Welcome back. Last year, a group led by OpenAI researchers dropped "AI 2027," a roadmap predicting what AI would look like over the next two years. Fast forward to today, and 19 of the 53 tracked predictions are confirmed or ahead of schedule. Check out the full timeline to see what’s changed.

Also: A detailed guide on harness engineering, fixing your home wifi with Codex, and how to humanize AI PRs of your dev team.

SpaceXAI’s new model targets long-running agents and large codebases: Elon Musk's AI lab just unveiled Grok 4.6, a model specifically designed to handle multi-step workflows. Whether it researches a codebase, structures an app, or iterates based on feedback, this model performs reliably over the long haul. It’s currently available via Cursor, Grok Build, and the API at a rate of $2 per 1M input tokens. Devs are already giving this model a shot. Shopify CEO’s post about the model's animated reaction during a maintenance task is making rounds on the internet. 

Alibaba open-sources its most powerful model to date: The Chinese tech giant just dropped a new version of Qwen3.8 on Hugging Face. The company claims the model goes head to head with the top frontier models on coding and agentic benchmarks. But there is a catch for devs: you will need a massive 4.9TB of storage to run it at full precision. Even with Unsloth’s 1-bit build (which trims that down to 397GB), you are still going to need some serious RAM. If you want to run this locally, a lighter 27B version will drop soon. More details here.

DeepSeek's flagship nears frontier coding at a fraction of the cost: The Chinese AI startup just moved V4 Pro from preview to general availability while maintaining its current pricing. It costs $0.435 per million input tokens and $0.87 per million output tokens, which represents a small fraction of the Fable 5 rate. The model supports a 1M token context window and tool calling, allowing it to integrate seamlessly into your agent pipelines. While it still trails slightly behind Fable 5 in agentic coding performance, these results come from the company's own unverified benchmarks. 

An open-source reviewer now places #2 of 42 on Martian's Code-Review-Bench, ahead of the commercial tools, at ~10x lower cost per review.

PR-AF plans the review per PR, runs reviewer agents in parallel, and verifies every finding against your source before it posts, so no hallucinated comments. 0.706 recall, roughly 3x more valid findings. Apache 2.0, self-hosted, runs on any open or closed model.

Star & Deploy

The wall everyone hits. Agentic coding can help your engineering team ship up to 10x faster, but the costs can spiral even quicker. If you aren't careful, your AI bill will eventually outweigh the actual value it delivers. Databricks calls this a solvable engineering problem. According to their internal playbook, teams that successfully manage these costs lean on four main strategies, ranked here by impact: 

The starting point. Combine these levers, and unit costs can fall by as much as 90% without a quality hit. But you should start small. First, look at your team's spending, identify the models and tools driving the bill, and then tackle the biggest source first.

P.S. Databricks' full walkthrough covers the tooling, while its open-source Omnigent meta-harness lets teams experiment with model switching themselves.

Soniox Text-to-Speech lets you create extraordinary voice experiences in 60+ languages with expressive control, exceptional precision, instant voice cloning, and ultra-low-latency streaming.

Direct whispers, laughter, hesitation, excitement, and emotion, or precisely speak phone numbers, email addresses, technical terminology, and multilingual text.

Built for voice agents, global products, customer support, media, and any application where speech needs to sound human and get every detail right.

$0.70 per generated hour.

Try it now

How to stop overpaying for agents by using Prompt Caching: You’ll learn how to build cache-friendly agent harnesses that keep prompts reusable and slash token spend. This tutorial covers how to avoid costly cache invalidation and make long-running agent workflows significantly cheaper to run across all major providers.

Omg.dev: You can now run coding agents like Claude Code or Codex locally on your machine and control them from your phone.

Field guide on Grok 4.6 (by a Cursor engineer): AI models often struggle with complex tasks when they rely on long prompts or assume the work is done without checking the results. This guide solves that by using concise Grok 4.6 prompts paired with strict verification steps, ensuring the AI actively tests and fixes its own output.

Claude Code performance tends to dip during long sessions. If you try a full refactor in one go, the second half usually ends up much sloppier than the first.

To fix this, a senior engineer shared a sizing rule using his /to-tickets command. It breaks a large spec into smaller tickets that each fit perfectly within a single session. 

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

Check out this GitHub repo to remove Claude’s watermarks from your content.

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