Google drops a double release for builders

Welcome back. Last week we explored Opus 5.5 (link below), and developers are already pushing it further. One dev used it to clone four Adobe apps, and the demos are sparking a fierce online debate. 

Meanwhile, Google has unveiled two new features for your tech stack, and French startup Mistral aims to take on DeepSeek and Qwen with its latest AI model. 

Also: Prompt engineering lesson from the creator of Claude Code, the most paid engineering skill is shifting in 2026, and a learning system for people with multiple interests. 

Google unveils two major updates for your tech stack: The search giant has two new models for your next project. First up, EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video into embeddings locally, letting you build search and RAG pipelines that work offline. You can use this at your organization for internal knowledge-based tools.You can start building here. On the creative side, Nano Banana 2.1 promises better visual design, mask-based editing, and more consistent subjects across images, according to Google.

Devs use Anthropic's new model to clone Adobe apps: Opus 5.5 has been out for just two weeks, and one developer has already taken a crack at Adobe's creative suite. Their post shows off PhotoCraft, VectorCraft, FilmCraft, and LightCraft, which are open-source Rust takes on Photoshop, Illustrator, Premiere, and Lightroom. They came together fast, though plenty of features are still missing. Explore the repos.

Mistral's best AI model yet takes on DeepSeek and Qwen: The French AI startup just debuted a preview of Mistral Large 4. The 1T-parameter model excels at cybersecurity, coding, manufacturing, finance, and multimodal tasks. calls Large 4, the top open-weight model from the US or Europe, claiming it beats DeepSeek V4 Pro and Qwen3.8-Max at agentic coding and trails only Claude Opus 5 in a blind code-quality test. Self-hosting weights arrive by month’s end, but you can try the preview API now.

Most voice AI still works best in English.

Soniox brings speech-to-text, text-to-speech, and translation together in one real-time platform across 60+ languages, built for accurate input, natural output, and seamless multilingual conversation.

Powering real voice products at Perplexity, Samsung, Wispr Flow, Vapi, Retell, and more.

Build with Soniox

Trust went nowhere. Engineering teams that stopped reviewing diffs failed to put any verification in place to replace that scrutiny. Because AI models do not reliably generate consistent outputs, teams must establish dependable safeguards around them. In a new essay, Anthropic engineer Addy Osmani highlights where the industry is heading and he identifies two main costs in this process:

If teams skip the initial checks, they save time upfront, but resolving downstream issues later requires far more effort.

The working version is boring. At Anthropic, an automated reviewer runs on nearly every PR. Engineers mark fewer than 1% of its findings as incorrect, and substantive review coverage rose from 16% to 54%. Based upon his experience, Osmani recommends a code review workflow built around the following principles: 

Price pressure has a cost too. Some teams claim they cannot stay competitive without blindly shipping code. But Osmani points out the heavy cleanup costs that follow. When you skip basic checks, you only push failures downstream, where they quickly snowball into production and harm your customers. The best teams recognize which diffs require deep human review, and they always empower someone to step in and say no. 

Try it yourself. If you want to add AI review to your own pipeline, Anthropic's code review docs cover their setup.

That's how long agents spent in TinyFish browsers in September, and the stack held.

Add $50, get $65 (30% extra) for agent runs through October 31.

Claim 30% Bonus

How to build a Claude DevOps agent: This tutorial shows how to build an AI SRE with the Claude Agent SDK that acts as your first responder when prod breaks. It digs through logs, metrics, and recent deployments, figures out what probably went wrong, and recommends what to do next while leaving the final call to you.

VMPal: A virtual machine built specifically for Apple silicon and AI agents. It runs macOS, Windows, and Linux with GPU acceleration, adds a Liquid Glass UI, and includes MCP support so agents can work either on your Mac or inside the VM.

How to run Claude Code in the cloud (by Anthropic): This field guide shows when to move coding work off your laptop and into separate cloud VMs. You’ll learn how to run tasks in parallel, let Claude keep working while you’re away, hand CI failures back to it, and move sessions between the cloud and your terminal.

Every line in AGENTS.md loads into context on every turn, and most files are stuffed with no-ops the agent ignores. A senior dev shared a prompt that cuts the waste.

Three subagents each produce a restructuring, from conservative to aggressive, merged into one PR. Review the diff, keep the version that fits.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Our dev team put together a deep dive on how developers can get the most out of Opus 5.5 across different projects. Check it out.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

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

Until next time — The Code team