Welcome back. Every ChatGPT user has the same complaint: it talks a bit too much. OpenAI finally fixed that problem with its latest model update, along with other upgrades. Plus: A new company is going viral for claiming they solved AI’s biggest bottleneck.

Also: How to run Claude Code locally on 24GB RAM, find the 5 AI engineering courses worth taking, and why the Coinbase CEO just laid off 14% of staff.

OpenAI rolls out a smarter default model for ChatGPT: The ChatGPT creator has officially released GPT-5.5 Instant as the new standard, replacing GPT-5.3 Instant across the board. This update significantly slashes hallucinations in critical fields like law and medicine while boosting math and multimodal reasoning. For developers, the GPT-5.5 model will be available through API as “chat-latest,” with 5.3 available as an option for paid users for only three months.

Subquadratic unveils a new model with a major breakthrough in LLM intelligence: The AI lab just launched SubQ, the first-ever sub-quadratic frontier model that runs on 1,000X less compute than its rivals. It packs a massive 12M token context window, which could theoretically keep agents on task for weeks without losing their edge. Developers can now plug it into Claude Code, Codex, and Cursor to map codebases, gather context, and tackle token-heavy questions faster.

Google ships multimodal file search and a faster Gemma 4: The search giant just dropped a multimodal update for Gemini API File Search using the new Embedding 2 model, letting developers query text and images together. The update integrates retrieval directly into the API, cutting out the need for separate vector databases and offering inline citations with 1 GB of free storage. Plus, new multi-token prediction drafters make Gemma 4 up to three times faster.

Ready to move past novelty chatbots? Google Cloud’s Startup School: Agentic AI covers the architecture you need to build robust, autonomous workflows.

Dive into Realtime Voice AI, enhanced Multimodal RAG, and bidirectional Vision Agents. Go from prototype to secure production without missing a beat.

Save your spot

The hidden cost. Anthropic dropped SKILL.md as an open standard last December, and coding agents adopted it almost instantly. Now, engineers everywhere are writing these files. But most are missing a critical piece. AWS engineer Lakshmanan Meiyappan just published a breakdown arguing that most people fundamentally misunderstand what they're actually building.

The 3x rewrite. Meiyappan saw this firsthand on his first attempt. He wrote one massive, sprawling file that worked, but it ate up 20% of the context window before the agent even got started. He ended up rewriting those same instructions as a lean "spine" that pointed to three reference files, and the cost dropped to 7%. The content stayed the same; the architecture was just better.

A loader in disguise. What really changed was how the system processes the information. Anthropic's runtime loads SKILL.md files in stages: it scans metadata with every turn, pulls in the main body only when a skill triggers, and waits to access references until they are specifically called. Most engineers treat SKILL.md like a single massive prompt, but the runtime just isn't built to work that way.

The upgrade trap. Meiyappan hit another roadblock when he upgraded to a more powerful model: instead of improving, his writing skill actually tanked. The “smarter” model took instructions way too literally and stopped using its own judgment. This is a perfect example of why skills tuned for one model don't always carry over to the next. You need to re-run evals every time you switch models. Think of every skill as its own small program that needs testing.

Most AI agents are built for engineers. Viktor is the agent for everyone else. Operations, finance, marketing, sales, support.

One Slack message and Viktor pulls from your tools, does the work, ships the output. 3,000+ integrations. SOC 2 compliant. Teams on Viktor build what used to take a developer, an analyst, and a week.

Get Started for Free

Many developers have accidentally committed a hardcoded API key or a “console.log” they forgot to remove (we’ll admit we might have done it once or twice ourselves). You usually catch it the next morning during review, then waste an hour rotating keys and scrubbing your Git history.

A simple three-line pre-commit hook lets Claude scan your diffs first. Save this code to “.git/hooks/pre-commit” and run “chmod +x .git/hooks/pre-commit” to set it up:

Every “git commit” now sends your staged changes straight to Claude. The review pops up right before the commit finishes, giving you a chance to hit “Ctrl+C” if Claude catches an actual issue.

P.S. You can find 50+ AI coding hacks here.

How to do multi-agent orchestration in Claude Code: This tutorial covers building a multi-agent system using an orchestration layer to streamline coordination and workflow structure. You'll learn how to design cleaner, more scalable AI systems that handle complex tasks without the chaos of multiple agents clashing.

ProxyUser: Continuously tests your app’s critical user flows. It turns plain-English descriptions like “user signs up and creates a widget” into automated monitoring that alerts you only when something truly breaks.

Files-to-prompt (2.7k ⭐): A command-line tool that combines all files in a directory into one formatted prompt for LLMs.

Building reliable agents with memory and compaction (by OpenAI): Long AI investigations often fail by hitting context limits and forgetting past workflow rules. Using "Compaction" to shrink active memory and "Memory" to save reusable habits creates reliable agents that smoothly process evolving evidence into accurate reports.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 260K+ engineers and 150K+ followers on socials. Get in touch.

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

Until next time — The Code team