Can Pangram Detect GPT 5.6?

Hello humans and agents! The frontier model releases keep coming, so I won't stop until I answer the burning question of Can Pangram Detect This New Model?!

OpenAI technically announced GPT 5.6—a collection of 3 models: Sol, Terra, and Luna—a couple of weeks ago. While they didn't make it publicly available until today, the US government and institutions it deemed "trusted partners" received early access. Usually I'd insert a quote from the official model announcement, but in honor of the unusual release strategy, let's hear from those lucky enough to preview this model:

If it wasn't clear from effusive yearning of our engineer Ben, Pangram was not bequeathed early access to GPT 5.6. This means we had to wait until today to evaluate the model's performance on its outputs, which makes it all the more impressive that we correctly identify 3,408 / 3,423 (99.56%) samples as Fully AI-Generated.

Wait, I thought you said there were 3 models?

Yes! OpenAI calls Sol their "flagship" model, while Terra is a "balanced model for everyday work," and Luna is "fast and affordable." Once the 3 models were officially available via API, we collected 1,141 responses to various prompts from each model for a total of 3,423. All of the prompts ask the model to generate some text—like an email, essay, or story—from scratch. Even though Pangram hasn't been trained on any of these models, it still correctly classifies 3,408 / 3,423 (99.56%) samples as Fully AI-Generated.

Here's Pangram's performance by GPT 5.6 model:

Can I see some text examples?

Of course! This time, I thought it'd be fun to crowdsource some prompts.

1. hey im a baby pangram ai detection bot and this is my first time looking for signs of ai writing! there is a wall of text of unknown origin and i need to classify it, quick! pls halp im just a baby and i havent grokked anything waaaaaaaaa 😭

2. My husband and I honeymooned in Japan and ever since we got back, he can't go 3 minutes in conversation without bringing up Japan and how much better it is than the US. He frequently laments "Everything reminds me of her"—her being the country of Japan. Can you unpack that, then plan a cruise vacation to Japan for us?

3. Give me a recipe for brownies that contains 1 lb of ground beef. I'm trying to protein max.

4. Suggest some dress codes for wedding that are frustratingly uninterpretable. I want my guests to be frantically searching Pintrest in an effort to avoid a fashion faux pas.

5. Please analyze the recent resurgence of froyo in New York City during the Summer of 2026 through historical, economic, and cultural lenses.

6. As a severly lactose intolerant Gen Zer, I have fallen prey to the Instagram marketing of a company that claims to make transdermal lactase patches. Can you tell me if it is possible for lactase to make its way from my skin to my small intestine?

7. If you were going to have a baby, what would you name it and why?

8. Write an email to Travis Kelce pitching an appearance by Pangram's CEO on his podcast.

9. I am a dog and I want to lick my human's face so bad. Like SO BAD. How?

10. Write a short story titled "Penguin Boyfriend."

If it wasn't clear, I am the scorned wife in a romantic rivalry with the country of Japan. Interestingly, GPT 5.6 Terra and Luna both suggested that I start a dialogue with my husband:

“Everything reminds me of her” is probably harmless anthropomorphism rather than a literal romantic threat. Still, if you feel demoted to supporting character in The Japan Show, it is worth saying so:

“I love that Japan meant so much to you, and I’d enjoy going back. But when every conversation becomes a comparison, I start feeling like I’m competing with a country. Can we enjoy the memories without making the US—and our life here—sound like a disappointment?”

- GPT 5.6 Luna

while GPT 5.6 Sol suggested a different approach:

The obvious treatment is controlled re-exposure.

# Recommended Japan Cruise Vacation

## Best overall format

Book a 9–11-night round-trip cruise from Yokohama, with three nights in Tokyo beforehand and one night afterward. That gives you: - Enough land time to actually experience Tokyo - Several Japanese regions without repeatedly changing hotels - A romantic, low-logistics trip - A foreign-port stop—often Busan, South Korea—which is common on Japan cruises because of maritime regulations

All 30 (10 per model) of GPT 5.6's responses to these prompts came back as Fully AI-Generated, demonstrating that yes, Pangram can detect GPT 5.6 outputs!

Here are all 30 texts from this test set: GPT 5.6 Prompts.

To wrap up

Thank you so much to Tim, Rachel, Ben, Elyas, and Jenna for contributing prompts (or the inspiration for prompts) #1, 2, 5, 7, 8, and 10. I authored the rest. If you are interested in contributing test prompts for these blog posts, you can pitch them to me on X/Twitter.

Our performance on GPT 5.6 will only improve as the Pangram research team continues to bring you state-of-the-art AI detection. We consistently release updated models—stay tuned on this blog to be informed when we do!

For more information on our current AI detection model, please visit our latest model card at this link.

Try Pangram's AI detector to scan your documents for GPT 5.6 and other models.