Welcome back. Over the weekend, developers across Silicon Valley were busy testing OpenAI's GPT-5.6. The model immediately earned praise for its long agentic runs and gained traction among developers. Demand was so high that OpenAI actually had to update its usage limits.

Also: CTO’s guide to using GPT-5.6-Sol without hitting limits, how agentic coding works in production, and Anthropic extending Fable 5 access (again).

OpenAI's new flagship model wins its first weekend: The ChatGPT maker took GPT-5.6 Sol public on Thursday. Developers jumped in quickly and praised its long agentic runs and subagent orchestration. Over the launch weekend, usage hit double the previous record. Codex and ChatGPT Work now have 6M active users combined, according to Codex lead Thibault Sottiaux. In response to the high demand, the company temporarily removed its five-hour usage cap. See the reaction.

Cursor lets you question your agent without stopping it: The AI coding startup just rolled out side chats in its 3.11 release. You can use /side or /btw to start a parallel conversation with your main chat's context. This lets you research libraries or double-check decisions while your agent keeps working. You can even @-mention the side chat later to bring those answers back into your thread. The update also adds transcript search and cloud agent hooks. 

Apple accuses OpenAI of stealing its hardware secrets: The iPhone maker filed a lawsuit on Friday, specifically naming Chief Hardware Officer Tang Tan, who allegedly asked job candidates to bring Apple prototypes to interviews. The company also claims that one engineer kept his Apple-issued laptop and exploited a security bug to download confidential technical files after leaving. The lawsuit hits just as OpenAI is rumored to be building a device to rival the iPhone. OpenAI says that it has no interest in its competitors' trade secrets.

The fastest-growing startups get compliant early. That's where Vanta comes in.

Used by over 16,000 companies like Ramp, Cursor, and Harvey, Vanta helps you get audit-ready quickly — and stay that way.

Don't let compliance be the reason a deal slips. Sign up with Vanta today, and get $1,000 off.

Get started

Choosing a model broke. In just the last month, five new frontier models have dropped. Each one offers multiple sizes and reasoning capabilities. That's too many options to evaluate on every task. Over the weekend, former CTO Yuchen Jin looked for a smart router to automate the selection process. He quickly discovered that the industry was already building one.

Here’s what one does. A router sends each request to the cheapest model that can do the job. More difficult cases automatically move to a stronger model. Cognition's new Devin Fusion does exactly this by switching models mid-task. According to their benchmarks, it maintained top-tier coding performance while cutting costs by 35 percent. The real challenge, though, is figuring out which requests are just “too much” for the cheaper models to handle.

That's an eval problem. To know if an output is actually good, you first have to define what 'good' looks like. That’s where your eval data comes in. It uses real prompts from your workload along with scored answers. The startup, Not Diamond, trains custom routers based on the eval sets customers upload, and that technology is what powers “Auto mode” on OpenRouter. Without your specific evals, every router is basically just guessing. 

The moat isn't the router. OpenRouter sells one. Microsoft includes one in Azure. GitHub is full of routers that can be quickly deployed. But none of them can provide you the data that shows a router what a good result looks like. Models become basic commodities every week, but your eval data grows in value over time. Veteran ML engineer Hamel Husain's guide to evals is the place to start.

Senior engineer’s verdict on GPT-5.6 vs Fable 5: This tutorial compares GPT 5.6 and Fable 5 across frontend and backend engineering tasks. You’ll see that GPT 5.6 is faster and cheaper. But it takes risky shortcuts and needs close supervision. Fable 5 remains the safer choice. It’s the higher quality option for reliable code. 

FetchSandbox: This tool lets you build, test, and validate API integrations in realistic environments before deployment.

Claude-code-templates (29.3K ⭐): Massive library of ready-to-use setups to automate your coding and integrate external tools directly into your CLI workflow. 

How to configure Claude’s 18 hidden power settings: Using Claude with default settings can lead to hidden costs and lower quality responses over time. This guide shows you how to adjust eighteen hidden settings to restore peak performance and significantly lower your token costs. 

Trying to run GPT-5.6 Sol usually means leaving Claude Code for the Codex app. The head of Codex at OpenAI, Thibault Sottiaux, shared a setup popularized by the creator of T3 Chat:

Running “claudex” launches Claude Code with GPT 5.6 Sol as the main model.

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

A senior engineer tested early access to GPT 5.6 by pushing it to its limits. He used it for everything from 20-hour tasks to full app rebuilds. His final verdict is pretty blunt.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 310K+ 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