Choose between Standard, Flash, and Jev

All three return structured judgments that your code can use: a yes/no probability, a choice, or a rubric score. Choose a candidate from the work it must do, then evaluate the resulting application decisions on your own examples.

Where to start

Standard and Flash belong to Perplexity; Jev is a separate TypeSafe service. If you are building with Perplexity and unsure where to begin, evaluate Standard first. Then test Flash if you have clear, repetitive judgments and want to assess its economics at your required quality level. If you are comparing providers for independent judgments on text, include Jev—especially when several questions use the same information.

The Standard and Flash guidance follows the Perplexity quickstart ↗. The Jev suggestion is this guide's, based on the question isolation described in the TypeSafe introduction ↗. These are not measured rankings. The categories overlap: all three can serve the support workflow below. None removes the need to evaluate your workflow.

Jev and Flash both evaluate questions independently. Jev ingests state once for a multi-question request; hosted Flash repeats and bills it for each question. That makes Jev worth including in a comparison of independent text judgments. Ingesting state once does not by itself establish a once-per-state billing rule. This distinction does not prove a quality, speed, or total-cost advantage; rates, actual billable usage, and task performance also matter.

How the requests differ

A request is one API call containing state and questions. A processing group is the set of questions Standard evaluates together; a large request can be split into several groups.

Imagine a support message with three questions: does it need human review, which team should own it, and how urgent is it?

Standard evaluates related questions together. Within a processing group, each question sees the state and the other questions. That can help when the questions describe facets of one decision. For independent candidate scoring, such as rating unrelated documents against the same criterion, start with separate requests. If a request is too large, the API splits its questions into consecutive groups and repeats the state for each; state plus any single question must still fit the limit.

The same question can receive a different score depending on its group. Keep grouping fixed while tuning thresholds; if grouping changes, re-evaluate those thresholds. Leave input headroom so automatic splitting does not unexpectedly change which questions see each other.

Hosted Flash evaluates each question separately. Start with one question per request. If several arrive in one request, the API still evaluates them separately and bills the state again for each question. Packing questions together does not create Standard's joint context or shared-state savings.

Jev evaluates questions independently against shared state. Questions can be submitted together, but each is evaluated in isolation. TypeSafe recommends atomic questions—one specific judgment per question—whose results your code combines. Put the necessary evidence in state; do not expect one question's instructions or answer to supply another's missing context.

Question isolation does not mean responses are guaranteed to be numerically identical across repeated calls. Measure whether variation changes important application actions.

See the same support workflow across all three. These descriptions concern the hosted services. A downloadable model checkpoint or a different serving implementation can have a different contract.

Inputs, limits, and cost

Documentation checked 20 October 2026. Prices below are USD per million input tokens; all three currently have free output tokens.

Sources: Perplexity models, grouping, and pricing ↗, TypeSafe models and limits ↗. Jev has an optional long-context mode; TypeSafe warns it is slower and less accurate, so assess it separately rather than treating its expanded limits as the default. Perplexity image input has additional format, size, and request-body limits; see the API reference.

For three questions over a 10,000-token state, the state alone costs about $0.0002 on Standard in one group versus $0.0003 on Flash. This is arithmetic from documented rates and repetition rules, excluding questions, other overhead, and retries. Tokenization and usage can differ by provider. Compare actual billed usage, quality, latency, and review workload—not rates alone. The guide's cost explanation shows how to estimate a complete workflow.

For Jev, 10,000 billable state tokens at $0.042/M would cost $0.00042 if counted once, above both Perplexity figures in this illustration. This is a conditional estimate, not verified billing: TypeSafe documents ingestion once, but the inspected API usage definition ↗ does not explicitly specify repeated-state billing. Compare actual returned usage and account for different tokenizers.

High-volume suitability is not a throughput guarantee: Perplexity's documented default limits ↗ are 4 million tokens per minute for Flash versus 15 million for Standard, with 10 requests per second shared across both; check your workload and account limits, and request higher limits if needed.

Account for optional web grounding

Grounding rules checked 5 November 2026.

Standard and Flash both support optional built-in web grounding. This adds retrieval before inference; their different question-processing rules still apply. Fetched page text and search evidence count as input tokens, repeated per Standard processing group or per Flash question. The state-only examples above exclude these retrieval charges and added evidence tokens. See where evidence comes from and the grounding reference.

On both Perplexity models, changing the grounded question set can change the shared retrieved evidence and potentially other answers. Flash still evaluates questions separately. Keep the grounded question set and retrieval configuration fixed while tuning thresholds; this retrieval implication has not been measured here.

Compare the complete evidence-and-decision workflow. A model's input-token rate does not establish which grounded workflow is cheapest or most accurate, and a separate service's shared retrieval does not establish joint model evaluation.

Interpreting the answers

• Yes/no probability estimates whether the supplied statement is true. Near 0.5 means uncertainty, not medium urgency. noul has no separate confidence field.
• Choice selects among the options you supplied. Good option coverage matters: a confident answer cannot repair a missing category.
• Score averages over ordered rubric levels. The average can hide uncertainty between extremes, so inspect the distribution when it matters.
• Confidence summarizes a distribution; it is not a demonstrated accuracy rate. Perplexity and TypeSafe use different score-confidence normalization, so thresholds should not be transferred blindly.

The worked example connects these signals to rules. Exact formulas, version identity, and migration checks are optional detail below.

Practitioner checklist

Check your understanding

General model behavior was checked on 20 October 2026; the web-grounding additions were checked on 5 November 2026. No model-quality or latency comparison was run for this guide. Recheck linked provider documentation before implementation and evaluate thresholds on your own data.

Back to the guide · Follow the worked example · Choose a prompt