GPT-5.6 Lands With a 13-Point Lead Over Claude Fable 5 and a 16x Cost Advantage

Hello everyone, I'm Kangaroo Emperor.

I've been using Fable 5 like crazy lately. It's good, but it's way too expensive!

But I'm not using it on the official Claude platform, because my latest account got banned again on July 1st, damn it (I've lost count of how many times this has happened).

The world has suffered under A Corp for too long...

So, I've been using Fable 5 on Cursor recently. Although I splurged on a $200 subscription, after just two days of running Fable 5, my total usage already exceeded 48%...

It's making me a bit afraid to use it. Unless it's a core task, I only dare to turn on auto mode (automatic model selection) at other times.

That's how frustrating it is.

Then, in the early hours of this morning, OpenAI finally made a move!

GPT-5.6 has been officially released.

I haven't stayed up late to follow a model release in a long time, but this time GPT-5.6 got me genuinely excited.

After reading the data on the official blog, I laughed out loud 🤣

Fable 5, it's time to say goodbye 👋

P.S.: OpenAI is so generous. I had run out of my Codex Pro quota yesterday, but with the new model release, they reset my quota. Awesome!!

How powerful is GPT-5.6?

OpenAI started teasing this GPT-5.6 model with small-scale releases back in June. It's not to be underestimated; it doesn't feel like a minor update.

Three models: Sol (flagship), Terra (balanced), Luna (cheapest)

Let's talk about the benchmarks first. Straight to the data.

Agents' Last Exam tests long-chain tasks across 55 professional domains. GPT-5.6 Sol scored 53.6 points; Claude Fable 5 scored 40.5 points.

Fable 5 is behind by 13 points.

This isn't the incremental, toothpaste-squeezing lead of the past; this is a complete blowout!

And Sol, in medium reasoning mode, already surpasses Fable 5 at roughly a quarter of the cost.

Coding ability is even more exaggerated:

On the Artificial Analysis Coding Agent Index, Sol scored 80 points; Fable 5 scored 77.2.

The gap doesn't look big, right? But Sol's token usage is only half of Fable 5's, its completion time is also half, and the price is a third cheaper.

Sol does a better job, faster, and cheaper~

On Terminal-Bench, DeepSWE, and other long-chain engineering tasks testing real codebases, Sol also achieved new top scores across the board.

Then I checked other areas: scientific research, cybersecurity, browser tasks, computer operation.

In basically every category, Sol is either first or close to first, and almost always achieves this with fewer tokens and in less time. 🐂🍺

Honestly, every time OpenAI released a new model before, I'd look at the benchmark tables and there were always a few items where Claude held the lead.

But after looking around this time, there really aren't many places where Claude Fable 5 can win anymore 🤔

How does GPT-5.6 perform on real products?

Beyond benchmarks, we still need to see how it actually performs in use.

Many of you have probably used Lovable, that AI website builder tool.

Their co-founder stated: after integrating GPT-5.6, the steps users needed to build an app decreased by 25%, tool calls dropped by 35-48%, and instances of getting stuck decreased by 15%.

Qodo, which does code reviews, tested and found GPT-5.6 to be the strongest code review model, with token usage 3 times less than GPT-5.5 and response latency cut in half.

Design capability has finally improved

Previously with GPT-5.5, I always heard friends complaining and criticizing its terrible frontend aesthetics.

But this time, GPT-5.6's frontend and design capabilities surprised me.

While browsing the official demos, I saw several works:

A sailboat game, a museum website, a clock village game, an interior design showcase. All were interactive works generated from just a single prompt.

And it doesn't just generate code and stop.

It also uses computer use to look at the actual rendered output, finding visual issues or interaction bugs and fixing them itself.

In the newly released ChatGPT Work (which is basically a 'work' mode added inside Codex), it can also directly generate interactive visualizations like spiral graphs, wave interference simulations, tokenizer principle explainers, and the effects are quite nice.

Note, this is generated directly inside the dialog box; previously it was always in the right-side preview.

This interaction method seems to be a first, right? 🤔 Quite novel. And the effect is indeed very good, and more convenient. I expect all the major agents will follow suit soon.

However, the PPTs generated by GPT-5.6 are still very ugly:

For making PPTs, you can completely disregard GPT-5.6.

But although the appearance is ugly, it seems to be quite accurate at grabbing data. Maybe foreign countries focus on data/presentation and don't care if the PPT looks good or not.

What about the price?

This is the main event.

The three models of GPT-5.6: Sol (flagship), Terra (balanced), Luna (lightweight).

Pricing is as follows:

Sol: $5 input / $30 output (per million tokens)

Terra: $2.50 input / $15 output

Luna: $1 input / $6 output

How awesome are Terra and Luna? OpenAI's exact words were:

Terra and Luna beat Fable 5 on some evaluations, but at roughly one-sixteenth the cost.

One-sixteenth!!

Anyone who has used Fable 5 should be able to appreciate the weight of those words.

I burned through 48% of my quota in just two days on Cursor before, and always felt constrained in my subsequent usage. If I switch to Terra or Luna, I can use it freely.

Seriously, with performance stronger than Fable 5, a cheaper price, faster speed, and easier access...

Probably no one will use Fable 5 anymore, right? 🤔 Feel free to discuss in the comments and share your thoughts.

How to access it?

ChatGPT Plus and Pro users can use GPT-5.6 Sol starting today, just set the effort level to medium or higher.

If you use the API, all three models are open today.

In ChatGPT Work and Codex, Free and Go users default to Terra. Plus and above users can manually switch between Sol, Terra, and Luna, and can also adjust the effort level.

There's also an interesting 'ultra' mode; it runs 4 agents in parallel simultaneously, suitable for particularly complex tasks.

Pro and Enterprise users can use it. But I haven't tried it yet; I plan to write a separate piece on this later, so I won't expand on it today.

「Finally」

In the past, I kept jumping back and forth between Claude and GPT. Ultimately, it was because GPT wasn't competitive enough; in many scenarios, Claude was indeed better to use.

But with GPT-5.6 this time, performance has comprehensively surpassed it, and the price is so much cheaper.

It feels like OpenAI has pinned Anthropic to the ground and is rubbing its face in it.

Actually, I'd rather OpenAI's models far surpass Anthropic's. Let A Corp stay closed off.

Friends who have actually tested GPT-5.6 are welcome to share your experiences in the comments~

Sleepy 🥱, going to sleep 💤, see you tomorrow.

I am Kangaroo Emperor, a blogger dedicated to helping you turn AI into productivity. See you next time~*

Those who can read this far are truly rare beings!

If you think it's good, feel free to like, click 'wow', and forward~

If you want to receive push notifications first, you can also give me a star ⭐

Thank you for patiently reading my article~