Pangram is the AI detector we'd trust most when a false accusation has real consequences. Where rivals chase headline accuracy, Pangram optimized for the number that matters in the real world: its false-positive rate is roughly 1 in 10,000, verified by university researchers. It tied first on a major public benchmark, added a smart four-tier scoring system, and priced itself right at Originality.ai's level. It's not perfect on edited AI text, but for institutions it's the strongest pick we've tested in 2026.
Verdict: The most trustworthy detector for high-stakes use, thanks to an exceptionally low false-positive rate and nuanced four-tier scoring. Held back only by weaker performance on lightly edited AI and a stingy free tier.
Best for: universities, publishers, and moderators who can't afford to wrongly flag a human writer.
What is Pangram?
Pangram Labs is a research-driven AI detector that launched to catch the flaw plaguing the category: false positives. Most detectors will occasionally call a human's writing "AI," and in a classroom that's a disaster. Pangram trained specifically to drive that rate as low as possible, and independent testing from the University of Chicago and University of Maryland put it near 1-in-10,000.
The current release, Pangram 3.0 (December 2025), replaced the usual single percentage with a four-tier verdict and supports 20+ languages. It identifies content from GPT-4o, Claude 3.5 Sonnet, Gemini, Llama 3, DeepSeek, and other major models.
Key features
Four-tier classification
Instead of a raw 0-100% score, Pangram 3.0 returns one of four labels: Fully Human, Lightly AI-Assisted, Moderately AI-Assisted, or Fully AI-Generated. That nuance matters — it separates "used Grammarly" from "pasted straight out of ChatGPT," which a single number never could.
Ultra-low false positives
This is the headline. A verified false-positive rate near 1-in-10,000 means the "Fully Human" verdict is genuinely trustworthy, which is the opposite of how most detectors behave. For any use where a wrong accusation is costly, this is the differentiator.
Multilingual detection
Pangram detects across 20+ languages, so it's usable beyond English-only classrooms and content teams. Coverage isn't as broad as Copyleaks, but it's solid for the major world languages.
API & institution plans
There's an API and dedicated institution tiers for schools and platforms that need volume and integration, so detection can run inside grading tools or moderation pipelines rather than one paste at a time.
Accuracy tested
Pangram tied first at 99.3% on the COLING 2025 benchmark and posts a false-positive rate around 1-in-10,000 in independent university testing — the best combination in the category. In practice that means very few missed AI passages and very few wrongly flagged humans, which is the balance every other tool struggles with.
The honest caveat is the same one that humbles every detector: AI-edited text. When a human writes a draft and an LLM polishes it, published accuracy drops to around 73%. So Pangram is superb at catching fully machine-written content and confidently clearing human work, but the muddy middle — heavily co-written text — remains hard. Use it as strong evidence, never as a verdict on its own.
Pricing
Pangram uses a credit model where one credit checks up to 1,000 words. Plans start around $20/mo, matching Originality.ai. The free tier is limited to five checks a day.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | 5 checks/day, basic verdicts |
| Starter | ~$20/mo | 600 credits (~600,000 words), four-tier scoring, history |
| Institution / API | Custom | High volume, seats, LMS/API integration |
At $20/mo Pangram costs about the same as Originality.ai but produces fewer false positives, which is the whole reason to choose it. If budget is the constraint, note that Copyleaks starts at $7.99 and GPTZero has a free monthly allowance — see the full GPTZero pricing breakdown for the low end of the market.
Pros & cons
Pros
- Lowest verified false-positive rate in the category (~1-in-10,000)
- Four-tier scoring adds real nuance over a single percentage
- Tied first on the COLING 2025 benchmark
- 20+ languages plus API and institution plans
- Priced in line with Originality.ai
Cons
- Accuracy drops to ~73% on AI-edited text
- Free tier is only five checks a day
- Fewer bundled extras than Originality.ai (no SEO/plagiarism suite)
- Fewer languages than Copyleaks
Who it's for
Use Pangram if a false accusation would be costly — universities, publishers, and content moderators who need to trust a "human" verdict. Its false-positive record is unmatched, and the four-tier labels make results easy to act on fairly. Look elsewhere if you want a broader suite with plagiarism and SEO tooling (Originality.ai) or the cheapest combined checker (Copyleaks).
See where it ranks in our best AI detectors roundup, or compare the incumbents in GPTZero vs Originality.ai.
Frequently Asked Questions
How accurate is Pangram?
Pangram is among the most accurate detectors tested, tying first at 99.3% on the COLING 2025 benchmark with a measured false-positive rate near 1-in-10,000. Its weak spot is AI-edited text, where published accuracy drops to about 73%.
How much does Pangram cost?
Plans start around $20/mo for 600 credits, where each credit checks up to 1,000 words. The free tier is limited to five checks per day, and there are separate institution and API plans for higher volume.
Is Pangram better than Originality.ai?
They're priced similarly at around $20/mo, but Pangram produces fewer false positives, which matters most in schools where a wrong flag has real consequences. Originality.ai bundles more extras like plagiarism and SEO tools.
What is Pangram's four-tier scoring?
Rather than a single percentage, Pangram 3.0 returns Fully Human, Lightly AI-Assisted, Moderately AI-Assisted, or Fully AI-Generated. It separates light tool use from wholesale AI generation, which a raw score can't.
Why does false-positive rate matter so much?
A false positive means a human's genuine writing gets flagged as AI. In education or publishing that can end in a wrongful accusation, so a detector's trustworthiness hinges on keeping that rate near zero — which is Pangram's whole design goal.
Can Pangram detect edited AI text?
Partly. It's excellent on fully machine-written content but drops to around 73% accuracy when a human has meaningfully edited an AI draft. That muddy middle remains hard for every detector, so treat results as evidence, not proof.
Does Pangram have a free plan?
Yes, but it's limited to five checks per day. For regular use you'll want the paid Starter plan at about $20/mo, or an institution plan if you need volume and integrations.