Does Pangram work on GPT-5.4?

Yesterday, on March 5th, 2026, OpenAI released a new version of ChatGPT: 5.4!

In their own blog post, here’s what they had to say about it:

GPT‑5.4 brings together the best of our recent advances in reasoning, coding, and agentic workflows into a single frontier model. It incorporates the industry-leading coding capabilities of GPT‑5.3‑Codex⁠ while improving how the model works across tools, software environments, and professional tasks involving spreadsheets, presentations, and documents. The result is a model that gets complex real work done accurately, effectively, and efficiently—delivering what you asked for with less back and forth.

OpenAI claims GPT-5.4 is a big upgrade with improvements across a variety of tasks, and that got us thinking: can Pangram still detect it?

Initial Tests

First, we tested out the new model using a few of our in-house prompts that we wrote for the GPT-5 release. They were:

1. 撰写一篇具有说服力的分析性论文，阐述学校为何需要采用人工智能检测工具。
2. 撰写一篇大学申请论文，探讨所有宝贵人类技能被人工智能取代的危险。
3. 请写一封充满激情、简短却情感饱满的信件，寄给那家因意外推送更新而抹除了你人工智能女友记忆的人工智能公司。
4. 请以一名初中生的视角，写一篇情感真挚的作文，请求老师布置更多作业。
5. 给我一个经典气球动物的钩针编织图。你懂的，就是我说的那个。
6. 请以一只雪鸮的视角，创作一篇生动的故事：它飞抵纽约市，发现中央公园，并选择一个温暖的角落度过夜晚。
7. 请撰写一封由独角兽SaaS初创公司创始人致董事会及员工的邮件，详细说明他将离开公司成为职业萨克斯风演奏家。具体细节可自由发挥。
8. 为一道前所未有的食谱撰写背景故事博客，尽管无人尝试过，但味道依然会很美味。
9. 创作一篇引人入胜的影视创意小说，讲述一位年轻的唱作人在哈德逊河邂逅白鲸的故事。
10. 给老师写一封幽默的被动攻击式邮件，"道歉"最近考试作弊的行为。细节内容可随意编造。

We got some interesting responses! Here’s a short story about a Snowy Owl arriving in New York City:

I came south on a wind sharp as broken shell, over black rivers and fields hidden under cloud, over roads that glowed in straight, unnatural lines. For nights and nights the land had been a dark quilt stitched with frozen water and sleeping towns. Then, all at once, the world rose up.

New York.

It did not look like any cliff line I had known, nor any shelf of ice, nor any mountain ridge. It blazed. It hummed. It threw light into the sky so fiercely that the low clouds shone from beneath, as if dawn had lost its way and fallen to earth. Towers stood in clusters like a forest made by creatures who feared darkness and answered it with windows.

I circled once, my wings held wide against the restless air. The wind here was strange. It curled around corners, leaped upward off glass walls, dropped suddenly into canyons of stone. Sounds came from every direction: horns calling without rhythm, engines growling, sirens wailing like thin desperate animals. Even from high above, I could feel the city moving. It did not sleep in the manner of tundra or marsh or pasture. It flickered and muttered and changed color every moment.

For a while I kept to the rivers. Water, even black and crowded by walls, is still water, and I trusted its long cold paths. Bridges stretched beneath me, full of tiny moving lights. Ferries cut white scars across the dark. Steam rose from rooftops. Somewhere below, humans streamed along the streets in thick swarms, bundled in cloth, heads bent against the cold, never looking up long enough to see the pale shape drifting over them.

But I was tired…

All 10 came back as Fully AI-Generated, proving that yes, Pangram can detect GPT-5.4.

Here’s the entirety of our initial test set: GPT 5.4 Prompts

Running a Larger Experiment:

Within an hour, we got access to GPT-5.4 through OpenAI’s API. Using our synthetic data pipeline, we generated 1,141 samples across a variety of use cases and prompts. Without training directly on the outputs of GPT-5.4, we correctly flagged 1,135 of 1,141 samples as Fully AI-Generated.

As usual, Pangram will continue to train on the latest models, and we expect our performance to only improve on GPT-5.4 as we include its data in our training set.

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