The Perplexity Decisions API is a scoring layer that picks an answer from a fixed set instead of generating text. It accepts text, JSON, or images and returns a choice, possibly with a probability for each option.
The endpoint skips parsing because it treats content as an opaque input, so no field extraction or rule-based logic is needed. That speed and simplicity come with a real tradeoff: the final label may hide the evidence behind it.
Agencies can use ordered rubrics to grade content or rank leads without writing a parser. The practical question is whether the answer object exposes enough probability detail to set thresholds and escalate weak matches.
I have enough context now. Here are my 9 subheadings:
The promise of scoring content without a parser hinges on a narrow design choice. The Decisions API does not generate text at all. It picks one answer from a finite set that a developer defines. That answer arrives in about 150 milliseconds, according to OpenAI's DevDay presentation on September 29, 2026.
A standard GPT-6 Luna call takes roughly 1.6 seconds by the same comparison. The speed gap is real, but it is not the whole story. There is still no public documentation or pricing for the endpoint as of the announcement.
The practical question is whether each decision returns per-option probabilities. Those probabilities would let a team set thresholds, abstain on weak matches, or escalate hard cases. Without them, a score is just a label.
The implication is simple: Treat the endpoint as a decision layer, not a content writer. Its value for agencies depends on what the answer object actually exposes.
What Is the Perplexity Decisions API and Why Skip Parsing?
The Perplexity Decisions API is a scoring layer, not a parser. It accepts text, JSON, or images and returns a choice from a defined set. That choice can include a probability for each option. The API skips parsing because it treats content as an opaque input.
There is no need to extract fields, clean markup, or build rule-based logic. A developer sends the raw material and the model returns a decision. The tradeoff is real. The endpoint does not expose intermediate search results or their metadata as distinct data, based on a MindStick question from October 15, 2026.
That question asked whether the broader Perplexity API separates search results from its written answer. It remained unanswered at the time. The pattern matters for agencies. A scoring endpoint that returns only a final label may hide the evidence behind that label.
Teams then have to trust the model without inspecting the reasoning. The value is speed and simplicity. The cost is less control over the underlying signal. That is the core reason to skip parsing. It is also the core reason to ask what the answer object actually returns.
Scoring Content With Ordered Rubrics: How the Score Question Works
An ordered rubric turns a fuzzy judgment into a fixed set of choices. The score question works by forcing a selection from that set. It never asks the model to write a sentence. It asks which option fits best.
The answer comes back as a label, not a paragraph. That label is the score.
- What does the score question actually return? It returns a chosen option from the developer-defined list, and it may include a probability for each option.
- How does order affect the decision? The order of the rubric options can act as a ranking scale, so the model compares the content against each defined level.
- When does a score question fail? It fails when the rubric levels overlap or when the content contains mixed signals that do not match any single option cleanly.
How Probability-Weighted Scores Work vs. Raw Classification Labels
A raw label names a category. A probability-weighted score expresses how strongly the model supports that choice.
| Criterion | Probability-weighted score | Raw classification label |
|---|---|---|
| Decision strength | Shows relative confidence across options | Shows only the selected option |
| Threshold control | Teams can set a minimum acceptance level | No built-in cutoff exists |
| Uncertainty handling | Weak matches become visible in low weights | Weak matches hide inside a single label |
Decisions API vs. Asking ChatGPT for a Label: At a Glance
A direct label and a calibrated score serve different routing jobs.
| Criterion | Decisions API | Asking ChatGPT for a label |
|---|---|---|
| Output type | Fixed-choice probabilities | Free-text label or phrase |
| Schema control | Typed schema constrains answers | Prompt wording guides the reply |
| Decision consistency | Same option set each call | Wording may drift across calls |
Steps to Grade and Route Agency Content Without Writing a Parser
How can an agency grade and route content without building a parser?
- Start with a fixed decision question rather than a free-form request. Define an ordered rubric with three to five levels, and give each level a short label tied to a distinct content state.
- Provide the content in its available form and keep preprocessing limited to what the workflow requires. This can reduce implementation work, but confirm that the selected endpoint and input format support the content you plan to submit.
- Treat the returned score or confidence information as a review signal, not as proof that the classification is correct. Send low-confidence, ambiguous, or high-impact cases to a human reviewer instead of routing them automatically.
- Map every rubric level to a downstream action before production use. For example, one level might enter a publishing queue, another might require revision, and a third might be held for review, making the routing logic easier to test and audit.
How Agencies Use Score Questions to Rank Leads and Prioritize Work
A score question can turn a mixed lead list into a clear work queue.
- What makes a lead rank higher than another? The rubric order itself becomes the ranking scale when each level matches a distinct readiness state.
- How does probability information change prioritization? Low weights on a high-value option can trigger a human review instead of an automatic pass.
- Where should agencies apply score questions first? Start with repetitive intake tasks where the cost of a wrong route is easy to measure and fix.
The Perplexity Decisions API can score content without parsing, but the answer is qualified. It returns a choice from a fixed set, not generated text. The main value is speed and simple routing. The main limitation is hidden evidence behind the final label.
Probability detail may exist for each option, but public documentation was not available at the announcement. Agencies should treat the score as a review signal, not proof. Send low-confidence or high-impact cases to a human reviewer before production use.
