Perplexity

Perplexity has two distinct meanings in artificial intelligence. In information theory and natural language processing, perplexity (PPL) is an intrinsic score for a probabilistic model: on a specified sequence, it is the exponential of the model's average negative log-likelihood. A lower value means that the model assigned more probability, on average, to the outcomes that were actually observed. The number is meaningful only together with the evaluated data, tokenizer or prediction unit, preprocessing, context policy, and likelihood definition.^[1][2][3]

Perplexity AI, Inc., usually shortened to Perplexity, is also the name of a privately held San Francisco-based company founded in 2022. It operates a web search and question-answering service that retrieves sources and uses language models to synthesize responses with links or inline citations. The company calls this product an "answer engine."^[14][15] These two meanings are related only by name: the company is not the metric, and a product answer does not receive a trustworthy quality score merely because its provider is called Perplexity.

This article treats the metric first and the company second. Company products, prices, financing reports, and litigation are described as of July 28, 2026, because those facts can change.

Perplexity as an evaluation metric

Definition

Let $x_{1:N}=(x_1,\ldots,x_N)$ be a sequence of evaluated units. For an autoregressive model $q$ that assigns a conditional probability to each observed unit, define the average negative log-likelihood in nats as

$$\mathcal{L}_{\mathrm{nat}} =-\frac{1}{N}\sum_{i=1}^{N}\ln q(x_i\mid x_{<i}).$$

The sample perplexity is

$$\operatorname{PPL}_q(x_{1:N}) =\exp\left(\mathcal{L}_{\mathrm{nat}}\right) =\exp\left( -\frac{1}{N}\sum_{i=1}^{N}\ln q(x_i\mid x_{<i}) \right).$$

Using base-2 logarithms gives the equivalent form

$$\operatorname{PPL}_q(x_{1:N}) =2^{-\frac{1}{N}\sum_{i=1}^{N}\log_2 q(x_i\mid x_{<i})}.$$

The product form,

$$\operatorname{PPL}_q(x_{1:N}) =\left(\prod_{i=1}^{N}\frac{1}{q(x_i\mid x_{<i})}\right)^{1/N},$$

shows that perplexity is the reciprocal of the geometric mean probability assigned to the observed units. Implementations normally add log probabilities rather than multiply probabilities, both because the log form is the natural loss and because multiplying many small probabilities causes numerical underflow.^[3]

The notation must be read literally. It describes the score of model $q$ on this particular sequence, under this conditioning convention. It does not describe a text independently of a model, and it does not by itself estimate whether the text is true, useful, grammatical, human-written, or safe.

A worked example

Suppose a model assigns probabilities $1/2$, $1/4$, and $1/8$ to the three observed next tokens in a short evaluation. Their negative log probabilities in bits are 1, 2, and 3. The mean is 2 bits per token, so

$$\operatorname{PPL}=2^2=4.$$

The same answer follows from the product form:

$$\left(\frac{1}{1/2}\cdot\frac{1}{1/4}\cdot\frac{1}{1/8}\right)^{1/3} =(2\cdot4\cdot8)^{1/3}=4.$$

The result can be described as an equivalent-choice value: the model's average surprise equals that of a uniform four-way prediction at every step. That analogy does not mean that the model literally kept four candidates or that every prediction had probability $1/4$. Very different probability profiles can have the same geometric mean.

Relationship to entropy and cross-entropy

Claude Shannon's 1948 work defined entropy as an average measure of uncertainty or information under a probability distribution.^[1] For a true distribution $p$ and a model distribution $q$, cross-entropy in bits is

$$H(p,q)=-\mathbb{E}_{X\sim p}\log_2 q(X).$$

When the expectations exist,

$$H(p,q)=H(p)+D_{\mathrm{KL}}(p\Vert q),$$

where $D_{\mathrm{KL}}$ is the Kullback-Leibler divergence. The cross-entropy is therefore at least the entropy of $p$, with equality when $q=p$ almost everywhere under $p$. Population perplexity is $2^{H(p,q)}$ when cross-entropy is measured in bits. A finite test-set score replaces the unknown expectation with a sample average.

For a language model trained by maximum likelihood, minimizing average cross-entropy and minimizing perplexity on the same observations are equivalent: exponentiation is monotone. Reporting perplexity can make a log loss easier to interpret, but it does not add information to the underlying average negative log-likelihood.

Two distinctions matter:

• Entropy is a property of a distribution under a stated unit and conditioning scheme.
• Cross-entropy or perplexity evaluates one distribution using probabilities supplied by another distribution or model.

Calling a test-set perplexity "the entropy of English," for example, would require assumptions far beyond the computation. The result reflects the model, the sample, the segmentation, and the available context as well as whatever uncertainty exists in language.

Bounds and the effective-choice interpretation

For a properly normalized probability model, each observed probability is at most 1, so finite sample perplexity is at least 1. A score of exactly 1 means that the model assigned probability 1 to every observed outcome at every scored position. It does not establish that the model is a perfect predictor on other sequences.

There is no finite upper bound equal to the vocabulary size. A uniform distribution over a vocabulary of size $V$ has perplexity $V$, but a model can assign an observed token much less than $1/V$ probability. As an observed-token probability approaches zero, the negative log-likelihood and perplexity can grow without bound. If the model assigns probability zero to any scored outcome, the sample perplexity is infinite unless the evaluation procedure changes that probability, for example through smoothing.

The phrase "effective branching factor" comes from interpreting $2^H$ as the number of equiprobable choices with the same entropy. Jelinek, Mercer, Bahl, and Baker introduced perplexity in a 1977 discussion of language-model difficulty for speech-recognition tasks, arguing that vocabulary size and simple branching counts did not adequately reflect unequal probabilities.^[2] The interpretation is useful if it remains an equivalence in average log loss. It is not a count of search-tree children, an estimate of how many words the model consciously considered, or a claim that the predictive distribution is uniform.

