In partnership with

Good morning. It’s Friday, April 25th.

On this day in tech history: In 1961, Robert Noyce was granted a U.S. patent for the silicon-based integrated circuit, a pivotal development that laid the groundwork for modern microchips and led to the founding of Intel in 1968.

In today’s email:

• OpenAI Expands Deep Research for Free Users

• Anthropic Delves into “AI Welfare”

• Gemini Gains On ChatGPT

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email. Though today I am in the hospital with my wife, we are about to have our second baby - so it might take me an extra day to respond!

OpenAI Expands Deep Research for Free Users, Launches GPT-Image-1 API for Developers

OpenAI has introduced a cost-efficient, lightweight version of its Deep Research feature for free-tier ChatGPT users, offering up to five queries per month. This version runs on the o4-mini model, providing comparable depth to the standard o3-powered tool, though responses may be shorter. Paid users on Plus, Pro, and Team plans can access additional queries, with automatic fallback to o4-mini once the o3 limit is exceeded.

In related updates, OpenAI's models continue to perform strongly on the LMSYS Arena leaderboard, with the o3 model securing the #2 spot and excelling in math, coding, and style control. The o4-mini also ranks highly, particularly in mathematical tasks.

OpenAI’s Open Model to Invoke Cloud for Complex Queries: OpenAI is set to release its first fully downloadable AI model in years, designed to outperform open-source models from Meta and DeepSeek. Internally dubbed a “reasoning” model, it may feature an optional “handoff” capability, enabling it to invoke OpenAI’s larger, cloud-hosted models for complex queries via API.

The handoff concept, proposed by a developer during a community event, has reportedly gained traction within OpenAI. While the exact tools and pricing structure remain unclear, the company is training the model from scratch rather than adapting an existing one.

Revenue Projections: OpenAI Eyes $174B by 2030: OpenAI expects its revenue to grow from $4 billion in 2024 to $174 billion by 2030, driven by AI agents, new products, and monetization strategies. Agents like “Operator” could generate $29 billion by 2029, with $50B from subscriptions and $22B from APIs.

GPT-Image-1 API Launches for Developers: OpenAI has released its GPT-Image-1 model —originally integrated into ChatGPT—via API, expanding access to high-fidelity image generation and analysis tools for developers. The model supports text-to-image synthesis with precise prompt alignment and image interpretation across PNG, JPEG, WEBP, and GIF formats. Input can be submitted via URL or Base64, with output resolution capped at 768×2000 pixels.

Pricing is usage-based: $5 per million tokens for text prompts, $10 for image inputs, and $40 for image outputs, with per-image costs ranging from $0.02 to $0.19 depending on detail. Safety measures match ChatGPT-4o, including C2PA metadata and content moderation, and API data is excluded from training.

Anthropic Targets AI Black Box Transparency and Welfare Research by 2027

Anthropic has launched a research program focused on AI "model welfare," investigating whether AI systems could experience human-like traits such as distress or consciousness. This initiative follows a report co-authored by philosopher David Chalmers, suggesting AI might develop consciousness and agency. The research will explore the ethical implications of these possibilities, looking for signs of distress or preferences in AI models.

Led by Kyle Fish, the research will analyze both behavioral responses—such as how models react to choices—and internal mechanisms that could align with theories of consciousness. While no consensus exists on whether AI can achieve consciousness, Anthropic is proceeding cautiously, acknowledging the uncertainty in the field.

Additionally, CEO Dario Amodei has outlined plans to improve AI model interpretability by 2027, urging the AI industry to ramp up efforts in mechanistic interpretability to mitigate risks of autonomous decision-making.

Google’s Gemini Surpasses 350 Million Users, Gaining Ground on ChatGPT

Google has teamed up with HHMI Janelia and Harvard to release ZAPBench, a comprehensive brain activity dataset aimed at refining predictive models of brain function. Capturing data from approximately 70,000 neurons in a larval zebrafish's brain, the dataset offers new insights into how neural activity correlates with brain structure. With ZAPBench, researchers can compare and enhance models that track brain activity over time. The next step will be incorporating a full connectome to increase model accuracy.

Meanwhile, Google's Gemini chatbot has reached 350 million active users, with the actual number likely higher due to its integration with platforms like Google Search and Android. While it still lags behind OpenAI’s ChatGPT, which has over 500 million weekly users, the latest Gemini 2.5 Pro strengthens Google's position in the AI race. Notably, in today’s earnings call, Google revealed that over a third of all code generated at the company is now produced by AI.

On the creative side, Google’s Music AI Sandbox has enhanced its features and expanded access for artists. The platform now includes Lyria 2, an advanced music generation model offering high-fidelity audio. Tools like "Create," "Extend," and "Edit" help artists explore genres and compose new tracks, while Lyria RealTime enables real-time music creation.

• Nvidia's G-Assist adds a plug-in builder so you can create your own PC AI automations

• MIT's Machine Learning "Periodic Table" Could Accelerate AI Discoveries

• Microsoft enhances 365 Copilot with new AI agents and image generation tools

• New Firefly Models from Adobe Include OpenAI Tools for Creative Workflows

• Perplexity brings multitasking AI voice assistant to iOS

• IGN and CNET owner Ziff Davis sues OpenAI

• LLMs Won't Scale to AGI, But Instead We'll Need Complementary AI Approaches"

• Apple to strip secret robotics unit from AI chief weeks after moving Siri

• United Arab Emirates first nation to use AI to write laws

• China’s Baidu Upgrades Ernie AI Models and Slashes Prices

• Krea AI Integrates ChatGPT Image for Chat-Based Editing and 3D Rendering

• Perplexity CEO says its browser will track everything users do online to sell 'hyper personalized' ads

• Tencent Hunyuan Launches AI Engine v2.5 with Ultra-HD 3D Modeling Capabilities

• Former DeepSeeker and collaborators release new method for training reliable AI agents: RAGEN

• Researchers warn models are "only a few tasks away" from autonomously replicating

• Relyance AI builds 'x-ray vision' for company data: Cuts AI compliance time by 80% while solving trust crisis

• Gemini is on track to being the first Al to beat Pokémon Red. It has beaten 6 gyms

• California Bar discloses AI was used to develop some questions in problem-plagued February exam

• New system quantifies language model personalities through linguistic analysis

• Algorithm can make AI responses increasingly reliable with less computational overhead

• Research shows humans are still better than AI at reading the room

• In Q1 2025, AI commanded 71% of total VC deal value, according to PitchBook

📄 VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models

📄 DreamID: High-Fidelity and Fast diffusion-based Face Swapping via Triplet ID Group Learning ​​

📄 Step1X-Edit: A Practical Framework for General Image Editing

📄 Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning

📄 Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!