Welcome back. For two weeks, the strongest agentic coding model on the market was locked behind a government approval process. Not anymore. GPT-5.6 is now available to everyone, and they didn't ship the model alone.

Also: How a senior engineer uses GPT-5.6, a checklist for frontend optimization, and ex-Google director’s advice for engineers.

OpenAI unveils its most efficient coding models yet: The ChatGPT maker just dropped GPT-5.6 in three tiers named Sol, Terra, and Luna. OpenAI claims flagship Sol leads the Artificial Analysis Coding Agent Index with a score of 80 while using less than half the output tokens of Claude Fable 5. All three models are live in the API with Programmatic Tool Calling and a parallel-subagent ultra mode. The models also power ChatGPT Work, a new agent with Codex built in that runs across your apps for hours, delivering finished sheets, slides, docs, and web apps.

Meta debuts its first paid AI model for agentic tasks: Zuckerberg's AI lab just shipped Muse Spark 1.1 along with the Meta Model API. This marks a shift from their usual open-weight Llama releases, moving to a hosted model with per-token pricing. The update is built for long agentic tasks. It handles planning across parallel subagents and manages a massive 1M-token context to help big projects finish faster. You can check out the API right here.

Ex-GitHub CEO's startup gives AI agents faster Git: The new venture, named Entire, just released its distributed Git network platform. Your code stays on GitHub, but agents can now clone and pull from a nearby mirrored copy. Founder Thomas Dohmke says this setup eliminates the rate limits and latency issues caused by billions of clones hitting a single server. It works with Claude Code, Cursor, Codex, and Copilot. See how it works.

You already know AI can write code — the hard part is trusting what it tells you.

You. com pairs coding help with grounded, cited search, so you're debugging real answers, not made-up APIs and phantom docs.

Try it free: new users get 100 credits, no strings attached.

Sign up now →

A benchmark from real PRs. Databricks created a private benchmark by using real tasks their engineers completed in their own codebase. They tested various AI agents on these tasks and graded them using the original pull request tests. The most surprising takeaway: open-source GLM 5.2 performed just as well as Opus 4.8 but cost about 30% less per task. 

But the bigger gap. Every model runs inside a harness like Claude Code or Pi. Databricks found that using different harnesses with the same model can double the cost with no meaningful change in quality (see the above image). The efficient harness sent less context each turn and required fewer runs to complete the job.

Sticker prices mislead too. In this benchmark, Sonnet 5 actually costs $2.09 per task, while Opus costs only $1.94. Even though Sonnet 5 has lower per-token rates, it ends up being more expensive per completed task. This is because Sonnet 5 takes longer and re-reads more context to get the job done. Ultimately, the only metric that matters is “how much it costs to get a code change that passes your tests.”

Your repo is already a benchmark. Databricks points out that teams merging PRs are sitting on a goldmine of evaluation data. Merged PRs capture real intent, and the ones with solid tests come with graders built in. No model has trained on them. The post outlines a step-by-step method: pick the right PRs, rewrite their intent into prompts, and hold out the tests for scoring.

Senior engineer’s updated agent skills for an improved dev workflow: In this tutorial, former Vercel engineer Matt Pocock shows you how to optimize your dev process using v1.1 of the skills repo. You'll find major updates like renamed flow skills (/to-spec, /to-tickets), improved grilling, a full development lifecycle, and the new Wayfinder skill for planning large projects.

agents-cli (by Google): A command-line tool to build, test, and deploy agents on Google Cloud. It's designed to work with your coding agent like Antigravity, Claude Code, or Codex.

Claude-Video (6.9K ⭐): Claude handles most inputs, but video was the missing link. "/watch" fixes that by ingesting your links and recordings. It analyzes frames and audio to answer technical questions just like a teammate.

How coding agents work (from the co-creator of Django): You can't get the most out of coding agents without knowing how they actually work. Django co-creator Simon Willison wrote a practical guide to fix that. It explains technical details like system prompts and execution loops without any fluff.

Codex usage limits can run out fast, but running it on a local model bypasses the quota entirely. Thibault Sottiaux, OpenAI's head of Codex, mentioned this in a post back in June.

You can now run Codex locally on your machine, meaning no API bills, no rate limits, and total privacy for your code. Full setup details in OpenAI's docs.

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

This open source tool lets you run AI locally with a single install, automatically choosing the best model for your hardware.

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