Gemini 3.5 Pro Out, DeepMind A24 & AI Oncology Advances

Google's highly anticipated Gemini 3.5 Pro model has officially entered its General Availability (GA) window, a crucial step for broader enterprise and developer adoption. The model, initially announced at Google I/O in May, is now expected to be widely accessible between June 23 and June 30, 2026. This release is marked by significant architectural enhancements, including an expanded 2-million-token context window, which doubles the capacity of its predecessor, Gemini 3.5 Flash^[1]. This substantial increase in context allows the model to process and reason over vastly larger amounts of information, enabling more sophisticated and coherent outputs for complex tasks. A key architectural innovation within Gemini 3.5 Pro is the "Deep Think" reasoning mode, which is gated to the Ultra subscription tier^[1]. While specific details on the internal workings of Deep Think are proprietary, it is understood to be a specialized processing paradigm designed to enhance the model's analytical capabilities, allowing for more in-depth problem-solving and nuanced understanding of intricate prompts. This positions Gemini 3.5 Pro as a powerful tool for applications requiring extensive data analysis, complex code generation, and advanced content creation. The model also maintains multimodal support, processing both text and images seamlessly^[1]. The opening of the GA window for Gemini 3.5 Pro is set to intensify competition within the frontier AI model space, particularly against offerings like Anthropic's Claude Opus and OpenAI's latest models. The extended context window and specialized reasoning mode are direct responses to the industry's demand for more capable and reliable AI systems that can handle real-world operational complexities. Developers and enterprises are eager to integrate these advanced features into their applications, ranging from automated customer support and sophisticated data analytics to research and development workflows, where the ability to manage vast context and perform deep reasoning is paramount. The initial limited Vertex AI enterprise preview has generated considerable interest, signaling a strong market appetite for these architectural breakthroughs^[1].

Evogene, a company at the forefront of AI-driven molecular discovery, has announced a significant milestone in its collaboration with Google Cloud, advancing its proprietary generative AI platform, ChemPass AI™, towards autonomous small molecule discovery and optimization.^[1] This achievement positions Evogene to significantly accelerate the identification and refinement of novel, highly potent small molecules, which are critical for both pharmaceutical and agricultural applications.^[1]

The ChemPass AI™ platform is specifically designed to navigate vast chemical spaces and generate optimized small molecules across multiple parameters. By leveraging Google Cloud's Gemini models and engaging in agentic AI, Evogene is enhancing its ability to streamline complex biological and chemical research. This^[1] partnership is not merely about computational power; it represents a strategic integration of advanced AI to fundamentally redefine the molecular discovery process, moving beyond traditional trial-and-error methods.^[1]

This development has profound implications for industries reliant on new molecular entities, such as pharmaceuticals, where faster discovery can lead to quicker development of new drugs, and agriculture, for creating novel crop protection or enhancement agents. The collaboration aims to scale these capabilities across Evogene's pipeline, highlighting the transformative potential of AI in converting scientific innovation into real-world industrial solutions.

In a significant move for the creative sector, Google DeepMind has announced a strategic partnership with independent film studio A24 to explore the application of artificial intelligence in supporting filmmakers and other creative professionals. The collaboration aims to develop novel creative workflows while ensuring that future AI tools are shaped by the very individuals who will use them, directly addressing industry concerns about generative AI's impact on creative control.^[1]

This initiative comes at a time when the entertainment industry is grappling with the transformative potential and ethical implications of AI. By focusing on co-creation and user-centric development, Google DeepMind and A24 are seeking to allay fears of AI replacing human creativity, instead positioning it as an empowering co-pilot. A source close to the arrangement, cited by Reuters, confirmed that filmmakers will retain full creative control, explicitly stating that the deal is not an intellectual property or data-training agreement, a crucial distinction given ongoing disputes over AI and copyright in the creative sector.^[1]

Alphabet, Google's parent company, has also underscored its commitment by reportedly investing $75 million (£59 million) in A24, the studio behind acclaimed films such as "Everything Everywhere All at Once," "Marty Supreme," and "Backrooms."^[1] This financial backing signals a long-term vision for integrating advanced AI capabilities into the filmmaking process, potentially revolutionizing areas from script development and visual effects to post-production and distribution. The partnership is poised to establish new precedents for how AI can ethically and effectively enhance cinematic artistry.

