ChatGPT co-inventor’s frontier lab drops new model

Welcome back. One of ChatGPT's co-creators is back with a new project and a big claim: A new type of AI model that's 200x faster with 0 hallucinations. Speaking of OpenAI, the ChatGPT maker wrestled back top spot on developer spend after 2 years of Anthropic dominance.

Also: An AI agent that tried to fund itself, an app to keep track of your Codex limit, and a Meta engineer’s reality check on the AI race.

ChatGPT co-inventor bets on models that can't hallucinate: The maker of ChatGPT Diogo Almeida, thinks frontier models need too much human oversight for automation. After raising $40M, his AI lab emerged from stealth with Jev, a model trained on a new method called RLCD (Reinforcement Learning for Calibrated Decisions). While this model cannot write code or generate prose, it can be mainly used to classify, rank, route tasks, pick tools or agents, and verify LLM outputs. Almeida claims it delivers responses as fast as 70 milliseconds, eliminates hallucinations, and runs up to 100x faster and cheaper than rivals. Check out the model here.

OpenAI just passed Anthropic in developer spending: For the first time in over two and a half years, OpenRouter users spent more money on OpenAI than on Anthropic last week, driven by surges in GPT-5.6 and GPT-6 Astra usage. Now, OpenAI is pushing teams even further into the new lineup. GPT-5.5 leaves Codex, ChatGPT, and Work on October 14, steering developers toward GPT-5.6 Sol or Astra. If your stack still relies on older benchmarks, it’s probably time to test your defaults again. 

Google collapses your voice stack into one model: The search giant just dropped Gemini 3.8 Live and 3.8 Live Extended Thinking to unify your voice AI stack. Now, you can build real-time voice apps without stitching together separate models for transcription, reasoning, and speech. Through the Gemini API and Google AI Studio, you can power conversations across 97 languages while background tools run without interrupting the audio stream. The same API also adds speech translation, text-to-speech, and music generation. Start here. 

AI agents are moving beyond writing code. They're planning, testing, reviewing, and increasingly taking on more of the software delivery lifecycle.

At {unscripted} Virtual, leaders from Anthropic, SpaceXAI, Workday, Google, and more will explore where agents can safely act, what humans should still own, and how teams can increase autonomy without sacrificing security, governance, or developer velocity.

September 30 | Virtual | 10:00 AM PT / CEST

Register now

We've all seen agents pull and dump a ton of irrelevant noise we never asked for. You point an agent at a GitHub repo to ask what happened, and it'll happily dump everything on you: every PR, every diff, every minor change, with zero priority ranking. 

Noise in, noise out. The real engineering challenge is teaching the agent what actually matters. James Samuel, an engineering manager at Reddit, just shared a clever fix to this. His setup runs on three markdown files and one scheduled task.

The ground rules. Samuel only saves something if it clears a simple bar: does it drive a future action, spot a trend, or help a teammate? Everything worth keeping goes into a Logseq (a note-taking app) detailing people and the projects he cares about. He applies that same filter today but lets an agent do the heavy lifting. 

Three files, three jobs. Before the agent starts searching, three markdown files tell it what to care about:

The 8 AM drop. Every workday, a scheduled Codex task runs daily.md. It ranks the important stuff, links every claim back to a source, removes duplicates across tools, and writes a Logseq-ready summary before the day starts.

The complete playbook. Part 1 explains what to capture and why. Part 2 ships the agent skills and the full daily.md prompt, ready to fork. It’s worth reading in order.

Coding faster is NOT-EQUAL-TO releasing faster.

That's because your SDLC is now legacy.

The teams winning with AI have agentified the full delivery pipeline. They have adopted ADLC.

In 90 minutes, see how to reimagine the SDLC and make engineering move at the speed of business ambition.

If you are an engineering leader with a mandate to ship more, this masterclass is for you.

Book your seat

A senior dev’s complete agentic setup: This tutorial shows a practical setup for weaving AI agents into both coding and admin work. You’ll learn how to use Codex as the main interface, how tools are connected behind it, and how longer-running jobs get handed off without constant babysitting.

Cline Desktop: Run coding agents on your machine with open-weight models. You can import tasks from Claude Code or Codex, schedule recurring PR reviews and security scans, plug in MCP servers, and use voice or web search.

How to keep Codex working from your phone (by OpenAI): This guide shows how to use Remote in the ChatGPT mobile app to keep engineering work moving when you’re away from your desk. You can start tasks, steer agents, review code, queue prompts, manage worktrees, and jump in only when a decision needs you.

A dev’s Codex setup went viral for cutting agent costs. Instead of running every subagent at max effort, the tree assigns effort levels chosen by weighing DeepSWE pass rates, average cost per task, and agent steps.

Astra at medium orchestrates, delegates, then integrates and verifies. Luna at max investigates, Sol at high implements and tests. The xhigh review only spawns when needed.

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

A founder just turned his bestselling book into code. The Minimalist Entrepreneur now runs as 9 Claude Code skills that guide you through building a startup step by step.

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