In partnership with Stability AI Ltd., Advanced Micro Devices Inc. announced a new image model that takes advantage of the BF16 format on their new SD3 Medium neural processing unit chip. This is notable because the BF16 (bfloat16) format should allow AI models to process data much faster and more efficiently, using less memory than traditional formats like FP32. Although the BF16 format is supported across a number of accelerators, it has traditionally been associated with high-end datacenters rather than on-device AI hardware. Its use here highlights how more accessible and power-efficient chips can now be harnessed for image generation tasks.

In other number format innovations, Alibaba Cloud Computing Co. released a major update to its Qwen3 family of open-source large language models and included a new FP8 version, designed to compress operations to use less GPU memory and power. The new Qwen3-235B-A22B-Instruct-2507 model performs significantly better for challenging reasoning tasks than earlier variations and works more effectively in less common languages.

AWS made a range of announcements at AWS Summit in New York. Amazon Bedrock AgentCore saw significant focus — an offering made of runtime, memory, identity, browser, observability, code interpreter and gateway services. Kiro is a stand-alone integrated development environment that is pitched as supporting both "vibe coding" (the ability to autonomously generate applications based on natural language) and "spec coding" (referring to a structured development approach built from a series of design documents and planning steps). Agent hooks are agentic actions that can be triggered by events, which could involve automatically generating tests, for example.

OpenAI has launched ChatGPT Agent, a new agentic system that integrates reasoning, web browsing, and code execution to autonomously perform complex, multi-step tasks. The agent appears to combine the core functionality of Deep Research and Operator, in being able to conduct research, interact with public websites, fill out forms, edit spreadsheets, and access third-party sources like email and document repositories, and should lead to the stand-alone Operator experience being depreciated.

Music generator Suno Inc. announced a new interface and a set of enhancements including lengthier audio uploads and the ability to split songs into isolated audio tracks for different instruments and effects. These announcements follow its V4.5 model release in May, which added greater variety and diversity of outputs, alongside a prompt enhancer.

Chinese AI startup Moonshot AI released Kimi K2, an open-source mixture-of-experts model, nominally optimized for agentic intelligence and coding, with Moonshot AI suggesting it outperformed OpenAI's GPT-4.1 and Anthropic PBC's Claude 4 Opus models on key coding and reasoning benchmarks. The company used the announcement to showcase its MuonClip optimizer, an enhancement to its previous Muon optimizer, which it suggests sets a strong foundation for stable, large-scale LLM development.

Adobe Inc.'s Firefly Video received a significant upgrade, introducing style presets, enhanced visual fidelity, text-to-avatar generation, support for composition references, and improved sound effects. Users can now create timed audio effects using either text prompts or voice commands.

Mistral AI SAS had a very busy mid-July. In the space of a week, the company announced a new medium variant of its Devstral coding model series, two speech-understanding models under the brand Voxtral and a range of enhancements to chatbot Le Chat. These enhancements include the addition of a voice model, a chat management capability known as "projects" and "deep research" functionality.

In late June, Eleven Labs Inc. announced voice assistant 11.ai in alpha, designed to integrate with various tools via the Model Context Protocol and perform actionable tasks. The release aims to demonstrate the potential of voice-driven productivity and cites the ability to connect to services like Slack Technologies LLC, Linear and Perplexity AI Inc.

At the end of June, Baidu Inc. open-sourced a suite of 10 text and multimodal models, known as the ERNIE 4.5 family, under an Apache 2.0 license. This marks a significant strategic shift, as Baidu had previously maintained a more closed approach compared with other leading Chinese AI companies. The models vary in size from 300 million to 424 billion parameters.

