Every model it had, replaced in one week

On October 20, 2026, Originality.ai shipped two new detectors, 5.0 and 5.0 Max, and in the same move finished retiring everything that came before them. Lite, Turbo and Academic were sunset in September 2026. Multi Language and AI Allowance followed in October. AI Allowance, the model that let you say "up to 15% AI is fine", had only launched in July. It lasted about three months.

The bigger change is what the result looks like. The old percentage confidence score is gone. A scan now returns one of three labels: Likely Original (under 10% AI), Mixed Writing (10-80% AI) or Likely AI (81-100% AI), with sentence-level highlighting underneath. Founder Jon Gillham framed it in the launch press release: "As AI becomes more integrated into the writing process, the way people want to understand AI involvement is changing." That is a fair point. A document that is half human and half ChatGPT never fitted a single "73% AI" number.

The two models split the job. 5.0 is the cheap, fast one for routine volume. 5.0 Max is slower and, in Originality's words, meant "for high-stakes decisions". Both auto-detect language across 30+ languages, both now flag AI-written code, and the launch post says they catch output from the GPT-6 family, GPT-6.1 Sol, Claude Opus 5.5 and Grok 4.7. Originality.ai has been at this since November 2022, when Gillham launched it to catch GPT-3 text, before ChatGPT existed. The company is based in Collingwood, Ontario.

Pricing, and the 10x catch on the accurate model

Prices below were checked on the pricing page on October 25, 2026. Everything is sold in credits, and one credit covers 100 words.

The headline says Pro scans 200,000 words a month. That is true only on plain 5.0. The credit cost per 100 words now depends on the model you pick:

• 5.0: 1 credit, or $0.10 per 1,000 words
• 5.0 Deep Scan: 5 credits, $0.50 per 1,000 words
• 5.0 Max: 10 credits, $1.00 per 1,000 words
• 5.0 Max Deep Scan: 50 credits, $5.00 per 1,000 words

So the model Originality itself recommends for high-stakes calls turns Pro's 200,000 words into 20,000 - about twenty 1,000-word articles a month. Run Max with Deep Scan and it is 4,000 words. Add a plagiarism check to the same scan and the burn doubles, since AI plus plagiarism costs 2 credits per 100 words on the base model. Subscription credits also expire at the end of each month with no rollover. A quiet month is money gone.

Two smaller things the plan cards blur. The free tier is described two ways on the same pricing page (100 credits once on the card, a daily-renewing allowance in the comparison table); the homepage backs the 100-credit figure. And scan history on Pro is 30 days according to the help centre, even though the pricing card promises a year. Download the PDF report if you need a record.

What the accuracy numbers actually prove

Originality's launch numbers are strong. On its own OpAI-Bench set of 1,000 mixed documents, 5.0 Max labelled 90.1% correctly as mixed, against 82.9% for 5.0 and 22.4% for Pangram 4. On HumanizerBench, text run through humanizer tools, Max scored 92.38% where GPTZero 4o managed 35.95%. On human-written student essays (PERSUADE 2.0) it left more than 99.95% alone.

Every one of those figures is the vendor's own test, and the launch post admits "these are not a matched head-to-head test". The model is five days old. No independent study of 5.0 exists yet.

The independent record is for older models, and it is mixed. The RAID benchmark (ACL 2024, more than 6 million texts, 12 detectors) found Originality the strongest commercial detector overall at 85.0% accuracy with a 5% false-positive rate, and it held up well against paraphrasing (96.7%). But one cheap trick, swapping letters for look-alike Unicode characters, dropped it to 9.3%: the largest collapse of any detector tested. Originality calls homoglyphs a "rarely used" technique. That was model 2.0 Standard, not 5.0.

A November 2025 study of slightly AI-polished Arabic articles rated Originality the best commercial model at 96% accuracy, with an 8% false-positive rate. GPTZero, a direct rival, published its own comparison in April 2026 giving Originality 83.0% accuracy and a 4.79% false-positive rate, roughly one human text in twenty flagged. Treat that as a competitor's claim, not a neutral benchmark, but it lines up with the most common complaint on Trustpilot (4.6/5 from 1,365 reviews): that it flags clearly human writing as AI. Originality's own help centre puts false positives at about 1.56% and says Grammarly, Quillbot and Microsoft Word's editor all push the AI score up.

The fair reading: good enough to sort a pile of freelance submissions, not good enough to be the only evidence against one person.

What you get besides the score

Even the free account includes the AI checker, a plagiarism checker, readability and grammar checks, a fact and hallucination checker and sentence-level highlighting. Pro adds Bulk Scan, file uploads (docx, pdf, doc), shareable reports, PDF downloads, team management and tagging.

Deep Scan is the new feature most likely to be misread. It explains why a passage was flagged and suggests edits to make it read "more human". That is useful to a writer, and an odd thing for a detection tool to sell. It is English only.

The Chrome and Firefox extensions are the most practical part for an editor. Inside Google Docs they record a character-by-character writing replay and spot text that was auto-typed instead of written. If a freelancer disputes a "Likely AI" label, a replay settles it faster than any score.

The REST API is Enterprise only, so programmatic scanning starts at $136.58 a month on an annual plan. Originality publishes no rate limits on any page we could reach. 5.0 also runs through Bulk Scan, an MCP server and a Moodle integration.

Try Originality.ai

Your scans train it unless you switch that off

The privacy policy, amended June 17, 2026, says: "Unless you opt out, we may use your scan history and scan results to help train and improve our models." It is opt-out, not opt-in. The switch is under Account & Settings, Data Opt-Out, where one toggle controls it. The help centre calls it "Enable Data Usage"; in the app it reads "Use anonymized data to improve our model". Student data is excluded from training general-purpose models. Some features pass text to OpenAI's API, and the policy notes OpenAI may keep it for up to 30 days. No fixed retention period is given for Originality's own copy. Anyone scanning client work under an NDA should flip that switch before the first scan.

Refunds are "at our own discretion", case by case, and any refund wipes your remaining credits. One Trustpilot reviewer this month asked for a refund after the Lite model they relied on was retired and was refused. With the whole model lineup replaced in two months, that is a real risk for anyone who bought a year up front for one specific model.

Verdict: who should buy it, and who should not

Buy it if you run a content site or an SEO agency and screen freelance copy at volume. 5.0 at $0.10 per 1,000 words is cheap triage, Bulk Scan and the plagiarism check share one bill, and the Google Docs writing replay gives you something firmer than a score when a writer pushes back. Pro at $14.95 a month covers a small team; use Max only on the pieces that matter.

Skip it if you plan to accuse one student or one writer on the strength of a single result. The independent false-positive figures for its earlier models run from about 5% to 8%, the homoglyph collapse is on record, and nobody outside the company has tested 5.0 yet. Educators who want a lower independent false-positive figure should look at GPTZero. If plagiarism is all you need, the checker in QuillBot or Grammarly costs less than credits that expire monthly. And if you are on the other side, trying to get AI drafts past a detector, our AISEO review shows how humanizers fare against tools like this one.