Historical development

Shannon connected probability, coding length, and uncertainty in communication systems in 1948.^[1] Statistical language modeling subsequently used those ideas to assign probabilities to sequences. In 1977, Jelinek and colleagues proposed perplexity as a probability-sensitive measure of the difficulty presented by a language model in speech recognition.^[2]

Early language models commonly used n-grams. By the chain rule,

$$q(x_{1:N})=\prod_{i=1}^{N}q(x_i\mid x_{<i}),$$

and an n-gram model approximates the full history with a fixed number of preceding units. Smoothing is essential because an unsmoothed count model assigns zero probability to unseen events, producing infinite test perplexity. During the n-gram era, held-out perplexity became a convenient intrinsic measure for comparing smoothing, interpolation, pruning, and model-order choices under matched evaluation conditions.^[3]

Neural probability models retained the same evaluation objective. Bengio and colleagues' 2003 neural probabilistic language model learned word representations and a word-sequence probability function jointly, reporting improved results over the n-gram baselines studied in that work.^[4] Recurrent and Transformer language models continued to report negative log-likelihood or perplexity, even as downstream, human, factuality, safety, and efficiency evaluations became necessary for modern systems.

This continuity can be misleading if it encourages comparisons across incompatible benchmarks. A word-level n-gram score on one Penn Treebank preprocessing, a subword Transformer score on another corpus, and a pseudo-perplexity from a masked model are not points on a single universal scale.

Computing perplexity correctly

Held-out data and the evaluation target

Training perplexity measures fit to observations used for parameter estimation. Held-out perplexity estimates predictive log loss on observations not used to fit the model. Only the latter is evidence about generalization to the sampled target distribution, and even it is conditional on a valid split.

An evaluation should specify:

• the exact corpus and version;
• the unit of splitting, such as documents, speakers, users, or time periods;
• duplicate and near-duplicate handling;
• all normalization and filtering;
• whether training data could contain the evaluation examples;
• whether development results influenced model, checkpoint, or prompt selection; and
• the population or domain to which the result is intended to generalize.

If the test material appears in training data, low perplexity may reflect contamination or memorization. If the test distribution differs from deployment, a precise in-domain estimate may still be irrelevant out of domain. If many alternatives are chosen using the same held-out sample, the winning result can be optimistically selected.

Training and test perplexity also answer different questions. A falling training score with a flat or rising validation score can indicate overfitting, but it is not self-proving. The gap can also be affected by stochastic regularization, preprocessing differences, small samples, or nonindependent records. Evaluation code should put the model in its intended inference mode and apply identical scoring conventions.

Tokenization and prediction units

Raw token perplexity depends on the tokenizer. One tokenizer may encode a word as one token, another as several. The number and difficulty of predictions change, so the resulting per-token averages generally cannot be compared directly. Vocabulary size alone does not correct the problem.

Even with the same nominal tokenizer, results can differ because of:

• tokenizer version or vocabulary;
• treatment of whitespace, Unicode, casing, and punctuation;
• inclusion of beginning-of-sequence and end-of-sequence symbols;
• whether special or padding tokens are scored;
• handling of unknown tokens;
• document-boundary resets; and
• truncation or segmentation of long inputs.

Cao and Rimell showed that evaluating a subword language model only on one deterministic tokenization does not, in general, compute the probability of the underlying character string. The marginal likelihood of the string sums probability over its possible tokenizations; their work develops ways to estimate that quantity.^[7] Most published token-level evaluations instead score the tokenizer's selected segmentation. That is acceptable when stated, but it should not be confused with a tokenizer-invariant string probability.

Bits per character and bits per byte

When the same raw text is available, total negative log-likelihood can be normalized by an external unit. Let

$$L=-\sum_{i=1}^{N}\ln q(x_i\mid x_{<i})$$

be total negative log-likelihood in nats. If the evaluated text contains $N_c$ characters and $N_b$ bytes, then

$$\operatorname{BPC}=\frac{L}{N_c\ln 2} \qquad\text{and}\qquad \operatorname{BPB}=\frac{L}{N_b\ln 2}.$$

Bits per character (BPC) and bits per byte (BPB) reduce the direct effect of how a model partitions the same string into internal tokens. They do not erase every evaluation difference. Character counting must specify Unicode normalization and boundary conventions; byte counting depends on the byte representation; model context, preprocessing, and whether tokenization probability is marginalized can still differ.

For a word-tokenized evaluation, bits per word is $\log_2(\operatorname{PPL})$. Converting between units is legitimate only when the denominator counts refer to the same evaluated text and total loss.

Fixed context windows

An autoregressive model with a finite context window cannot condition every token in a long document on its entire preceding history. The evaluation policy then becomes part of the metric.

A simple but inefficient method advances one token at a time, retaining the longest allowed preceding context and scoring only the new token. A common approximation uses overlapping windows with a selected stride: each segment provides context, but loss is counted only for newly introduced tokens. A disjoint-chunk method is cheaper, yet tokens at every new chunk lose context that would have been available in continuous use. Hugging Face's documentation demonstrates that changing from disjoint blocks to a strided sliding window changes the reported GPT-2 perplexity, precisely because the available context changes.^[5]

A reproducible report should state:

• the maximum context length;
• stride or overlap;
• which positions in each window contribute to loss;
• whether state resets at documents or only at corpus boundaries;
• whether the first token in a document is scored and with what context; and
• how losses are weighted when the last segment is shorter.

Averaging segment-level perplexities is generally wrong because segments can contain different numbers of scored tokens. Sum their negative log-likelihoods, divide by the total number of scored units, and exponentiate once.