Funding and M&A

AI coding startup Windsurf, formerly known as Codeium, experienced a turbulent July. The company was apparently in acquisition talks with OpenAI for a $3 billion deal, which some reports suggested primarily collapsed due to intellectual property concerns involving Microsoft Corp.'s right to use OpenAI's technology. Shortly afterward, Google DeepMind hired Windsurf co-founder and CEO Varun Mohan, along with other senior executives and researchers, in a deal that appeared to include Google LLC securing a nonexclusive license to Windsurf's technology, although not acquiring a stake in the company. Days later, Cognition, the startup behind AI coding agent Devin, announced a definitive agreement to acquire Windsurf, including its intellectual property, product and remaining team.

XAI raised $5 billion in debt and $5 billion in equity. This latest injection is seen as a route to further bolstering its huge infrastructure investments. This brings the company's total funding to over $22 billion.

Shanghai Xiyu Jizhi Technology Co. Ltd., better known as MiniMax, announced that it had received nearly CNY 2,152 million ($230 million) in a new funding round. This brings its total funding to $1.15 billion.

Meta continued its hunt for AI talent by acquiring audio generator PlayHT and suggesting in an internal memo that the entire team would join Meta. Details are sparse, but the memo reportedly suggested that the technology would align with Meta's work in areas like AI characters and wearables. PlayHT had received $21.5 million in total funding since its 2019 incorporation and had a head count of about 35.

Confident Security, an AI privacy startup based in San Francisco, came out of stealth with $4.2 million in seed funding. The startup promises to be able to evidence that all prompts and interactions with an AI system are entirely private.

Beijing Knowledge Atlas Technology Co. Ltd., better known as ZhipuAI, announced just under $140 million in funding in a new round, valuing the startup at $5.58 billion. The company develops language and video models, as well as smartphone voice agents.

Multimodal AI startup Reka AI Inc announced $110 million in funding, which included new investors NVIDIA Corp. and Snowflake Inc. The company raised the round at a post-money valuation of $1 billion.

Progress Software Corp. acquired retrieval-augmented-generation specialist Nuclia, officially incorporated as Besutech XXI S.L. The company has achieved $5.89 million in total funding and a head count of 25.

Swedish software startup Lovable AB has secured $200 million in a series A round led by Accel Partners. The investment also saw participation from byFounders VC Management, Creandum Advisor, Hummingbird Ventures Management, Visionaries Club, The Twenty Minute VC, and angel investors including Sebastian Siemiatkowski, Job van der Voort, Stewart Butterfield and Dharmesh Shah. The round values the company at $1.8 billion.

Politics and regulations

Following the Senate's vote to strip a proposed 10-year moratorium on state-level AI regulation from the current US administration's budget bill, AI companies are expected to return to aggressively pushing for federal legislation, in order to avoid the fragmented regulatory landscape that is already emerging across individual states. Republican Senator Thom Tillis was the lone lawmaker who voted to retain the ban.

The US administration issued several executive orders, including measures to bar AI companies perceived as "ideologically biased" from receiving federal contracts, streamline datacenter permitting and expand AI technology exports by lifting compute capacity restrictions on certain countries. The administration released an AI Action Plan featuring 90 recommendations aimed at securing "US leadership in the global AI race."

In July, a public hearing was held on the Advance Notice of Proposed Rulemaking titled "Guidelines to Prevent Deceptive Trade Practices in the Use of Artificial Intelligence Decision-Making that Impacts Rights, Opportunities, or Access to Resources or Services," issued by the Rhode Island Attorney General. The proposal aims to address discriminatory outcomes and misleading results produced by AI systems.

In two landmark decisions, federal judges in California ruled in favor of AI companies Anthropic and Meta, finding that their use of copyrighted books to train large language models qualifies as fair use under US copyright law. In both instances, the outputs of the models were deemed "transformative." A separate trial is pending over Anthropic's alleged use of "pirated" books.

In recent weeks, the European Commission confirmed that the AI Act will proceed as scheduled, despite notable pushback from some industry groups and politicians advocating for a delay. To support compliance, a voluntary set of guidelines was released on July 10. While OpenAI publicly committed to the code, Meta has publicly declined to sign.

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.