A new report by Forrester, released in partnership with 4As at Cannes Lions 2026, reveals that generative AI is now pervasive across US marketing agencies, with nine out of ten agencies utilizing the technology. The primary drivers for this rapid adoption are enhancing productivity and achieving cost efficiencies.^[1] However, the report also issues a stark warning: this focus on efficiency is inadvertently "undermining marketing effectiveness, creativity, and long-term brand growth." ^[1]

The findings indicate that while agencies have successfully embedded AI into workflows, particularly for building creative content, crafting media and SEO strategy, and improving internal productivity, the industry risks "mistaking efficiency for effectiveness."^[1] Jay Pattisall, VP and principal analyst at Forrester, urged agencies and Chief Marketing Officers (CMOs) to pivot their focus. He emphasized the critical need to "reinvest efficiency gains into talent, training, and AI-powered marketing operating systems that enable differentiated experiences, stronger performance, and sustainable revenue growth." ^[1]

This report highlights a crucial inflection point for the creative industries. While generative AI offers undeniable benefits in automating routine tasks and accelerating content production, there's a growing concern that an overemphasis on cost-cutting could stifle the very innovation and unique human insight that drives compelling marketing. The call to action is clear: for AI to truly realize its full potential, it must be strategically integrated not just to do things faster, but to enable new forms of creativity and deliver more impactful, differentiated experiences for consumers.

Electronic Arts (EA), a major player in the video game industry, is experiencing a "real rise in creativity" and "faster prototyping" within its studios due to the integration of AI tools, according to Laura Miele, EA's president of enterprise development.^[1] Miele's remarks, made at The Game Business Live during Summer Game Fest, highlight how AI is effectively removing "tedious tasks" from developers' jobs, allowing them to focus more on innovative aspects of game creation. ^[1]

The adoption of AI in game development aims to make processes more efficient, empower developers, and accelerate the journey to the "fun" aspects of game design.^[1] While EA anticipates over 50% of its development processes will be positively impacted by generative AI advancements, the technology remains a contentious topic within the wider industry. A survey conducted in early 2026 by GamesIndustry.biz indicated that a significant majority of game industry respondents believed no amount of AI-generated content is acceptable in development, underscoring ongoing debates around authorship, job displacement, and creative integrity. ^[1]

Despite these reservations, EA's experience suggests that when deployed strategically, AI can enhance human creativity rather than diminish it. The company previously obtained permission from football commentator Guy Mowbray to replicate his voice via AI for generating player names, illustrating a practical application of the technology.^[1] This ongoing integration comes as Electronic Arts faces a major acquisition by a consortium of investors, including Saudi Arabia's Public Investment Fund, a deal expected to finalize by the end of June. ^[1]

A significant advancement in generative AI's application to software development has been highlighted with the Atoms platform, which was featured in reports on June 24, 2026. Atoms is an AI platform that translates natural-language descriptions into fully deployable applications by orchestrating a team of specialized AI agents.^[1] This represents a leap in generative AI architecture, moving beyond single-task code generation to encompass a comprehensive, end-to-end application development pipeline. The core architectural innovation of Atoms is its "AI Engineer" paired with a broader agent team.^[1] This team comprises distinct AI agents responsible for deep research, architecture design, product management, SEO, advertising, and data analysis. When a user provides a natural language prompt describing the desired application, Atoms intelligently distributes tasks among these specialized agents. These agents then collaborate to generate the frontend, backend, necessary integrations, and even hosting configurations automatically, resulting in a deployable application complete with authentication, a database, and payment processing. A^[1] particularly notable training and operational methodology employed by Atoms is "Race Mode," which enables the platform to run multiple models or agent teams in parallel on the same prompt.^[1] This allows for a direct comparison of different approaches and outputs, facilitating the selection of the best result. This parallel processing and comparative evaluation mechanism likely plays a crucial role in refining the agents' collaborative strategies and optimizing the overall application generation process. The impact of such agentic orchestration is transformative for software engineering, offering a pathway to significantly accelerate prototyping, development, and deployment cycles, potentially democratizing application creation for a wider range of users.^[1]