Conditional and sequence-to-sequence models

Perplexity can be defined for a conditional model. If $y_{1:M}$ is a target sequence conditioned on input $x$,

$$\operatorname{PPL}(y\mid x) =\exp\left(-\frac{1}{M}\sum_{j=1}^{M}\ln q(y_j\mid y_{<j},x)\right).$$

For machine translation or summarization, this evaluates reference-target likelihood under teacher forcing. It is not a direct measure of the quality of free-running generated outputs. A system can assign high probability to references yet decode poorly under a particular search method, or generate adequate alternatives that receive low reference likelihood. End-to-end task metrics and human evaluation remain necessary.

The denominator must also be clear. Some software reports loss per nonpadding target token; other code can accidentally include padding, prompt tokens, or source tokens. Those choices change the score.

Masked language models and pseudo-perplexity

A standard bidirectional masked language model does not factor a sentence left-to-right into normalized next-token probabilities. Ordinary autoregressive perplexity is therefore not directly defined for it. Substituting its training loss into the autoregressive formula without qualification mixes different objects.

Salazar and colleagues proposed scoring masked language models with pseudo-log-likelihood: mask each token in turn and sum its conditional log probability given the remaining observed tokens. Exponentiating the negative mean yields pseudo-perplexity.^[6] Pseudo-perplexity is useful for ranking or comparing sentences under a specified masked model, but it is not the chain-rule likelihood of the sequence. It should be named and reported separately.

Variants that mask multiple tokens at once, reuse hidden states, or approximate the score can be computationally cheaper, but they define further protocol choices. Comparisons require the same masking and normalization method.

Topic models and other probabilistic models

The definition extends beyond next-token neural models whenever a model assigns probability to observations and a normalization unit is defined. Topic models, for example, can report held-out document likelihood or a perplexity derived from it. That score evaluates predictive fit under the topic model; it does not directly measure whether humans find the discovered topics coherent or useful.

Chang and colleagues compared topic-model likelihood with human interpretation tasks and found that models with better held-out likelihood could produce topics that people judged less semantically coherent in their studied settings.^[8] Their result is not that perplexity is useless. It shows that likelihood and human interpretability are different evaluation targets.

Speech-token models, image-token models, and multimodal sequence models create similar questions. A token may encode a learned acoustic or visual unit rather than a linguistic word. Global token perplexity then inherits the representation's rate, redundancy, and segmentation choices. The unit and its relationship to perceptual quality must be validated rather than assumed.

Comparing close results

A small numerical difference is not automatically a meaningful improvement. Because perplexity is exponential in mean log loss, comparisons are clearest in the underlying log-loss unit. If model A has perplexity $P_A$ and model B has perplexity $P_B$ under the same token protocol, their difference in bits per token is

$$\Delta=\log_2 P_A-\log_2 P_B =\log_2\left(\frac{P_A}{P_B}\right).$$

This transformation makes the scale explicit. Moving from 20 to 19 and moving from 200 to 190 represent the same relative ratio, even though the absolute perplexity changes differ.

Uncertainty should be estimated at a unit that reflects data dependence. Tokens within one document are not independent observations, so treating millions of tokens as independent trials can produce unrealistically narrow intervals. Depending on the sampling claim, documents, speakers, conversations, users, or time blocks may be more defensible resampling units. Paired comparisons are useful because both models score the same examples.

Aggregate results can also hide heterogeneous effects. A model may improve on frequent boilerplate while worsening on rare names, another language, long documents, or the final portion of a context window. Useful diagnostics include per-document log loss, quantiles, domain and language slices, position within context, token frequency, and the fraction of total loss contributed by outliers.

Finally, significance is not importance. A very precise reduction in log loss may have no measurable benefit in the intended application, while a change concentrated on safety-critical or rare events can matter despite a small corpus-wide average. The conclusion should therefore state both the statistical evidence and the task consequence.

Uses and limits

Intrinsic model development

Perplexity is attractive because its underlying token log losses aggregate in log space, require no separate human labels beyond the observed sequence, and are efficient to compute relative to many downstream evaluations. Under one fixed protocol, it can help compare:

• smoothing and interpolation methods;
• model architectures or parameter settings;
• checkpoints during training;
• context-length or retrieval ablations;
• domain adaptation;
• tokenizer choices, if reported with appropriate normalization; and
• compression or quantization changes that alter predictive probabilities.

The strongest statement supported by a lower score is narrow: the model achieved a lower average negative log-likelihood on that evaluation under that protocol. Any broader claim needs further evidence.

Speech recognition

Perplexity originated in the evaluation of language-model difficulty for speech recognition.^[2] A lower language-model perplexity can be helpful because the language model assigns more probability to the reference transcripts. It does not guarantee lower word error rate (WER), because recognition combines acoustic evidence, language-model scores, search, pruning, and calibration.

Chelba and Jelinek documented poor correlation between conventional perplexity and WER for some language-model comparisons and proposed an acoustically sensitive alternative that weighted confusions using acoustic similarity.^[9] Consequently, a speech-system claim should report WER or another end-to-end measure. Perplexity remains an intrinsic diagnostic for the language-model component, not a substitute for recognition accuracy.

Domain fit and data selection

A model trained predominantly on one domain often assigns higher perplexity to material from another. Researchers use this behavior for corpus selection, domain filtering, and anomaly detection. The interpretation must be relative: high perplexity means "unlikely under this scoring model and protocol," not "bad text."

Legitimate reasons for high perplexity include a new topic, another language or dialect, names, code, equations, poetry, transcription errors, or adversarial corruption. Legitimate reasons for low perplexity include common boilerplate, duplicates, memorized passages, highly formulaic writing, or genuine in-domain prose. A threshold learned for one corpus does not automatically transfer to another.

