Most coverage of GPT-5.6 Sol reads like a scoreboard: beat this test, lost that one, costs less than Fable 5. All true, and all missing the bigger shift.

The real change is where Sol lives now: one app that can research, write, code, and keep working on something for hours without you re-explaining the task at every step. Here’s what’s actually new, where the benchmarks disagree, and the one thing worth checking before you give it real access to your files.

One Topic: GPT-5.6 Sol Is Less About the Model, More About Where It Lives

What Actually Makes Sol Different

GPT-5.6 Sol isn’t just a bigger model with a new number stapled on. Most of what changed happened in training, not size, and it shows up in three features.

Programmatic Tool Calling lets Sol write small, temporary programs that coordinate its own tools, filter out the noise, and hand back only the finished result. Max reasoning gives it more time to double check hard problems. Ultra mode splits a task across several subagents working in parallel, then merges what they find. The practical effect: less time babysitting the model through each step.

Is GPT-5.6 Sol Actually The Best Model

OpenAI calls GPT-5.6 Sol its best coding model yet, and one number backs that up. On the Artificial Analysis Coding Agent Index, Sol scored 80, 2.8 points ahead of Claude Fable 5, using under half the output tokens and finishing faster.

That’s not the whole picture. Developer Simon Willison, who tests nearly every major release, called Sol highly capable but not clearly ahead of Fable 5 for his own complex coding work. Fable also kept its lead on SWE-Bench Pro, though OpenAI has questioned the quality of some tasks in that benchmark.

My take: GPT-5.6 Sol’s edge isn’t raw intelligence, it’s reasoning, tool coordination, and cost per task combined. For volume and speed, GPT-5.6 Sol wins clearly. For the single hardest problem on your desk, it’s a genuine toss-up with Fable. Another factor is usage limits, which make a huge difference because GPT-5.6 Sol is available on all paid plans with generous limits.

Three Models, Three Clear Jobs

Use GPT-5.6 Sol for the hardest work: complex coding, research, security review. Use GPT-5.6 Terra for everyday professional tasks. Use GPT-5.6 Luna for fast, high volume work like sorting files or first-pass summaries.

Pricing follows the same logic: GPT-5.6 Sol costs $5 per million input tokens and $30 per million output, GPT-5.6 Terra runs at roughly half that, GPT-5.6 Luna is cheapest of the three. Sol powers the reasoning options in regular ChatGPT chats. Terra and Luna mostly live inside Work, Codex, and the API, depending on your plan.

One App, Not Just One Fewer Icon

OpenAI merged Chat, Work, and Codex into a single desktop app for Mac and Windows. Chat (earlier ChatGPT) is for questions and thinking out loud. Work uses connected apps, local files, a built-in browser, and scheduled tasks, and it generates documents, spreadsheets, presentations, and simple hosted web apps called Sites. Codex stays the dedicated space for repositories, tests, and code review.

The real value isn’t the merger, it’s what the merger removes: the handoff. Picture prepping for a board update the old way: research in one tab, a brief in another app, a deck built by hand, then a separate request to a developer for a live dashboard, each step starting from zero. Inside the unified app, that same research becomes the brief, the brief becomes the deck, and a switch to Codex mode builds the dashboard from the same data, no re-explaining required.

How To Try This This Week

1. Route by task, not habit: GPT-5.6 Sol for hard problems, Terra for daily work, Luna for background tasks.
2. Start with Medium reasoning for routine coding. Save Max or Ultra for planning, not unsupervised execution.
3. Turn one real research task into a brief, then a deck, inside the unified app.
4. Set approval prompts before any agent can delete, move, or overwrite a file.

None of this makes GPT-5.6 Sol unsafe to use. It makes it a tool that earns bigger permissions the way a new hire does: slowly, and inside real limits. I tried for last 10 days with full access, and it’s mind blowing.

Interested in travel or photography, read last week’s LensLetter newsletter about why people quite landscape photography?

Read last week’s JustDraft about loop engineering?

Two Quotes to Inspire

The leaders who win this decade will not be the ones with the smartest model. They will be the ones who know exactly when to slow it down.

Autonomy without accountability is not delegation. It is just risk wearing a nicer name.

One Prompt to Steal

A ready one for anyone giving an AI agent real file or system access this week.

Before you touch any file, folder, or database, tell me exactly what you'reabout to do and why. Wait for me to say "yes, go ahead" before runninganything that deletes, moves, overwrites, or renames something. If youcan't find the exact file or resource I named, stop and ask me. Neversubstitute something similar on your own.

Paste this into your system prompt or custom instructions for Sol, Codex, Claude Code, or any coding agent before you turn on autonomous mode. It won’t slow down the 95% of tasks that go fine. It just puts a human back in the loop for the 5% that don’t.