Pangram 4 AI detector claims 99.66% accuracy, one error per 24,000 docs

Pangram, a company focused on detecting AI-generated text, has released Pangram 4, its latest detection model. The model is six times larger than its predecessor and delivers significant accuracy improvements: it reduces false positives by a factor of 14 and catches missed AI texts six times more often.

According to Pangram's own benchmarks, Pangram 4 correctly identifies 99.66 percent of all AI-generated text. On the flip side, it falsely flags human writing just 0.0041 percent of the time—which the company translates to roughly one error per 24,000 documents. Beyond simple detection, Pangram 4 can distinguish lightly AI-polished text from fully AI-generated text, a capability relevant to the growing prevalence of hybrid human-AI workflows. The model also resists "humanizer" tools designed to disguise AI writing as human-authored; it catches the AI component across 13 common humanizers 98.83 percent of the time.

From a commercial standpoint, Pangram has adjusted its pricing model. The API now costs $0.05 per 100 words, representing a two- to tenfold price increase depending on document length. To offset this increase, all plans now include image scanning at no extra cost. The company is maintaining backward compatibility: Pangram 3 remains available until September 30, 2026, giving existing users time to evaluate and migrate to the new version.

Pangram's market traction underscores growing demand for AI detection. According to the New York Times, annual revenue has grown 35x year over year. Monthly active users jumped significantly, from 2,700 in June 2025 to 120,000 in June 2026. These figures suggest both enterprise adoption and mainstream awareness of AI text detection as a business and content-integrity tool.

Pangram's release of Pangram 4 arrives as enterprises and content platforms struggle with AI-generated text flooding digital channels. The model's reported accuracy metrics—99.66% true-positive rate and 0.0041% false-positive rate—represent a meaningful advance if independently validated. The company's ability to distinguish lightly AI-polished text from fully AI-generated content addresses a real user need: much AI-assisted writing falls between pure human and pure machine, and nuance matters for moderation and compliance decisions.

The pricing restructuring signals confidence in the product. A two- to tenfold increase is substantial, yet Pangram's user growth—from 2,700 to 120,000 monthly users in one year, and 35x annual revenue growth—suggests customers perceive sufficient value to absorb the cost. The inclusion of image scanning at no extra cost appears designed to lock in that value and discourage switching, particularly as multimodal AI generation becomes commonplace.

One caveat: Pangram's own benchmarks are the source of these accuracy figures; independent third-party validation would strengthen claims. The stated deprecation of Pangram 3 through September 30, 2026 gives users a 15-month window to migrate, a relatively generous timeline that may reflect Pangram's effort to retain existing customers during the pricing transition.