OpenAI just gave devs two big shortcuts

Welcome back. We've got a ton of exciting updates for you this Friday. There's plenty to dig into over the weekend, from OpenAI's new voice model to fresh releases from Cognition and DeepSeek that solves bugs at a fraction of the cost. Let's dive in. 

Also: The creator of Claude Code shared his anti-slop playbook, Google’s revamped SWE interviews, and an argument against “AI is ending humanity.”

OpenAI just gave developers two big shortcuts: The ChatGPT maker just launched GPT-Live-1 to simplify voice apps. It handles listening and speaking natively in a single model, deciding in real time when to pause, interrupt, call a tool, or hand off complex reasoning. Meanwhile, the new Agents API brings the Codex harness into a single call, managing subagents, tools, and context automatically. You can run it inside OpenAI's sandbox, your own setup, or through partners like Vercel and Cloudflare. 

Cognition’s new model matches frontier labs at a fraction of the cost: The Devin maker just debuted SWE-2, its most advanced coding model yet. On the FrontierCode benchmark, which measures code quality rather than just raw accuracy, it beats Grok 4.6 and matches the performance of GPT 5.6 Sol and Fable 5.1 at a fraction of the cost. The real hook is the price tag, the model delivers almost identical results for far less money with fewer wasted steps. Plus, a new training approach sharpens performance across different reasoning levels in a single run. Try it here.

DeepSeek open-sources a leaner agentic model: The Chinese AI lab just dropped V4.1-Flash, a leaner model that slashes agent running costs. By cutting back on compute and memory compared to the previous version, it lets teams tackle the same workloads without racking up massive API bills ($0.30 per million input tokens). On OpenDesign benchmarks, it reached 98% of GPT-6 Astra’s score at 1.4% of the cost on everyday design tasks based on user requests. It is also beating models like Claude Opus 5 while slashing costs down to $1.80 for fixing 24 real bugs. Try it on Hugging Face.

A lot of AI tools guess. You.com grounds every answer in real, cited sources—so you get accurate results, instead of confident-sounding fiction. Web search, research, and get work done with an AI platform built for trust.

New users get 100 free credits to try it out.

Sign up now →

Past the hand-wringing. Enough has been said about agents writing the code, how review has become the bottleneck, and why human judgment still matters. Now in his latest deep dive, veteran engineer Gergely Orosz looks at what teams are actually doing about it. The standout strategy he found was reviewing by “blast radius.” The logic is simple, you just need to prioritize reviews based on how much damage a change could actually do if it goes sideways.

Put humans where mistakes cost the most. OpenAI and Anthropic already run versions of this. Their low-risk changes often rely on AI review, while anything risky still gets a human sign-off. Also, the potential impact of a change, rather than its size, determines who reviews it. 

One small team went further. Orosz shares an example of Duckbill Group, a cloud and AI contracts management company. Its five-person engineering team hit 60 open PRs, then rebuilt the process around risk. Here’s how it works:

With that in place, weekly merges jumped from 80 to 154. Low-risk PRs merged in about an hour, versus 26 hours with a human reviewing.

Draw the line first. Decide what needs a human eye and let the rest flow through. The review time you save is better spent nailing down the plan and tests before the agent starts.

P.S. If you need a step-by-step guide to set up risk-based code review on your own team, check out this breakdown by Cortex.

Grok Bot is your team of AI teammates that get work done. Learn how to get the most out of Grok Bot at this three day event.

See live demos featuring the team behind the product and discover best practices for customizing your bots based on your role and department.

Register now

Moving from Mac to Linux for AI development: If you’re running multiple agents at once, Linux can give you a faster, more reliable setup without tying everything to your laptop. This tutorial shows how to offload long-running jobs to a remote Linux machine, keep them running even after you close your Mac, and connect everything with tools like Tailscale and tmux. It also covers the less-fun bits, including syncing context, files, and secrets across machines.

I-have-adhd (39.5k ⭐): Stop coding agents from burying the answer under walls of text. It forces them to lead with the next action, number multi-step tasks, suppress tangents, cap long lists, and end with one concrete next step.

Building low-latency voice agents (by Google): Learn how to make voice agents feel a lot less sluggish. Instead of waiting for someone to finish speaking, Gemini 3.5 Transcribe Live streams the conversation in real time, so the agent can start figuring out what to do in the background. The guide walks through a debate agent, but the same setup works for things like call centers, meeting assistants, and other live voice apps.

When a long Codex session hits its context limit, it compacts everything into one summary. The reason you made a decision three days ago gets squished out.

A developer shared a config flag that fixes this for GPT-6 Astra.

Codex now writes its own notes across context windows and searches earlier messages and tool results when it needs a detail.

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

Remember the Anthropic researcher who quit, warning AI could end humanity? A current safety lead responded, and his take is worth reading.

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