Catch up on select AI news and developments from the past week or so (in no particular order):
OpenAI launches o3 and o4-mini models that 'think with images'. OpenAI has released o3 and o4-mini reasoning models with advanced image analysis capabilities. The models can interpret sketches, diagrams, and whiteboards—even at low quality—and integrate that data into reasoning processes. O3 is optimized for math, science, and image understanding, while o4-mini is faster and cheaper. These models can now independently use tools like Python, Web browsing, and image generation. Importance for marketers: These multimodal capabilities support more visual campaign planning, content creation, and analysis, enhancing marketing teams' productivity and ideation workflows.
Google's Veo 2 AI video generator now available to Gemini Advanced users. Google's powerful Veo 2 video generator is now accessible via its $20/month Gemini Advanced subscription. Previously limited to enterprise customers, Veo 2 creates 720p, 8-second videos in MP4 format using AI prompts. Its model, trained on vast YouTube content, generates stylized clips like crochet animals and cartoon cats. Google also launched Whisk Animate to enhance prompt-based animation creation. All Veo videos will include SynthID watermarks for transparency. Importance for marketers: Easier access to AI video tools like Veo 2 allows marketers to create dynamic video content for campaigns without high production costs or technical expertise.
Claude can now search Google Workspace and conduct autonomous research. Anthropic upgraded Claude with autonomous research tools and Google Workspace integration, positioning it as a virtual collaborator. Claude can independently conduct layered web searches and cite sources, while accessing user calendars, emails, and files (with admin approval). Anthropic emphasizes fast research speed and enterprise-grade security. Initial testing suggests significant productivity gains. Importance for marketers: This empowers teams to automate research, competitive analysis, and internal knowledge retrieval—boosting marketing velocity and supporting content and campaign planning.
OpenAI introduces Flex processing for lower-cost, noncritical AI tasks. OpenAI has launched Flex processing, a discounted API option for its o3 and o4-mini models. Designed for lower-priority tasks like data enrichment and asynchronous processing, Flex halves the usual input and output token prices. The tradeoff is slower performance and occasional unavailability. Flex responds to increasing AI cost pressures and competition from budget models like Google's Gemini 2.5 Flash. OpenAI also now requires ID verification for users accessing o3 and other models, particularly those in the first three pricing tiers. This move aligns with OpenAI's broader security efforts. Importance for marketers: Flex pricing offers a cost-effective option for marketers performing non-urgent tasks such as A/B testing, tagging, or enrichment.
Microsoft unveils CPU-friendly BitNet model for efficient AI performance. Microsoft researchers have introduced BitNet b1.58 2B4T, a hyper-efficient "bitnet" model that runs on CPUs, including Apple's M2. With 2 billion parameters and trained on 4 trillion tokens, it surpasses similarly sized models in performance benchmarks. BitNets quantize weights to -1, 0, or 1, offering greater efficiency in memory and processing. While performance is strong, BitNet requires Microsoft's custom bitnet.cpp framework and lacks GPU compatibility. The model is open source under the MIT license. Importance for marketers: Marketers working on lightweight or embedded systems may benefit from this new class of efficient AI, especially for on-device applications.
Microsoft's Copilot Studio can now automate desktop and Web tasks. Microsoft has added a "computer use" feature to Copilot Studio, allowing AI agents to interact with apps and websites by mimicking user actions such as clicks, form entries, and menu selections. This means agents can operate without an API, even when interfaces change dynamically. Microsoft claims agents can handle tasks like data entry and invoice processing independently. This mirrors similar functionality in OpenAI and Claude. A comparable "Actions" feature also launched in the consumer-facing Copilot earlier this month. Importance for marketers: This expands marketers' ability to automate tasks on non-API systems, speeding up operations like lead enrichment or form submissions.
Apple to train AI models with synthetic emails while preserving user privacy. Apple is using synthetic email data to improve AI-driven email summaries for opted-in users. Unlike rivals like Meta, which train on user posts unless opted out, Apple uses machine-generated messages combined with differential privacy techniques to enhance model performance while safeguarding user data. Though synthetic data introduces potential bias and limitations, it ensures sensitive information isn't exposed. Apple's method balances user privacy with the need for effective training inputs, especially in light of recent criticism over its AI capabilities. Importance for marketers: Apple's privacy-forward approach may shape expectations for responsible AI, especially in customer communication tools.
Anthropic to launch Claude voice assistant with multiple voice options. Anthropic is preparing to roll out a "voice mode" for its Claude AI chatbot, with three voices—Airy, Mellow, and Buttery. The feature will allow users to interact via voice, similar to OpenAI's ChatGPT. Voice mode had been mentioned previously by Anthropic's chief product officer and confirmed by app researchers. This addition positions Claude more competitively in the AI assistant landscape. The launch could occur as soon as this month, according to Bloomberg. Importance for marketers: Voice interaction capabilities open up new channels for user engagement, especially for content delivery, customer service, and accessibility-focused experiences.