For data curation, a useful design reports the scoring model, its training data, tokenizer, threshold-selection procedure, and the distribution of scores by language and domain. Human review and downstream ablations are needed to determine whether filtering improves the eventual model or simply removes valuable diversity.

What perplexity does not measure

Perplexity is not a general intelligence or answer-quality score. By itself, it does not measure:

Scroll sideways for more →

The Stanford text distinguishes intrinsic evaluation from extrinsic evaluation and notes that application performance is the relevant test when the model is a component of a larger system.^[3] Modern language-model evaluation therefore combines likelihood with task benchmarks, controlled human studies, robustness tests, calibration analysis, safety evaluation, and cost measurements.

AI-generated text detection

Likelihood-related signals can help distinguish some machine-generated and human-written text under a matched experimental setup. GLTR, for example, visualized the rank and probability of each observed token under a language model and found that these cues helped people identify generated passages in its 2019 experiments.^[10] DetectGPT later used the curvature of a model's log-probability under perturbations, rather than a single raw perplexity threshold, to perform zero-shot detection in the model and data settings it studied.^[11]

No universal rule says that human writing has high perplexity and machine writing has low perplexity. The score depends on the generator, decoding method, domain, language, prompt, editing, and detector's scoring model. A human can write conventional prose that is highly predictable; a generator can sample with high temperature, use an unfamiliar domain, or be edited and paraphrased.

Research has documented practical and theoretical limits. Sadasivan and colleagues showed that paraphrasing can substantially degrade several detectors and analyzed detection limits in terms of the statistical distance between human and machine text distributions.^[12] Liang and colleagues found high false-positive rates for non-native English writing in the detectors they evaluated; increasing lexical diversity reduced those false positives in their experiment.^[13] These studies do not prove that every detector always fails. They show why a perplexity-like score should not be treated as proof of authorship or used alone for punitive decisions.

The safe interpretation is evidentiary, not categorical:

• a detector score is a model output under a documented operating point;
• error rates should be measured on representative contemporary generators and human populations;
• false-positive and false-negative rates change with the threshold and base rate;
• edited and mixed-authorship documents need separate testing; and
• high-stakes conclusions require process evidence, provenance, and human review.

Reporting checklist

A perplexity result is reproducible only if enough context accompanies the number. A complete report should include:

1. Model identity: checkpoint, parameter version, numerical precision, and evaluation mode.
2. Data identity: dataset name, release, split, license if relevant, and a cryptographic or content identifier where possible.
3. Contamination controls: how duplicates, train-test overlap, and benchmark exposure were examined.
4. Prediction unit: word, subword, character, byte, acoustic token, or another representation.
5. Tokenizer: implementation, vocabulary/version, normalization, and special-token policy.
6. Likelihood: autoregressive log-likelihood, conditional target likelihood, marginal likelihood, or pseudo-log-likelihood.
7. Context policy: maximum length, stride, overlap, reset boundaries, and positions excluded from loss.
8. Denominator: exact count of scored tokens, words, characters, or bytes.
9. Aggregation: total negative log-likelihood divided by total units, not an unweighted average of batch perplexities.
10. Uncertainty: document-level variation, bootstrap intervals, or another justified measure when comparing close results.
11. Selection history: whether the evaluation split affected checkpoint or hyperparameter choice.
12. Companion metrics: task, human, safety, calibration, and efficiency measures required by the intended use.

Without these details, a stand-alone number is often impossible to audit and unsafe to compare.

Perplexity the company

Identity and founding

Perplexity AI, Inc. is an American artificial-intelligence company headquartered in San Francisco. It was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski.^[17][18] Srinivas serves as chief executive. The company launched its initial public answer product in December 2022 and has since expanded it into consumer subscriptions, enterprise plans, developer APIs, a browser, and an agentic task product.^[14][18]

The company's own description emphasizes a search-first workflow: a user asks a question, the service searches the web, synthesizes material from retrieved sources, and presents an answer with citations and links.^[14][15] "Answer engine" is a product label, not a technical guarantee. Citations make verification possible, but they do not prove that each sentence is entailed by the linked source. Users still need to inspect sources for consequential claims.

Early financing and reported private valuations

Because Perplexity is privately held, its financial position is not disclosed through public-company filings. Funding totals, revenue, valuation, and user figures should therefore be separated by source and status.

Perplexity announced a $25.6 million Series A in March 2023. Its announcement also said that the company had previously raised a $3.1 million seed round in September 2022.^[18] In January 2024, it announced a $73.6 million Series B and said that cumulative funding had reached $100 million.^[19] These are company announcements, not audited financial statements.

Later valuations were reported through unnamed-source coverage. In September 2025, The Information reported investor commitments of $200 million at a $20 billion valuation; Reuters repeated the report and explicitly said that it could not immediately verify the information and that the company had not immediately commented.^[35] The baseline article's later figures of about $21 billion or $22.6 billion were not supported by an identifiable completed financing announcement in the sources reviewed for this revision, so they are not presented as facts here.

Similarly, the Financial Times reported in April 2026 that annual recurring revenue had exceeded $450 million in March.^[34] That is a media report about a private-company operating metric, not an audited result. The exact figure is therefore best described as reported, not confirmed. A 2025 statement that the service received 780 million queries in one month came from Srinivas, according to TechCrunch; query volume is not the same as monthly active users.^[36] This revision does not repeat the former claim of 45 million monthly active users because the cited statistics aggregator did not provide a sufficiently strong, independently verifiable basis.

