Welcome back. Anthropic’s new mini model packs a mega punch. Claude Haiku 4.5 is here for devs, and you get close to Sonnet level performance at an incredibly faster speed and less than half the price. And OpenAI isn’t far behind — their new web search model is now 60% cheaper.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Anthropic launches mini model with frontier-level coding at half the cost: Claude Haiku 4.5 is a smaller and faster model that delivers coding performance comparable to Claude Sonnet 4 — but at one-third the cost and more than twice the speed. Developers can also orchestrate multiple Haiku 4.5 instances in parallel for complex multi-agent projects. Here’s a quick guide on how to leverage it.

OpenAI launches cheaper web search model for developers: The ChatGPT maker just rolled out an updated web search model in Chat Completions that's 60% cheaper than its predecessor at $10 per 1,000 calls. The new gpt-5-search-api includes domain filtering capabilities similar to the web search feature in the Responses API, making it easier for developers to limit search results to specific trusted sources.

Alibaba unveils GPT-5 Nano rival: Alibaba introduced Qwen3-VL 4B and 8B, compact and dense vision-language models available in Instruct and Thinking variants. Despite their size, they outperform models like Gemini 2.5 Flash Lite and GPT-5 Nano, and often beat them on benchmarks spanning STEM, VQA, OCR, video understanding, agent tasks, and more.

How Box is building an AI-first company—without cutting jobs: Box CEO Aaron Levie made a bold move almost overnight, transforming his 2,000-person company into an AI-first operation. But unlike other CEOs predicting mass layoffs, Levie's approach focuses on doing more, not cutting headcount.

Where developers actually want AI to support their work: New research from Oregon State University surveyed 860 developers to map exactly where AI tools fit into their workflow. The findings? Developers embrace AI for high-demand grunt work like documentation and environment setup, but strongly resist automation in mentoring and strategic planning.

Essential software engineering practices: Most AI coding failures aren't technical bugs, they're architectural disasters waiting to happen. A new guide from builder Jenny Ouyang reveals AI's core optimization bias: it favors prompt completion over system health, assumes perfect conditions, and ignores consequences.

Building Live Voice Agents with Google’s ADK: A new free course from DeepLearning AI teaches developers how to use Google's Agent Development Kit to build conversational AI agents that can handle natural, real-time voice interactions.

The 'no-BS' guide to Agentic Coding: A senior developer's brutally honest workflow post is going viral across the internet. He runs up to 8 instances of GPT-5-Codex simultaneously in terminal windows, lets them make atomic Git commits.

Building ChatGPT apps with Next.js: Vercel released an in-depth tutorial teaching developers how to integrate Next.js applications with ChatGPT's Apps SDK using the Model Context Protocol.

Coral NPU: Google launches Coral NPU is a full-stack, open-source, AI-first hardware platform designed to overcome performance and privacy limits in edge AI. It delivers ultra-low-power GOPS performance for always-on ambient sensing and enables the first open, standards-based NPU to bring LLMs to wearables.

Do LLMs Really Need 10+ Thoughts for "Find the Time 1000 Days Later"?: This Google DeepMind paper shows that long step by step thinking often wastes compute on easy questions. The paper makes overthinking measurable and stoppable. Shows where time is wasted and gives clear rules to stop early without losing accuracy.

Recursive Language Models: A new inference strategy where LLMs can decompose and recursively interact with input prompts of seemingly unbounded length, as a REPL environment.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team