Sakana AI has unveiled its Fuku model, an innovative development in generative AI that prioritizes multimodal orchestration through a single API. Reported on June 24, 2026, the Fuku model distinguishes itself not as a monolithic generative model, but as a sophisticated router and orchestrator designed to coordinate various existing AI models to deliver optimal results^[1]. This architectural shift aims to leverage the strengths of specialized models, allowing for a more dynamic and potentially more accurate approach to complex, multimodal queries. The core of Fuku's innovation lies in its ability to parse user queries and intelligently route them to a variety of different, often smaller, models based on the specific requirements of the task^[1]. This system operates with a relatively compact 7 billion parameter model as its central decision-maker, enabling it to efficiently direct user requests. The promise of Fuku is to enhance the quality of outputs by orchestrating competitive and complementary AI models, a trend that has been gaining traction in the industry. For instance, similar approaches have been seen with systems that use multiple LLMs and a "judge" to identify the best answer^[1]. The implications of such an orchestration architecture are significant for inference engineering and the broader AI ecosystem. By moving beyond a single-model dependency, Fuku offers a pathway to potentially more robust and adaptable generative AI solutions, particularly in the wake of discussions around model reliability and output quality, such as those seen with "fable discussions"^[1]. While still an emerging trend, this approach could lead to more efficient resource utilization by selectively engaging highly specialized models only when needed, rather than relying on one massive, general-purpose model for every aspect of a multimodal task. Early benchmarks suggest that while it may consume more tokens currently, the architecture holds promise for future optimization and improved performance comparable to leading frontier models^[1].

New research published on arXiv in June 2026 indicates significant advancements in the architecture and training methodologies for multi-model AI systems, particularly focusing on emergent collaborative deliberation. One notable paper, "Emergent Collaborative Deliberation in Multi-Model AI Systems: A BFT-Derived Protocol for Epistemic Synthesis," proposes a novel protocol for enabling multiple AI models to collaboratively synthesize knowledge.^[1] This work addresses a critical challenge in the development of increasingly complex AI applications: how to effectively coordinate diverse AI agents or models to achieve a shared understanding or outcome. This architectural breakthrough draws inspiration from Byzantine Fault Tolerance (BFT) protocols, traditionally used in distributed computing to achieve consensus even when some components fail or act maliciously. By adapting these principles, the proposed protocol allows different AI models, potentially with varying capabilities or perspectives, to engage in a form of "epistemic synthesis" - a process of combining and refining their individual knowledge to form a more robust and comprehensive understanding.^[1] This suggests a move beyond simple ensemble methods towards a more interactive and deliberative form of AI collaboration, impacting how future AI systems are designed to handle ambiguity, conflict, and incomplete information. The implications for training methodologies are substantial. Such protocols necessitate training not just individual models, but also the meta-level coordination mechanisms that govern their interaction. This could involve reinforcement learning for collaborative behaviors or specialized training for "judge" or "arbiter" models within the multi-model framework. The ability for AI systems to engage in emergent collaborative deliberation is crucial for building more reliable, resilient, and intelligent agentic systems, especially in high-stakes environments where complex decision-making is required. This represents a foundational step towards more sophisticated collective AI intelligence, moving beyond individual model capabilities to harness the power of distributed, interacting intelligences.

BostonGene, a prominent developer of AI models in tumor and immune biology, announced its participation and presentation at The BIO International Convention (BIO 2026), highlighting groundbreaking AI innovations in precision medicine. On June 24, 2026, BostonGene revealed its focus on advanced multimodal AI models that integrate diverse biological data to transform clinical development and accelerate personalized cancer treatments^[1]. This represents a significant advancement in applying generative AI to complex scientific domains, particularly in healthcare. The core innovation lies in BostonGene's model architecture, which is designed to integrate multi-dimensional molecular, cellular, and patient data across tumor and immune biology. This multimodal approach allows the AI to process and synthesize information from various sources, including genomic, transcriptomic, immune, and clinical signals^[1]. By converting these complex biological datasets into clear, actionable insights, the models improve the accuracy of predicting patient responses to therapies and assessing toxicity risks. This capability is crucial for de-risking clinical trials and expediting drug development across a range of cancer indications^[1]. BostonGene’s solutions demonstrate how specialized training methodologies, leveraging vast and diverse biological datasets, can lead to highly impactful generative AI applications. Through detailed case studies, the company plans to illustrate how its multimodal AI models can prioritize indications by modeling the tissue-based immune microenvironment, optimize clinical differentiation strategies by predicting indication effect sizes, and refine patient stratification through blood-based immune-profiling scores.^[1] This targeted application of advanced AI architecture and data integration promises to significantly enhance the efficiency and success rates of oncology clinical trials, ultimately leading to more effective and personalized treatments for patients. ^[1]

