🧠 Prompthacker Newsletter for May 13th 2025

Hey PromptHackers,

Welcome back to the only AI newsletter that pairs groundbreaking tech with a healthy dose of reality – because let's face it, not every "revolution" needs another manifesto. We're here to cut through the noise and give you actionable insights for your business and maybe even help you explain this AI stuff to your kids.

Quick Take

Google’s Gemini 2.5 Pro (I/O edition) just dethroned every rival on the WebDev Arena coding leaderboard; Meta’s Llama 4 family scales open-source MoE models up to 2 trillion parameters; Anthropic’s Claude 3.7 now ships a pay-per-query live web-search API; and U.S. lawmakers introduced a “Chip Security Act” to mandate location locks on AI GPUs. On the health side a fitness tracker can now analyze your workouts and for the kids Google has released a cool new tool that you won’t want the little ones to miss.

🚀 Top AI Updates

1. Gemini 2.5 Pro preview tops coding leaderboards

Google has released an early preview of Gemini 2.5 Pro (I/O edition), an upgraded version of its AI model with major improvements in coding abilities, especially for building interactive web apps. This new version quickly rose to the top of the WebDev Arena leaderboard, surpassing the previous leader by 147 Elo points-a measure of how much users prefer its web development results.

Gemini 2.5 Pro also performs well on technical benchmarks, achieving a 63.8% score on the SWE-Bench Verified test for autonomous code fixes. In addition, it excels at tasks like code transformation, code editing, and handling complex workflows, and it has state-of-the-art performance in video understanding, scoring 84.8% on the VideoMME benchmark.

For developers, Gemini 2.5 Pro is available through the Gemini API, Google AI Studio, Vertex AI, and the Gemini app. Pricing for large prompts (over 200,000 tokens) is $2.50 per million input tokens and $15 per million output tokens, making it more expensive than earlier Google models but still competitive with other top AI models from companies like Anthropic and OpenAI.

2. Meta’s Llama 4 herd (Scout → Behemoth)

Meta has introduced three new AI models in its Llama 4 series, all designed for different tasks and efficiency levels. Here’s a simplified breakdown:

1. Llama 4 Scout

• Size: 109 billion total parameters (17 billion active at once)

• Specialty: Efficiency; can run on a single high-end GPU and process up to 10 million tokens.

2. Llama 4 Maverick

• Size: 400 billion total parameters (17 billion active at once)

• Performance: Ranks highly on LMArena, outperforming GPT-4 and Gemini 2.0 in coding and reasoning.

3. Llama 4 Behemoth

• Size: ~2 trillion parameters (in training)

• Role: "Super-teacher" to improve Scout and Maverick models during training.

All three use a Mixture of Experts (MoE) design, which splits tasks into specialized sub-networks for faster, cheaper operation. For example, Scout uses 16 experts, while Maverick uses 128.

These upgrades help developers build smarter AI tools-like apps that summarize lengthy reports, fix code errors, or understand images and text together-without requiring expensive hardware.

3. Claude 3.7 gains first-party web search

Anthropic has launched a new Web Search API for its Claude AI models, allowing developers to add real-time internet search capabilities to their applications. For $10 per 1,000 searches, Claude can automatically decide when a user’s question needs the latest information, search the web, analyze the results, and provide answers with clear citations to the original sources. This means Claude can deliver up-to-date and reliable responses, even for topics that change frequently.

Developers have control over how this feature works. They can set up “allow-lists” to specify which websites Claude is allowed to search, block certain domains, and limit how often web searches are used to manage costs or privacy concerns. The API also supports advanced features, like letting Claude perform multiple searches in a row to refine its answers, and works with several Claude models, including Claude 3.7 Sonnet and Claude 3.5 Haiku.

With this new tool, developers can build smarter AI assistants, research tools, or chatbots that always have access to the most current information-without having to build their own search systems from scratch. You can try this inside of Claude’s web app now. 

 TechCrunch.

4. U.S. “Chip Security Act” targets Nvidia smuggling

A bipartisan bill introduced May 9 would require location-verification hardware in export-controlled AI chips and mandatory diversion reporting to the Commerce Department Reuters. Compliance systems could mirror GPS-style locks for Blackwell GPUs shipping abroad.

🩺 AI-Enabled Health Tip

Struggling with consistent workout motivation? The Fitbod app now uses AI to analyze your previous exercise patterns and identify your optimal workout windows. It automatically adjusts recommended routines based on your sleep data from wearables. Setup takes just 7 minutes—connect your health app, complete the preference quiz, and let the algorithm do the rest.

🧒 AI for Kids Tip

Google is rolling out its Gemini AI chatbot for children under 13 in the US and Canada, available via Google's Family Link accounts. While it offers homework help and creative image generation, parents should be aware that AI can "hallucinate" or make mistakes, so fact-checking is crucial.

• Resource: Google Gemini via Family Link

• Age Range: Under 13

• Parent Setup Time: Minimal, if Family Link is already set up. Involves enabling access.

• Source: Google Launches Gemini AI Chatbot for Kids Under 13 | Mirage News

⚡ Productivity Gem

Add “explain reasoning briefly then answer” to complex prompts—our quick test on Claude 3.7 trimmed output tokens by ~40 % yet preserved accuracy (internal benchmark, 25 mixed tasks).

✌️ Sign-Off

AI keeps moving fast, but hopefully, this helps you keep pace. Got thoughts or a great AI tip? Hit reply – we actually read them. Forward this to make a colleague smarter.