GPT-5 Pro detector

The route posts a sample to our server, a hosted model reads it, and the response describes the writing properties present: distribution of sentence length, density of hedging, degree of structural symmetry, completeness of topic coverage, and the report-like scaffolding that reasoning-tier output tends to carry. Every observation is returned with its justification.

The name of this page contains a model tier. The page does not claim to identify one. Those two facts sit in tension and the rest of this page is written to resolve it honestly rather than to exploit it.

The procedure is trivial. The interpretation is where the care goes.

Step 1

Supply the sample. The endpoint accepts 1 to 12,000 characters per run, enforced server-side.

Step 2

Run the analysis. The text is posted to our server and read by a hosted model under a fixed system instruction.

Step 3

Read the findings against what you already know about the text. A signal can corroborate a hypothesis about origin; on its own it establishes nothing.

GPT-5 Pro sits at the top of the GPT-5 compute range, above the standard model and above Thinking mode. The explainers ranking for this term are consistent on the mechanism: Thinking runs a linear reasoning pass, Pro runs multiple reasoning threads in parallel and converges on an answer, at substantially higher latency and cost. One of them reports roughly 22 per cent fewer major errors than Thinking on difficult tasks, and describes agentic behaviour in which the model issues several searches or tool calls at once and synthesises the results into a long report.

What none of them describes is a Pro writing style. That absence is the most useful finding available on this route, and it is worth stating rather than working around. The tier is an inference-time compute decision. It changes how an answer is arrived at, and it changes the shape of the artefact, longer, more sectioned, more exhaustive, but the sentence-level habits are inherited from the underlying model.

Each of these is a property of a distribution rather than of any single sentence, which is why a paragraph tells you almost nothing and a full page tells you a good deal.

Coverage completeness

Every branch of a topic addressed, including branches with no bearing on the question. Human writing omits things, for reasons of space, interest or ignorance.

Structural imposition

Report scaffolding applied to material that is not a report: headings, nested enumeration, a summary at both ends.

Uniform section weight

Comparable length allocated to sections of unequal importance, which is what happens when structure precedes argument.

Hedge density

Epistemic qualifiers attached to claims that carry no genuine uncertainty.

Low length variance

Sentence lengths clustering tightly, the single most durable machine-prose signal across model generations.

Punctuation preference

The em dash habit persists across GPT-5 output, and published testing on 100 rewrites found it in most of the machine-written set and none of the human control set.

Tier identification requires the tier to leave a recoverable signature in the output text. There is no published evidence that it does, and the mechanism described by the explainers argues that it should not: parallel reasoning changes what the model concludes, not the token-level style with which it writes the conclusion down.

The broader problem is worse. Detection research on cross-model generalisation reports that classifiers trained on one model's output perform poorly on another's, so even model-level attribution, a much easier task than tier-level, is unreliable. Our own constraint is enforced rather than promised: the system instruction this route runs under forbids claiming certainty about authorship or about which model produced a text, and requires observable signals instead. This page targets the search term and then tells the reader the truth about it.

Competing detectors publish numbers: one advertises a 97 per cent detection rate with under 3 per cent false positives, another over 99 per cent accuracy. Each figure comes from an evaluation the vendor designed, on a dataset the vendor assembled, against a threshold the vendor chose. None is comparable across tools and none is reproducible by a reader.

We have run no such evaluation, so no figure appears here. What appears instead is the method, the instruction the model runs under, and an explicit account of the conditions under which the check is wrong. That is less persuasive as marketing and considerably more useful when you have to defend a decision you made using it.

The failure profile is specific enough to name. Short samples fail because there is insufficient distribution to read. General subject matter fails because generic content has no distinguishing texture. Conventional essay structure fails because the structure is conventional for everyone. A detector that publishes its own false-positive guidance identifies exactly these three, and they follow from the mechanism rather than from any implementation detail.

Add two more that matter here. Prose written to a house style has had its variance edited out deliberately. Second-language English writing frequently holds a level, cautious register that produces the same profile. None of this is tunable away, because at the level these tools operate the evidence genuinely is ambiguous. The correct use is as a prompt to examine process, drafts and history, and never as a finding about a person.

The endpoint enforces 1 to 12,000 characters per run and 5 runs per rolling 24 hours per browser, with no account. For context, one established detector caps free queries at 2,000 characters and reserves 100,000 for paying customers. Our ceiling is fixed rather than tiered.

The ceiling interacts awkwardly with this particular route, and it is worth saying so. Reasoning-tier output is long by construction, so a full Pro answer will frequently exceed 12,000 characters. Split on section boundaries and check the opening and closing material first, since structural imposition and terminal restatement are both end-loaded.