How to use GPT-5.6 Sol inside Codex to run your work and your life

Picture this: it's 2pm, your inbox has 30 unread emails, and you haven't touched it since morning. That never happens. But that's exactly the situation Every CEO Dan Shipper found himself in after rebuilding his entire workflow around OpenAI's Codex and the new GPT-5.6 Sol model. Not because he ignored his inbox — because an agent had already handled it for him.

That's the promise behind the latest wave of Codex-native software, and it's a very different pitch than "AI chatbot that answers questions." This is AI that sits inside your actual tools, watches your actual inbox, and hands you decisions instead of busywork. Here's how it works, why it's a bigger deal than another model bump, and how to set up something similar for yourself.

What exactly is GPT-5.6 Sol and why does it matter?

On July 9, 2026, OpenAI shipped one of its biggest updates in a while. OpenAI introduced one of its biggest ChatGPT updates yet by launching GPT-5.6, a new ChatGPT Work agent, an upgraded desktop app with Codex built in, and a hosted sites feature for paid users. Alongside the model, OpenAI also retired the old naming convention. The company introduced a new naming system for its AI models with Sol, Terra, and Luna, and GPT-5.6 arrives with three different capability tiers designed for different types of users.

Each tier has a job. Sol serves as OpenAI's flagship model for advanced work, Terra focuses on balanced everyday performance, and Luna delivers faster responses at a lower cost. If you're unsure which one to reach for, the official Codex models documentation is refreshingly blunt about it: if you are unsure, start with Sol, since it's built for ambiguous, difficult, or high-value tasks that need extra analysis, judgment, or polish, such as complex code changes, deep research, or polished documents.

On the performance side, Sol punches well above its price tag. GPT-5.6 Sol comes within one point of Anthropic's Claude Fable 5 on the Artificial Analysis Intelligence Index while costing roughly one-third as much per task, and it leads the Coding Agent Index, outperforming Fable 5 on software engineering benchmarks including Terminal-Bench 2.1. Pricing for developers building on the API is public too: API pricing starts at $5 per one million input tokens and $30 per one million output tokens for Sol, while Terra costs $2.50 and $15, and Luna costs $1 and $6 respectively.

What is Codex, and why did OpenAI just merge it into ChatGPT?

Codex used to be its own separate desktop app for developers. Not anymore. OpenAI has merged Codex into the ChatGPT desktop app for both macOS and Windows, allowing users to switch between regular ChatGPT conversations, ChatGPT Work, and Codex from a single application. If you already had Codex installed, nothing gets lost in the shuffle — existing Codex users can keep their projects, settings, and workflows after updating, while macOS users can still choose the familiar Codex app icon.

This isn't Codex getting demoted, either. Sam Altman was direct about it, according to reporting from kie.ai's launch analysis: Codex is not deprecated, with Sam Altman stating "codex is the core of our new work product" and "codex is not going anywhere." It's more accurate to think of Codex as the engine room inside a bigger house — one that now also includes Chat and Work under the same roof, available to everyone. In the ChatGPT desktop app, Chat, Work, and Codex are available on every plan, including Free, and is available globally to download on Windows and Mac.

For developers specifically, the update brings real quality-of-life upgrades. You can edit Markdown files and source code directly inside the app with inline annotations, review GitHub pull requests in a built-in sidebar alongside reviewer comments, and work across multiple repositories within a single project.

How does Dan Shipper actually run his life through Codex?

This is where things get genuinely interesting. Dan Shipper, the CEO of Every, has spent weeks stress-testing GPT-5.6 Sol inside Codex, and according to BigGo Finance's writeup of his findings, the results changed how he works day to day.

How does the card-based email system work?

Instead of opening his inbox and triaging emails one by one, Shipper built a system where Codex does the triage and hands him decisions. Soul processes each inbound email inside ChatGPT Codex, decides what action to take, and presents Shipper with a card and a suggested response, and Shipper approves or rejects — over time, the system learns what matters to him and what doesn't.

He describes his own role almost like a manager reviewing a junior employee's drafts: "Really what I'm doing is I'm the one who's tuning 5.6 inside of Codex to tell it what to pay attention to, what the next actions are, and when it presents me with decisions to make, I just make decisions."

This system has a name — Tend — and Every published it as an open-source prompt. It's an open-source prompt from Every that makes Codex your decision layer, sweeping Gmail, Cora, Slack, and more and turning each item into a card with a summary, a drafted reply, or a next action for you to take. Nothing goes out the door without a human checking it first — from each card, you approve, edit, or redirect, and nothing sends without your sign-off. You can grab the Tend prompt directly from Every if you want to replicate the setup.