BostonGene, a leading developer of AI models for tumor and immune biology, is set to present its latest innovations at The BIO International Convention (BIO 2026). The company will showcase how its advanced multimodal AI models are transforming clinical development and accelerating precision medicine, particularly in the context of cancer. ^[1][2]

The presentation, titled "Clinical Trial Advantage Through Patient Biology," will highlight BostonGene's AI solutions that integrate multi-dimensional molecular, cellular, and patient data. These models convert complex biological datasets into clear, actionable insights, improving the accuracy of predicting patient responses and toxicity risks.^[1][2] This capability is crucial for accelerating drug development and de-risking clinical trials across various cancer indications. ^[1][2]

Through three detailed case studies, BostonGene will demonstrate how its AI models operate across different modalities to: prioritize indications by utilizing tissue-based immune microenvironment modeling for high-yield cohort selection; optimize clinical differentiation strategies by creating integrated models to predict indication effect sizes, thereby reducing trial enrollment requirements; and refine patient stratification by developing ready-to-use immune-profiling scores from blood-based signatures to enhance therapy response rates and optimize trial designs.^[1][2] This work underscores the increasing role of generative AI in translating complex biological data into practical strategies for developing more effective and personalized cancer treatments.

ImagineArt has launched a comprehensive AI creative suite designed to consolidate disparate tools for content creation, marking a significant step towards unifying the generative AI industry. The new platform integrates functionalities for image generation, video production, advertising content creation, and film design, featuring expanded studios and full Model Context Protocol (MCP) integration.^[1] This release positions ImagineArt as a leading contender in providing a holistic creative AI solution, directly tackling the challenge of "tool fragmentation" that has long plagued creators and marketing teams. ^[1]

Historically, creative professionals have relied on a multitude of specialized software, each with its own learning curve, pricing model, and quality inconsistencies. This patchwork approach often led to disconnected workflows and increased operational complexity. ImagineArt's all-in-one suite aims to streamline this process by offering every essential creative function within a single workspace.^[1] This allows solo creators, marketing departments, and large enterprises to manage entire campaigns, from initial concept to final output, without switching platforms.

The launch reflects a broader industry trend towards integrated solutions that enhance efficiency and creative output. By offering a unified platform, ImagineArt seeks to democratize access to advanced creative capabilities, enabling teams that previously lacked specialized resources to produce high-quality visual and audio content.^[2] The company anticipates that this consolidation will become standard practice, allowing businesses and creators to achieve professional-grade results with greater ease and reduced overhead.

Q.ANT, a pioneer in commercial photonic computing, has successfully demonstrated the execution of complex, production-relevant AI workloads on its second-generation Native Processing Unit (NPU). At ISC High Performance 2026 in Hamburg, Germany, Q.ANT showcased the operation of both a diffusion model for generative image synthesis and a recurrent neural network for sequential time series prediction on its photonic hardware. This^[1][2] breakthrough validates that Q.ANT's photonic architecture can support the full spectrum of modern AI capabilities.

The^[1] significance of this achievement lies in the remarkable energy efficiency offered by photonic computing. By performing computations with light rather than traditional transistors, Q.ANT's hardware aims to reduce energy consumption at the source, targeting 30 times the energy efficiency of classical processors for equivalent matrix operations at the photonic circuit level. This^[1] addresses a critical bottleneck in the future of AI, as the energy demands of increasingly complex AI models continue to grow.

The^[1] demonstrations at ISC High Performance build upon earlier progress this year, where independent developers successfully compiled and deployed an object detection model from PyTorch directly onto Q.ANT's photonic processor. This^[1] indicates that the NPU has advanced beyond foundational algorithms to support genuine commercial applications. Lukas Fischer, Head of Applied Research at NXAI, a company leveraging Q.ANT's technology, commented that the xLSTM architecture running on photonic systems "could redefine what energy-efficient AI even means," particularly for applications like financial market analysis, supply chain optimization, and weather forecasting.^[1][2]