Meta resumes training AI on EU users' public posts. Meta will again use publicly available posts from EU users to train its AI models, resuming efforts paused last year over privacy concerns. The company says it won't use private messages and will provide users with opt-out options. This follows regulator review and coincides with Meta launching its AI assistant in Europe. Meta cites similar practices by OpenAI and Google as precedent. Importance for marketers: Meta's resumed training means AI models may better reflect European user behavior—potentially improving relevance in EU-targeted campaigns and user modeling.
US considers penalties to block DeepSeek's access to American tech. The Trump administration is exploring penalties to prevent China's DeepSeek from acquiring US technology and may restrict American access to its services. DeepSeek's low-cost AI models, possibly trained on US-sourced hardware, have alarmed regulators. Nvidia's H20 AI chip shipments to China have already been limited, and US lawmakers are pressing Nvidia on compliance. Export controls aim to prevent Chinese military use of advanced AI. Importance for marketers: If restrictions tighten, global marketers may see disruptions in China-focused AI services and competitive shifts in available models or pricing structures.
Perplexity AI assistant to appear in upcoming Motorola Razr. Motorola's next Razr will reportedly integrate the Perplexity AI assistant, alongside Google's Gemini, with a dedicated user interface to encourage usage. The deal is part of a broader trend, with Perplexity also in early talks with Samsung to appear on its devices. Perplexity's AI agents may soon handle tasks like booking flights via voice commands. The Razr announcement is expected April 24. Importance for marketers: Native AI assistants in smartphones could redefine how users discover and act on information, impacting mobile marketing, SEO, and customer journey mapping.
ChatGPT adds image library to organize AI-generated visuals. OpenAI has rolled out an image library in ChatGPT, allowing users to view, access, and create AI-generated images in one place. Available on mobile and web, the feature shows a grid of previously generated images and includes an option to create new ones. This helps users manage their visual content without losing track of outputs. Importance for marketers: Marketers who rely on ChatGPT for visual ideation can now easily browse and reuse assets, supporting brand consistency and streamlining creative workflows.
OpenAI reportedly developing a social network focused on AI-generated content. OpenAI is exploring a new social network akin to X (formerly Twitter), featuring a social feed tied to ChatGPT's image generation. Although still in early development, the internal prototype suggests OpenAI may integrate real-time user content for data training. The effort would directly compete with Meta and Musk's xAI. Importance for marketers: A new OpenAI-owned platform could provide a high-quality channel for AI-generated content sharing and data insight—potentially shifting how marketers experiment and distribute creative assets.
China's Zhipu AI begins IPO process with state backing. Chinese AI startup Zhipu AI has begun preliminary steps toward a public offering, aiming to become China's first next-gen AI firm to go public. Backed by state funding, Zhipu claims its GLM4 model outperforms GPT-4 on several benchmarks. The company also open-sourced some of its smaller models and touts significantly lower operating costs. Importance for marketers: Zhipu's growth signals a strengthening of China's AI ecosystem, which may drive pricing and innovation pressure on Western vendors marketers currently rely on.
OpenAI prepares GPT-4.1 and smaller models for imminent launch. OpenAI is set to release GPT-4.1, a refined version of its GPT-4o model, along with smaller mini and nano variants. These will improve multimodal performance across text, audio, and image processing. Delays in rollout may occur due to infrastructure limitations. The o3 and o4-mini reasoning models are also part of this update cycle. Importance for marketers: Enhanced models with more efficient architecture may offer faster performance and better handling of mixed media content—key for campaign generation, creative development, and analysis.
OpenAI to phase out GPT-4 in ChatGPT by April 30. OpenAI will retire GPT-4 in ChatGPT, replacing it with GPT-4o. GPT-4 will remain available through API access. OpenAI states that GPT-4o outperforms GPT-4 in writing, coding, STEM, and conversation. GPT-4 had previously cost over $100 million to train and is involved in ongoing copyright disputes. The phase-out coincides with a broader model refresh. Importance for marketers: Marketers relying on GPT-4 should migrate workflows to GPT-4o and evaluate performance, especially for creative generation and data analysis in content and campaign workflows.
Irish regulator investigates X over AI training with EU personal data. Ireland's Data Protection Commission has launched an investigation into X's (formerly Twitter) use of publicly available EU user data to train its Grok AI system. X previously agreed to stop using such data without consent. The DPC's action may result in fines up to 4% of global turnover. The move underscores EU regulators' scrutiny of AI model training practices. Importance for marketers: Tighter enforcement of data privacy in the EU may affect how social platforms handle user data and limit available training signals for ad targeting and content personalization.
You can find the previous issue of AI Update here.
Editor's note: GPT-4o was used to help compile this week's AI Update.
