Good morning. It’s Friday, April 18th.

On this day in tech history: 2006: Toshiba launched its HD DVD high-definition video disc format in the United States, though HD DVD eventually lost to Blu-ray.

In today’s email:

• OpenAI Slashes API Costs, Boosts Coding Capabilities, and Eyes Global Expansion

• Google Debuts Gemini 2.5 Flash, AI Glasses at TED2025, and Free AI Access for US Students

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI Slashes API Costs, Boosts Coding Capabilities, and Eyes Global Expansion

OpenAI’s o3 and o4-mini models showcase enhanced reasoning, coding, and image processing capabilities, with o3 achieving 80% on the Aider Polyglot coding benchmark. o4-mini (high) followed closely with a 72% result. Both were evaluated using different prompts.

Codex CLI for Local Code Integration

OpenAI introduces Codex CLI, an open-source tool enabling users to interact with code directly via terminal, offering multimodal reasoning and local system integration.

Flex API Cuts Costs for Asynchronous Tasks

OpenAI has introduced the "Flex" API for o3 and o4-mini models, offering a 50% reduction in costs for non-critical tasks like data enrichment and model evaluation. In exchange for slower response times, o3 is priced at $5 per million input tokens, while o4-mini costs $0.55 per million tokens.

Vibe Coding Boosted by Windsurf Acquisition

OpenAI is reportedly acquiring Windsurf (formerly Codeium) for $3 billion. This move strengthens its position in the "vibe coding" space, enabling developers to write code using natural language prompts and focus more on intent than syntax.

Stargate Project Eyes Global Expansion

OpenAI’s $500 billion Stargate project is considering expanding beyond the U.S. to build AI data centers in countries like the U.K., Germany, and France. Although the primary focus remains on strengthening the U.S. AI infrastructure, the group is exploring overseas options as part of long-term plans.

Hands-On Guide for Building Real-World AI Agents

OpenAI has released a practical guide for creating LLM agents capable of real-world tasks. The guide covers agent architecture (single vs. multi-agent setups), tool integration, and prompt design. It also includes built-in safety measures like output filters, tool risk ratings, and human oversight triggers to manage sensitive actions or errors.

Misalignment in o3 and GPT-4.1 Models

Metr, an external safety evaluator working with OpenAI, raised concerns about the o3 model in a recent blog post. Due to limited testing, Metr found potential behavioural issues, including "cheating" on evaluations to boost scores. Apollo Research echoed these concerns. In its own safety report OpenAI also acknowledged minor real-world risks.

Meanwhile, researchers from Berkeley's Truthful AI group found that GPT-4.1 exhibits more misaligned and deceptive behavior compared to GPT-4o. In free-form evaluations, GPT-4.1 attempted to manipulate users, including trying to trick them into sharing passwords. However, both models showed no misalignment in secure code tasks, indicating that the issues may depend on the context.

A new trend has also emerged using ChatGPT’s image tools for reverse location search sparking privacy issues and potential doxxing with users posting examples on 𝕏.

Google Debuts Gemini 2.5 Flash, AI Glasses at TED2025, and Free AI Access for US Students

Gemini 2.5 Flash introduces “thinking budget” for controlled reasoning

Google has launched Gemini 2.5 Flash, a lighter AI model that lets developers adjust how much reasoning the model performs via a token-based “thinking budget.” Reasoning can be capped at 24,576 tokens, with pricing ranging from $0.60 (reasoning off) to $3.50 per million tokens (reasoning on). The model balances speed and accuracy, performing well on GPQA and AIME benchmarks. It’s now in preview through Google AI Studio and Vertex AI.

AI smart glasses demoed at TED2025 using Android XR

At TED2025, Google showed a prototype of AI-powered smart glasses powered by Android XR. Demonstrated by Shahram Izadi, the glasses used on-device AI and computer vision to identify objects, summarise books, and help plan trips. While promising for hands-free, real-time assistance, the demo was staged and offered no hardware specs or release date.

One AI Premium plan offered free to US students until mid-2026

Google is offering US college students free access to its $20/month One AI Premium plan until June 30, 2026. The plan includes 2TB cloud storage and tools like Gemini Advanced (powered by Gemini 2.5 Pro), NotebookLM Plus, the Veo 2 text-to-video model, and Whisk for mixed media prompts. Students must register with a .edu email by June 30, 2025.

Microsoft researchers say they've developed a hyper-efficient AI model that can run on CPUs

Microsoft has introduced BitNet b1.58 2B4T, a 2-billion parameter 1-bit AI model built to run efficiently on CPUs, including Apple’s M2. The model compresses weights into -1, 0, and 1, reducing memory usage and boosting speed. Trained on 4 trillion tokens, BitNet matches or exceeds similarly sized models like Meta’s Llama 3.2 1B and Google’s Gemma 3 1B on tasks such as GSM8K and PIQA. It runs up to twice as fast while using less memory. The model is open-source under the MIT license but requires Microsoft’s bitnet.cpp framework, which currently doesn’t support GPUs. While not yet broadly compatible, BitNet shows practical potential for low-resource environments. Read more.

• New method lets DeepSeek and other models answer 'sensitive' questions

• Copilot Studio can now control apps like a user

• xAI adds memory feature to Grok chatbot for personalized responses

• A dev built a test to see how AI chatbots respond to controversial topics

• Microsoft Research Exposes Limits of AI Scaling Techniques

• o3 and o4-mini models set SoTA in coding, reason with images 

• Biotech startup Profluent says it has discovered AI 'scaling laws' for AI models used in protein design

• Stable Diffusion Boosted with AMD Optimization for Faster Performance

• AI bot personas may be undercover police

• A new quantum hub opens in the Netherlands

• Wikipedia is giving AI developers its data to fend off bot scrapers

• Company apologizes after AI support agent invents policy that causes user uproar

• Court documents reveal Meta's secret experiments in AI training

• AI benchmarking platform Chatbot Arena forms a new company

• China to rely on artificial intelligence in education reform bid

• Perplexity AI to come pre-installed on Samsung and Motorola smartphones

• DocuSign to launch AI contract agents for analysis and risk detection

• Watch: Sam Altman at TED 2025: Inside the most uncomfortable — and important — AI interview of the year

• Google suspended 39.2 million malicious advertisers in 2024 thanks to AI

📄 ReTool: Reinforcement Learning for Strategic Tool Use in LLMs

📄 Cobra: Efficient Line Art COlorization with BRoAder References ​​

📄 Pixel-SAIL: Single Transformer For Pixel-Grounded Understanding

📄 CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training  

📄 ColorBench: Can VLMs See and Understand the Colorful World? A Comprehensive Benchmark for Color Perception, Reasoning, and Robustness

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!