Search and answer pipeline

Perplexity does not publish a complete production architecture, and its product can change without a paper or versioned specification. Its help center describes a broad sequence: interpret the question, search current web sources, summarize the retrieved information, and attach numbered links to sources.^[15] A Vespa case study supplies more detail about the retrieval layer. According to that vendor account, Perplexity uses Vespa to combine lexical, semantic, and structured retrieval; filter and rank candidates; select passages; and continuously update an index before material is supplied to a language model.^[16]

A generic representation of this kind of retrieval-augmented generation pipeline is:

1. Query processing. The service interprets a natural-language request and may decompose or reformulate it.
2. Retrieval. Search infrastructure selects candidate documents or passages using lexical, vector, and structured signals.
3. Ranking and filtering. The system ranks candidates for relevance, freshness, and other internal criteria.
4. Context construction. Selected passages, metadata, and links are assembled within the language model's context.
5. Generation. A model synthesizes a response conditioned on the query and retrieved context.
6. Citation rendering. The interface associates answer spans or sentences with source links.

Only some details in that diagram are publicly documented for Perplexity. It should not be read as a claim that every query follows an identical path, that generation is constrained to copy only retrieved facts, or that citations are aligned by a particular unpublished algorithm.

Retrieval can reduce dependence on information stored in model parameters and can expose current sources, but it introduces its own failure modes. The search layer can miss a decisive document, retrieve a low-quality copy, rank a source for relevance rather than authority, or return mutually inconsistent pages. The generation layer can misread the passage, blend claims from different sources, omit qualifications, or attach a citation that supports only part of a sentence. A linked source is therefore an audit trail, not a warranty.

For verification, a reader should ask:

• Does the cited page contain the claimed fact?
• Is the page primary or merely repeating another report?
• Does its publication date fit the time-sensitive claim?
• Does the answer preserve scope, uncertainty, and whether a statement is an allegation?
• Are contrary sources or later corrections missing?
• Does a numerical answer use the same unit, period, and definition as the source?

These checks are particularly important for medicine, law, finance, breaking news, and private-company metrics.

Consumer search modes

The service's standard search interface returns conversational answers with links. Follow-up questions can use the preceding thread as context. Perplexity's May 2026 help article listed regular Search, Pro Search, Research, and Create files and apps among the available modes.^[14][15]

Pro Search performs a more involved search process than the standard mode. Perplexity's July 2026 documentation says that it can search multiple source types, maintain conversational context, use code interpretation, and allow paid users to select among supported models.^[20] Model names and quotas are product settings rather than durable properties; an encyclopedia article should not freeze a list that can change weekly.

Deep Research, launched in February 2025, iteratively searches and reads sources before producing a longer report. The launch announcement described a loop that refines the research plan as it gathers material and allows export to a document or PDF.^[21] Claims in that announcement that the feature reads particular numbers of sources or completes work in a particular time are vendor-reported operating descriptions, not guarantees for every query. Likewise, benchmark scores published by the company should be interpreted with the benchmark version, endpoint, and evaluation method, rather than as a general accuracy percentage.

Create files and apps turns research or instructions into artifacts. As of July 16, 2026, the help center listed documents in DOCX or PDF form, XLSX spreadsheets, PPTX presentations, and HTML applications, with plan-dependent access and limits.^[24] Artifact generation creates additional correctness concerns: formulas, citations, layouts, and executable code need validation in the resulting file, not only in the chat preview.

These modes share the company's core retrieval-and-generation concept but have different time, tool, and compute budgets. A mode label does not define a stable scientific system, so reproducible product evaluation should record the date, account tier, selected model, prompt, location, and any files or connectors used.

Comet browser

Perplexity Comet, launched in July 2025, is a Chromium-based web browser with Perplexity search and assistant features integrated into browsing.^[22] The current help documentation describes support for standard browser functions and Chrome extensions, contextual questions about an open page, summarization, natural-language tab commands, and optional access to email or calendar data through connectors.^[22]

Comet also supports agentic actions in which the assistant operates websites or workflows for a user. That capability is distinct from ordinary search in three important ways.

First, an agent may interact with authenticated services and personal data. The scope of browser history, stored credentials, connectors, and page content therefore matters. Users should review permissions and confirm consequential actions.

Second, web actions can be ambiguous or irreversible. A shopping request, message, booking, or form submission requires correct interpretation of the user's intent as well as correct factual retrieval. Interface changes, anti-bot controls, or hidden page state can cause failures that a text answer would not encounter.

Third, authorization by the user does not automatically settle whether a website authorizes automated access. This distinction became central in Amazon's lawsuit over Comet's access to password-protected sections of Amazon's site, discussed below.^[33]

The baseline article gave exact platform rollout dates, voice-model versions, and lists of models available to particular subscribers. Those details were removed because they are volatile and did not materially explain the browser. The stable facts are that Comet is Chromium-based, integrates Perplexity's assistant, and can combine browsing with question answering and selected actions.

Perplexity Computer

Perplexity introduced Computer in February 2026 as an agentic product for carrying out multi-step tasks rather than only returning answers.^[23] Its launch post said that the system can decompose an outcome into subtasks, create subagents, use a browser and filesystem in an isolated environment, call connected services, and run work asynchronously. The July 2026 help center described document creation, code, web research, persistent memory, schedules, and connectors to external applications.^[23]

Those are provider descriptions of capabilities, not independent measurements of reliability or safety. Multi-agent orchestration can help parallelize work, but it also creates more execution paths and more opportunities for an error to propagate. An agent that researches, analyzes, creates a document, and sends it can fail at any transition. A responsible use pattern therefore includes:

• least-privilege connectors;
• confirmation before external communication, spending, deletion, or publication;
• review of generated files and calculations;
• visibility into source material and tool results;
• boundaries on recurring or long-running tasks; and
• a way to stop, revoke, or inspect the workflow.

