Introducing Pangram 4 (2 minute read)

# Introducing Pangram 4 | Pangram Labs Source: [https://www.pangram.com/blog/introducing-pangram-4](https://www.pangram.com/blog/introducing-pangram-4) Today we're announcing Pangram 4: our most powerful and accurate AI detector to date\. Pangram 4 is over six times larger than its predecessor, and it demonstrates significant improvements in humanizer detection, identification of AI\-assisted text, and boundary detection on interspersed human and AI content\. It is also a robust detector of the latest frontier models\. We believe it to be the best of its kind in the world\. For the full evaluation numbers, training details, and known limitations, see the[Pangram 4 technical blog post](https://www.pangram.com/blog/pangram-4-technical)and the[Pangram 4 model card](https://www.pangram.com/research/model-card/pangram-4)\. You can also[view the full technical report here](https://pangram-public.s3.us-east-1.amazonaws.com/pdf/pangram_4_technical_report.pdf)\. ## Results On internal benchmarks, Pangram 4 achieves a false positive rate of just 0\.0041%, or roughly one false positive for every 24,000 documents\. Pangram 4 also has a marked reduction in false negatives, achieving a false negative rate of 0\.3396%, down from Pangram 3's false negative rate of 1\.99% on our new challenge datasets\. To our knowledge, Pangram 4 is the first AI detection model that is able to differentiate AI\-edited writing from interspersed human\-AI content simultaneously in a single pass\. It is robust against humanizers and adversarial prompting, identifying AI involvement in humanized output 98\.83% of the time across 13 popular commercial tools\. It generalizes across frontier models: every model family is detected with an FNR below 0\.7%, with no significant correlation between a model's release date and its detectability\. ## Pricing In addition to the launch of our new models, we're adjusting how credits work and how many are in your plan\. Instead of each scan rounding up to 1,000 words, one scan will be used per 100 words, letting you make the most of your credits\. We're also adjusting scan volume to better reflect our model upgrades and current user behavior\. Finally, we're excited to add image scans to every plan with no price increase\. Here's a before and after breakdown of the plan changes: **Monthly plans before** - **Individual:**600 credits of 1,000 words - **Pro:**3,000 scans of 1,000 words **Monthly plans after** - **Individual:**Scan up to 300,000 words \+ 100 image scans - **Pro:**Scan up to 1,500,000 words per month \+ 500 image scans The same credit change applies to the API\. API calls are priced at $0\.05 per 100 words, which represents a price change of 1\-2x for short documents and up to 10x for long\-form documents\. We will continue to support Pangram 3 for the next two months, with a scheduled deprecation date of September 30, 2026\. Default requests that don't specify a model will continue routing to Pangram 3 using the previous billing system, until Pangram 3 is deprecated\. For the full evaluation numbers, training details, and known limitations, see the[Pangram 4 technical blog post](https://www.pangram.com/blog/pangram-4-technical)and the[Pangram 4 model card](https://www.pangram.com/research/model-card/pangram-4)\. You can also[view the full technical report here](https://pangram-public.s3.us-east-1.amazonaws.com/pdf/pangram_4_technical_report.pdf)\.