Welcome back. When SpaceX bought Cursor, every major engineering team asked the same question: what happens when the world's largest training cluster meets the hottest coding startup? We finally have our answer. Grok 4.5 has arrived — and Elon's calling it "Opus-class" at a fraction of the price.

Also: Developer toolkit you need for the agents era, how to automate your local AI setup, and Uber’s AI adoption playbook.

SpaceXAI's new model undercuts Opus on speed and price: Elon Musk's rebranded AI lab just dropped Grok 4.5, which was co-trained with Cursor for coding and agentic work. Internal benchmarks show it's already beating Claude Opus 4.8 on several coding tests. It finishes tasks in less than half the steps for a fraction of the cost. It’s now the default model for Grok Build and is available to all Cursor users. Try it here.

OpenAI gives voice AI a second brain for deeper work: The ChatGPT maker just released GPT-Live, a full-duplex voice model that listens and speaks at the same time. When a question needs web search or heavier reasoning, it hands the task to GPT-5.5 behind the scenes and keeps talking while it waits for the result. It's rolling out to ChatGPT users globally now. Developers can sign up for early API access here.

Cognition trains a frontier coding model on a budget: The team behind Devin just rolled out SWE-1.7, post-trained from Kimi K2.7 inside its agent harness. It matches GPT-5.5 within a point on Cognition's FrontierCode benchmark at roughly $2 per task, though Opus 4.8 stays slightly ahead. The model can summarize its progress and pick up right where it left off. This process lets coding sessions run for up to six hours. You can use it in Devin today.

An open-source code reviewer just outranked CodeRabbit, Copilot, and Devin. PR-AF places #2 of 42 on Code-Review-Bench running one open model.

The open-source harness plans a custom review per PR, runs reviewer agents in parallel, and verifies every finding against your source before posting inline comments. About 10x cheaper per review than closed-source tools. Self-hosted, one API call, Fork and deploy!

Star & Deploy

A job market running on crossed wires. Veteran engineer Gergely Orosz spent the first half of 2026 speaking with over 50 hiring managers and engineers, and his verdict went viral. Companies are struggling to fill senior roles, yet the very engineers they need can't even land an interview. It feels like a downturn, but software engineering openings are still rising.

The channel broke, not the market. One Seattle startup pulled 800 strong resumes for a single role in three months, many from Big Tech alumni, and they're going unread. AI-polished applications all sound identical, and some candidates turn out to be completely fake, so many hiring managers have quietly stopped checking inbound altogether. Talent and demand are both there. The pipe between them is dead.

Everything rerouted to referrals. Nearly every senior hire in Orosz's reporting came through personal networks or direct outreach. Most companies already run formal referral programs, yet one study found only 2% say theirs actually hit their hiring goals. The industry bet its pipeline on its worst-maintained channel.

Rebuild for the AI era. Two engineers who are building an interview-screening startup studied the problem for an Harvard Business Review playbook. After interviewing 120 recruiting leaders and analyzing thousands of screening calls, the advice is clear. Screen for live reasoning and judgment, because polished credentials are now way too easy to fake. For a broader look at the market, Orosz's field notes cover the rest.

Anyone can make an AI demo. The hard part is shipping it, monitoring it, and knowing when it breaks (and how to fix it).

Datadog's Developer Toolkit for the AI Era gives you the playbook for building, deploying, and observing AI applications at scale. It covers: CI pipelines, LLM observability, feature flags, testing, and modern AI delivery workflows.

Get the toolkit and ship with confidence.

Side tasks can be a bit clunky in Claude Code. If you run them in the main thread, they end up burying your context. If you hand them off to a subagent, you're stuck spending three prompts just getting it up to speed.

Delba de Oliveira, an engineer on Anthropic's Claude Code team, shared the fix: the /fork command. It spawns a subagent that inherits the entire conversation, works in the background, and reports back once it's done.

Since the fork is an exact match of the parent, its first request reuses the prompt cache, making it cheaper than starting a fresh subagent.

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

5 CLI tools that changed how a senior dev works: You'll learn how to boost your command-line productivity with 5 essential CLI tools. This tutorial covers everything from searching files with fzf and managing multiple terminal windows via tmux, to using advanced SSH port forwarding, securely storing passwords with pass, and integrating AI into your workflows using fabric.

Mosaic: This tool brings multiplayer support to Claude Code, allowing multiple users to share a terminal while connecting their Claude instances to communicate directly.

CubeSandbox (9.2K ⭐): A fast, secure, lightweight sandbox for AI agents. It scales easily, works with the E2B SDK, and launches hardware-isolated environments in under 60ms with minimal memory use.

Databricks benchmarks coding agents on internal tasks: Databricks benchmarked AI coding agents to find the best balance between performance versus cost for their engineers. They found that affordable models manage daily tasks effectively, though the testing environment significantly affects total efficiency and spend.

A Meta engineer is asked weekly if it's too late to break into tech. He shared 50 things he'd do if he were starting over today, without any of the doom and gloom.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ 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