The email tool underneath all of this is Cora, which is built to be agent-friendly rather than just human-friendly. Cora gives Codex two ways to act — commands for archive, summarize, and draft, plus a browser UI it can click through like a person. The workflow doesn't let the agent freelance in the inbox itself, either — Codex creates a running document for the inbox sweep, and every draft and decision goes there first before Shipper replies inline with simple calls like archive, reply, or send.

Why do people say Sol writes like a human now?

One of the most repeated complaints about AI-written emails is that they sound like AI wrote them. Shipper's take on Sol pushes back on that. "GPT 5.6 is just to the point. It's simple. It clearly expresses the thing it needs to express. It doesn't have a ton of AI-isms," Shipper said. For anyone who spends half their day drafting the same kinds of messages, that's the difference between editing every output and actually approving it. For anyone whose job involves composing text dozens of times a day, this is the feature that makes Soul sticky — you stop editing, you start approving.

How does the in-app browser let Codex collaborate inside your tools?

The card system for email is only half the picture. Shipper also has Codex working directly inside browser-based tools rather than just producing text for him to paste elsewhere. His use of tools like Codex, Claude Code, Google Docs, and PostHog shows how agents can orchestrate existing software workflows, using external tools through the agent's in-app browser for seamless, context-rich collaboration.

That extends to meetings too. Instead of sitting through every call, he lets the agent handle the follow-through. The same pattern extends to company meetings — Shipper leaves meetings early, has Soul read the transcript, extract key decisions, and push updates to the right people. And it isn't limited to work. An app reads his meals from voice notes via Monologue and photos via Apple Photos, calculates macros, and logs them automatically, and he uses it for Facebook Marketplace purchases and apartment decorating.

Where does Sol fall short, and how does pairing models fix it?

No model is great at everything, and Sol has a known weak spot: design sense. Where Soul still falls short is design — Shipper demonstrated a before-and-after with the same prompt and same image model underneath, but Soul's output was "too complicated" and "doesn't look well considered" next to Fable's version.

Rather than picking one model and living with its blind spots, Shipper runs two models together depending on the job. Dan Shipper has been pairing OpenAI's newly released GPT-5.6 Soul and Anthropic's Fable for a month — his workflow is to go into Fable for a difficult coding project, then tell it to use Soul as a sub-agent, calling them "a match made in heaven." His framing of the two models is useful shorthand if you're deciding what to reach for on a given task: Fable delivers "world-changing genius" at high cost and high latency, while Soul delivers speed, affordability, and reliability for daily-driver work.

How do you actually get started with this yourself?

You don't need Shipper's exact setup to benefit from the same ideas. Here's a practical path:

1. Download the new ChatGPT desktop app for Mac or Windows. It now includes Chat, Work, and Codex in one place — available globally for Mac and Windows, with Chat, Work, and Codex available on every plan, including Free.
2. Read the Codex models guide before you start prompting, so you know when to reach for Sol versus Terra versus Luna. As the docs put it, for narrower tasks, define what done looks like to keep the work focused.
3. Write your own operating manual. Every's breakdown of Shipper's system starts with a simple document, not a complicated tool. Write a one-page operating manual for your inbox including VIPs, what to auto-archive, what to summarize, scheduling rules, escalation rules, and your reply style.
4. Connect an agent-friendly tool. For email specifically, Cora is built for this. For collaborative documents that agents and humans can both edit, check out Proof.
5. Grab the Tend prompt from Every and adapt it to your own tools instead of building the card-based system from scratch.
6. Start every new task at Terra, since it is a natural starting point for work you previously gave GPT-5.5, and only bump up to Sol when a task needs real judgment.

For a deeper technical walkthrough of Codex-native knowledge work, Every's own power user's guide to Codex is worth the read — it frames the whole shift well: it's a workspace where you and AI agents can work side by side across your inbox, documents, data sources, and connected tools, and you bring the context, judgment, and review.

Why is "maintenance" becoming the real product in the AI era?

Building a small app with Codex used to be the hard part. Increasingly, it isn't. Shipping something is fast now — Sol can scaffold a working SaaS tool in an afternoon. The bottleneck has shifted to what happens after launch: who tunes the agent's judgment, who reviews its drafts, who decides when a card gets approved versus rejected.

That's the quiet argument sitting underneath Shipper's whole setup. The system doesn't get better because the model got smarter in a vacuum — it gets better because someone keeps feeding it decisions. Tend compounds — every call you make enhances the loop automatically, teaching it for tomorrow, so the next sweep needs less of you. That's not a one-time build. It's ongoing tending, which is exactly what the tool is named for.

If you're building anything with Codex right now — an email router, a meeting-notes pipeline, a tiny internal app — don't just ship it and walk away. The version of you six weeks from now, with a fine-tuned decision layer that already knows your VIPs and your reply style, is going to move faster than the version of you that builds a new tool from scratch every time. The model is only half the story. The other half is what you teach it while nobody's watching.