Perplexity's support page says that Computer runs in a cloud sandbox and may retain context through persistent memory.^[23] That is different from a claim that all data remain only on the user's device. The previous article conflated Computer with later product descriptions of a local "Personal Computer" for Mac and asserted detailed privacy behavior without a stable primary specification. This revision omits those claims.

Developer platform

Perplexity offers APIs separately from its consumer subscriptions. As of July 28, 2026, the documentation described three core interfaces:

• an Agent API for model responses that can use Perplexity search tools;
• a Search API that returns ranked web results without requiring a generated answer; and
• an Embeddings API for text representations used in search and retrieval.^[25]

The documentation also retained a legacy Sonar API. The Sonar family provides web-grounded generated answers, while the Agent API can route to supported third-party models and tools. These boundaries and the available model list have changed over time, so integrations should use the current API documentation rather than an encyclopedia's model table.

Pricing can combine model tokens, search or tool invocations, request fees, and other units.^[25] The baseline article embedded exact token prices and context windows but mixed older and newer products. This revision omits that table. A date-stamped cost analysis should record:

• exact endpoint and model identifier;
• input, output, reasoning, and citation-token rules;
• search-context or request fees;
• number of tool calls;
• caching and batch rules;
• rate-limit tier; and
• whether a failed or empty request is billed.

An API response that includes citations still requires validation. Developers should preserve source URLs and timestamps, expose uncertainty, handle unavailable sources, and avoid presenting model output as a primary source.

Subscriptions and business model

Perplexity earns revenue from consumer and enterprise subscriptions and API use. Exact terms vary by region, promotion, platform, and date. At the cutoff for this revision, the provider's pages listed the following US headline prices:

Scroll sideways for more →

The Max prices came from a help article updated July 16, 2026, and the enterprise prices from an article updated July 22, 2026.^[26][27] The $20 Pro price is shown on the provider's current Pro material.^[37] These figures are included only as a dated snapshot, not as permanent properties. API charges are separate, and promotions or app-store billing can differ.

The product mix shows several revenue routes:

• individual subscriptions for enhanced search and creation;
• per-seat organizational subscriptions;
• pay-as-you-go developer APIs;
• purchased credits for some Computer tasks; and
• commercial partnerships.

The previous article asserted that advertising had been discontinued in February 2026 and described several commerce and distribution deals. Those claims were not retained because a durable primary source establishing their status at the cutoff was not available in the reviewed evidence set. Absence here should not be read as proof that a revenue route does or does not exist.

Publisher program and source relationships

Perplexity launched a publishers' program in July 2024. In a December 2024 update, the company named participating outlets and said that members would share advertising-derived revenue, receive API and developer support, obtain a period of Enterprise Pro access, and receive content-performance analytics.^[28] This was the company's description of a voluntary commercial program, not a settlement of the legal questions raised by other publishers.

Le Monde announced a separate multi-year agreement with Perplexity in May 2025. The publisher said that its content would be used to answer queries with attribution, that the arrangement included revenue sharing, and that the content would not be used to train foundation models.^[29] Terms for one participating publisher should not be generalized to all web content.

These relationships illustrate three different mechanisms that are sometimes blurred:

• indexing makes a page discoverable to a search service;
• retrieval and summarization use material to answer a particular query; and
• model training adjusts model parameters using a dataset.

Robots directives, licenses, contracts, copyright exceptions, and site terms can address different mechanisms. A statement about not training foundation models does not by itself answer whether retrieval or summarization is authorized.

Crawling dispute

In August 2025, Cloudflare published results of tests in which, it said, Perplexity accessed new domains after the declared Perplexity crawlers had been blocked by both robots.txt and web-application-firewall rules. Cloudflare attributed the requests to undeclared user agents and rotating network sources and characterized them as an attempt to evade site restrictions.^[30] Those are Cloudflare's observations and attribution, not a court judgment.

Perplexity's current help article gives a different account of present behavior. Updated July 16, 2026, it says that PerplexityBot respects robots.txt, that a former feature allowing users to summarize a specified blocked URL was disabled, and that agreements with third-party crawlers were updated to require compliance, particularly for news sites.^[31] The article also says that a blocked page's domain, headline, and a short factual summary may still be indexed.

The two sources address different dates and come from interested parties. A precise summary is therefore:

• Cloudflare reported noncompliant and concealed crawling behavior in tests published in August 2025.
• Perplexity says that its current declared crawler and third-party arrangements respect robots.txt and that the earlier direct-URL behavior has been disabled.
• The sources reviewed here do not provide an independent, continuous audit covering all requests after those changes.

This framing avoids treating either company's assertion as a final factual determination about every crawl.

Copyright and related litigation

Several publishers and platform operators have sued Perplexity. Complaints contain allegations; they are not findings of fact. Orders on preliminary relief assess an early legal record under a particular standard and do not finally decide liability.

Dow Jones and New York Post

Dow Jones and NYP Holdings filed a complaint against Perplexity in the US District Court for the Southern District of New York on October 21, 2024. The docket classifies the case under copyright and trademark law and records an amended complaint and subsequent motion practice.^[32] The plaintiffs alleged unauthorized copying and substitution for their publications. Perplexity disputed the claims. The existence and procedural history of the case are verifiable from the docket; the merits were not finally resolved in the reviewed record.

The New York Times, Wirecutter, and The Athletic

The New York Times Company, Wirecutter, and The Athletic Media Company filed a copyright action against Perplexity in the same district on December 5, 2025.^[33] The complaint alleged copying, use of publisher content in a retrieval index, and distribution of answer text that substitutes for the plaintiffs' works. Those descriptions remain allegations unless admitted or adjudicated. The baseline article stated that Perplexity had filed a February 2026 motion to dismiss based on a weak news source; this revision relies on the docket for case status and does not characterize unresolved motions beyond what the public record establishes.

