GPT-5.6 Dropped

OpenAI didn't just ship a model this weekend.

They went straight for Claude's throat.

GPT-5.6 dropped. Three of them.

Sol — the flagship.
Terra — the balanced one.
Luna — fast and cheap.

The prices are the first gut punch.

Sol: $5 in / $30 out per million tokens.
Terra: $2.50 / $15.
Luna: $1 / $6.

Now the part that actually matters — how they handle real code.

TerminalBench 2.1 tests actual coding work.

Planning, iterating, juggling tools.

Here's the board:

Think about that for a second…

Sol is beating Claude's best model outright…

Terra ties Claude Fable 5... at half the price.

And Luna — the budget option — beats last generation's flagship.

But the benchmark isn't the real story.

"Ultra" mode is.

Ultra isn't one model thinking harder.

It's a team of AI agents working as subagents on the same problem.

An orchestrated system, not a single brain.

That's the whole direction now.

Not one model getting smarter...

a coordinated team of them solving the hard stuff together.

Here's what this means for you:

Basic tasks?

Luna at $1/$6 is almost free.

Everyday work?

Terra matches Claude Fable at half the cost.

Need raw horsepower?

Sol Ultra is king right now.

And that’s not even the best part…

It's all included in your ChatGPT subscription.

Which means it can run your OpenClaw and Hermes agents...

without burning a dollar on extra API credits.

I'll be honest — I'm fired up to put these through real work this week.

Now if someone hands me a local model this good...

I've got 6 DGX Sparks sitting here waiting on it. 😏

Did You Know?

Only 10% of a successful AI rollout is the actual tech. 

BCG's split is 10% to algorithms, 20% to technology and data, and 70% to people and processes.

Good news for you: the hard 70% is change management and workflow, not coding. You don't need to be an engineer. You need to be the one who handles the part the engineers can't.

🗞️ One More Thing 🗞️

One more thing, while we're on code.

Every one of those benchmarks measures how well the model writes code.

None of them measure the part that actually wrecks your projects…

The brief.

If you've used Claude Code or Codex you already know this...

You type "build me X."

First 5 minutes?

Magic.

Then it drifts…

Builds the wrong thing…

Forgets half of what you told it…

Breaks something else you diddn’t tell it to touch…

Now you're eight prompts deep re-explaining your own project to a robot.

The model was never the problem there.

What you handed it was.

So I built something to fix exactly that.

It's called AI Project Scoper.

You talk your idea out in plain English. It interviews you — your stack, your users, the edge cases you'd forget at 11pm.

About a minute later it hands you a full CLAUDE.md or AGENTS.md scaffold. Tasks, deployment steps, a project wiki. No placeholders, no TBDs.

Drop the zip in a fresh repo, tell Claude Code or Codex "review project instructions," and it picks up right where the scoping left off.

Works with both. Run both? They get matching instructions that actually agree with each other.

Founder pricing's open right now if you want it — locked in for life.

👉 aiprojectscoper.com

It'll save you the exact part of building that makes you want to remortgage your house so you can hire a dev.

Over to You...

That's it for today.

Go put the new models to work.

Jeff J. Hunter 
Founder, AI Persona Method | TheTip.ai

“AI Money Group” to Learn to Make Money with AI