Amazon and agentic shopping

Amazon.com Services sued Perplexity in the Northern District of California in November 2025 over Comet's agentic access to password-protected parts of Amazon's site. On March 9, 2026, Judge Maxine M. Chesney granted a preliminary injunction. The order barred specified automated access until the action is resolved or the court orders otherwise, found Amazon likely to succeed on the claims analyzed at that stage, and denied Perplexity's request for a stay pending appeal while granting a seven-day administrative stay.^[33]

The order matters because it distinguished a user's permission from the site operator's authorization. It is still preliminary. It did not enter a final judgment on all claims, and its reasoning is tied to the conduct, record, contracts, and causes of action before that court. The baseline article described the case as a definitive bellwether and gave an appeal date not established by the cited order. Those embellishments were removed.

How to read the litigation record

A neutral account keeps four levels separate:

Scroll sideways for more →

No case listed above establishes that all AI search is lawful or unlawful. Questions can differ by the work copied, amount, source access, contractual terms, transformation, market effect, jurisdiction, and remedy.

Selected chronology

The following chronology is deliberately limited to events supported by company publications or court dockets. It excludes rumored financings, transient model menus, and product details that could not be independently stabilized.

Scroll sideways for more →

This chronology does not imply that unlisted events did not happen. It reflects a source-quality threshold.

Evaluation and research considerations

Independent evaluation of an answer engine should separate retrieval, citation, synthesis, and user outcomes. A single end-to-end score can hide which component failed.

Retrieval evaluation asks whether the system found the documents needed to answer the query. It can measure recall of relevant or authoritative sources, freshness, language coverage, and diversity. Ground truth is difficult because the open web changes and multiple source sets may be sufficient.

Citation evaluation asks whether a cited document supports the associated claim. Useful dimensions include entailment, completeness, source authority, correct date, and placement. Counting citations is not enough: an answer can cite many irrelevant pages or leave its most consequential claim unsupported.

Synthesis evaluation asks whether the answer accurately combines retrieved evidence, resolves conflicts, states uncertainty, and avoids unsupported additions. This requires claim-level review rather than overall stylistic preference.

Task evaluation asks whether the user completed the intended work correctly and efficiently. For a browser or agent, this includes action accuracy, permission handling, reversibility, and recovery from interface changes.

Safety and privacy evaluation examines prompt injection, malicious web content, data sent through connectors, retention, cross-user isolation, confirmation boundaries, and unauthorized actions. Retrieval gives untrusted web pages a path into the model context, so quoted page instructions must not silently override the user's intent.

Product comparisons also require a fixed date and configuration. Perplexity can expose different models, retrieval depth, or limits by plan and region. A study should preserve prompts, raw outputs, source links, timestamps, account tier, selected mode, and repeated trials. Otherwise a result may be impossible to reproduce after a product update.

Distinguishing the company from the metric

The shared word leads to several recurring mistakes:

• "Perplexity improved" may mean a model's evaluation score decreased, or that the company changed its product.
• "Perplexity search" normally refers to the company's service, not a search algorithm that minimizes the metric.
• "Perplexity score" should refer to the probabilistic metric only when a model, data, unit, and protocol are specified.
• The company name should not be abbreviated as PPL in technical writing when that could be confused with the metric.
• The redirect named Perplexity AI points to this combined article; it does not denote a second independent company page.

In formal results, write token perplexity, word perplexity, pseudo-perplexity, or the appropriate unit. In company contexts, write Perplexity or Perplexity AI, Inc.

See also

• AI search
• AI content detectors
• Aravind Srinivas
• Cross-entropy
• Entropy
• Evaluation metrics
• Language model
• Large language model
• Loss function
• Retrieval-augmented generation
• Search engine
• Test set
• Tokenization

References

1. ^Shannon, C. E. (1948). "A Mathematical Theory of Communication." *Bell System Technical Journal*, 27(3), 379-423, and 27(4), 623-656. doi.org/...j.1538-7305.1948.tb01338.x
2. ^Jelinek, F., Mercer, R. L., Bahl, L. R., and Baker, J. K. (1977). "Perplexity - A Measure of the Difficulty of Speech Recognition Tasks." *Journal of the Acoustical Society of America*, 62(S1), S63. doi.org/...1.2016299
3. ^Jurafsky, D., and Martin, J. H. (2026). "N-gram Language Models." *Speech and Language Processing*, third-edition draft, January 6, 2026. web.stanford.edu/...3.pdf
4. ^Bengio, Y., Ducharme, R., Vincent, P., and Jauvin, C. (2003). "A Neural Probabilistic Language Model." *Journal of Machine Learning Research*, 3, 1137-1155. jmlr.org/...bengio03a
5. ^Hugging Face. "Perplexity of fixed-length models." *Transformers Documentation*. Accessed July 28, 2026. huggingface.co/...perplexity
6. ^Salazar, J., Liang, D., Nguyen, T. Q., and Kirchhoff, K. (2020). "Masked Language Model Scoring." *Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics*, 2699-2712. aclanthology.org/2020.acl-main.240
7. ^Cao, S., and Rimell, L. (2021). "You Should Evaluate Your Language Model on Marginal Likelihood over Tokenisations." *Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing*, 2104-2114. aclanthology.org/2021.emnlp-main.161
8. ^Chang, J., Gerrish, S., Wang, C., Boyd-Graber, J., and Blei, D. (2009). "Reading Tea Leaves: How Humans Interpret Topic Models." *Advances in Neural Information Processing Systems 22*. proceedings.neurips.cc/...cd64ab20fd554ff-Abstract
9. ^Chelba, C., and Jelinek, F. (2006). "Acoustic Sensitive Language Model Perplexity for Automatic Speech Recognition." *IEEE Transactions on Speech and Audio Processing*, 14(3), 682-687. research.google/...or-automatic-speech-recognition
10. ^Gehrmann, S., Strobelt, H., and Rush, A. M. (2019). "GLTR: Statistical Detection and Visualization of Generated Text." *Proceedings of ACL 2019: System Demonstrations*, 111-116. aclanthology.org/P19-3019
11. ^Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D., and Finn, C. (2023). "DetectGPT: Zero-Shot Machine-Generated Text Detection Using Probability Curvature." *Proceedings of the 40th International Conference on Machine Learning*, PMLR 202, 24950-24962. proceedings.mlr.press/...mitchell23a
12. ^Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., and Feizi, S. (2023). "Can AI-Generated Text Be Reliably Detected?" arXiv:2303.11156. arxiv.org/...2303.11156
13. ^Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). "GPT Detectors Are Biased against Non-Native English Writers." *Patterns*, 4(7), 100779. doi.org/...j.patter.2023.100779
14. ^Perplexity. "What Is Perplexity?" *Perplexity Help Center*, updated May 1, 2026. perplexity.ai/...10352155-what-is-perplexity
15. ^Perplexity. "How Does Perplexity Work?" *Perplexity Help Center*, updated May 1, 2026. perplexity.ai/...10352895-how-does-perplexity-work
16. ^Vespa.ai. "Why Perplexity Chose Vespa for AI Search." Accessed July 28, 2026. vespa.ai/...perplexity
17. ^Forbes. "Perplexity: Company Overview." Accessed July 28, 2026. forbes.com/...perplexity
18. ^Perplexity. "Perplexity Raises Series A Funding Round." March 28, 2023. perplexity.ai/...nding-round-and-mobile-app-launch
19. ^Perplexity. "Perplexity Raises Series B Funding Round." January 4, 2024. perplexity.ai/...ity-raises-series-b-funding-round
20. ^Perplexity. "What Is Pro Search?" *Perplexity Help Center*, updated July 21, 2026. perplexity.ai/...10352903-what-is-pro-search
21. ^Perplexity. "Introducing Perplexity Deep Research." February 14, 2025. hub-prod.perplexity.ai/...perplexity-deep-research
22. ^Perplexity. "Introducing Comet: Browse at the Speed of Thought." July 9, 2025; and "Getting Started with Comet," updated July 16, 2026. hub-prod.perplexity.ai/...introducing-comet ; perplexity.ai/...11172798-getting-started-with-comet
23. ^Perplexity. "Introducing Perplexity Computer." February 25, 2026; and "What Is Computer?", updated July 16, 2026. hub-prod.perplexity.ai/...cing-perplexity-computer ; perplexity.ai/...13837784-what-is-computer
24. ^Perplexity. "Creating Assets with Perplexity: Overview." *Perplexity Help Center*, updated July 16, 2026. perplexity.ai/...g-assets-with-perplexity-overview
25. ^Perplexity. "Quickstart" and "Pricing." *Perplexity API Documentation*. Accessed July 28, 2026. docs.perplexity.ai/...quickstart ; docs.perplexity.ai/...pricing
26. ^Perplexity. "Perplexity Max." *Perplexity Help Center*, updated July 16, 2026. perplexity.ai/...11680686-perplexity-max
27. ^Perplexity. "Enterprise Pricing and Billing: Frequently Asked Questions." *Perplexity Help Center*, updated July 22, 2026. perplexity.ai/...illing-frequently-asked-questions
28. ^Perplexity. "Perplexity Expands Publisher Program with 15 New Media Partners." December 5, 2024. hub-prod.perplexity.ai/...th-15-new-media-partners
29. ^Le Monde. "Artificial Intelligence: Le Monde Signs Partnership Agreement with Perplexity." May 14, 2025. lemonde.fr/...greement-with-perplexity_6741262_115
30. ^Cloudflare. "Perplexity Is Using Stealth, Undeclared Crawlers to Evade Website No-Crawl Directives." August 4, 2025. blog.cloudflare.com/...website-no-crawl-directives
31. ^Perplexity. "How Does Perplexity Follow robots.txt?" *Perplexity Help Center*, updated July 16, 2026. perplexity.ai/...does-perplexity-follow-robots-txt
32. ^*Dow Jones & Company, Inc. et al. v. Perplexity AI, Inc.*, No. 1:24-cv-07984 (S.D.N.Y.), docket. dockets.justia.com/...630270
33. ^*The New York Times Company et al. v. Perplexity AI, Inc.*, No. 1:25-cv-10106 (S.D.N.Y.), docket; *Amazon.com Services LLC v. Perplexity AI, Inc.*, No. 3:25-cv-09514, Document 81 (N.D. Cal. March 9, 2026). dockets.justia.com/...654190 ; law.justia.com/...81
34. ^Financial Times. "Perplexity ARR Tops $450M after Pricing Shift, FT Reports." April 8, 2026, syndicated by Yahoo Finance. ca.finance.yahoo.com/...ops-450m-pricing-132500539
35. ^Reuters. "Perplexity Finalizes $20 Billion Valuation Round, The Information Reports." September 10, 2025. investing.com/...d-the-information-reports-4234608
36. ^Wiggers, K. (2025). "Perplexity Received 780 Million Queries Last Month, CEO Says." *TechCrunch*, June 5, 2025. techcrunch.com/...lion-queries-last-month-ceo-says
37. ^Perplexity. "What Is Perplexity Pro?" *Perplexity Help Center*, updated July 21, 2026. perplexity.ai/...10352901-what-is-perplexity-pro