Decision model
Perplexity Decider 27B v1.2 API · with web
with web · 1 question per request
- Accuracy, all 2,198 questions
- 68.7%66.8% to 70.6%, 95% interval; 1,511 right
- Answers kept at 99% confidence
- 351 of 2,1981 wrong; share right 98.4% to 99.9%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- ≤ $1.034$1,034Basis and price under Cost
- Latency
- 0.482 smedian request, 1 question per request
The same model with Perplexity's web search off is Perplexity Decider 27B v1.2 · API: Web grounding compares the two.
How we called it
- Kind
- Decision model
- Access
- Perplexity Decisions API
- Paid as
- Provider service
- Model ID
pplx-decider-v1.2- Variant
- with web
- Parameters
- 27B dense, as the maker states it (counted from the released files: 26.09B).language model (no LM head) + 0.46B vision encoder + 255-way readout; source: counted from the checkpoint's safetensors, perplexity-ai/pplx-decider-v1.2@225c913 (the same weights as cce0f3f), huggingface.co/perplexity-ai/pplx-decider-v1.2; read 9 November 2026
- Note
- Perplexity's hosted Decider v1.2 through its Decisions API with web grounding on: a web search per question, its results added as evidence. The search sends the question out, and for some questions it returned pages of the bank's source book (the Web grounding page). Its cost is an upper-bound estimate: one $0.001 search billed per request, while Perplexity bills only completed searches.
- Confidence from
- Option score
- Run
- 640, 5 November 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
b0c7b9874aab- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 68.7% | 66.8% to 70.6% | 1,511 of 2,198 |
| Reading | 88.9% | 86.2% to 91.1% | 583 of 656 |
| Grammar | 60.2% | 57.7% to 62.6% | 928 of 1,542 |
Comparison group: zABCDE. Compare with another setup.
Accuracy by topic
English
The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.
- Right answers
- 495 of 547, first English run (639)
- Accuracy
- 90.5%
- Non-hallucination with “I don't know”
- 15.2%: 10 “I don't know” of 66 not right (run 644)
- Answered with “I don't know”
- 98.2%: 537 of 547
- Error rate with “I don't know”
- 10.4%: 56 of 537 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- ≤ 1.036≤ 1,036at list prices
- Latency, median request, first English run
- 0.484 s; 1 request at once
- English run
- 639
- Machine, English runs
- hosted API
- Date
- 5 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 351, of which 1 wrong (share right: 98.4% to 99.9%, 95% interval); 1,847 skipped, the model less than 99% sure
- Calibration gap
- 3.6 points
- AUROC, all questions
- 0.83 0.81 to 0.84
- AUROC, reading
- 0.92 0.89 to 0.94
- AUROC, grammar
- 0.76 0.74 to 0.79
- Brier score
- 0.157
- Accuracy of its surest answers
- most confident fifth: 99.5% right; most confident half: 90.9% right; most confident four fifths: 77.0% right
- Typical score of a right answer
- 90.0%
- Typical score of a wrong answer
- 48.7%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.482 s (this run)
- Latency, 90th percentile
- 0.684 s
- Requests at once
- 1
- Throughput
- 106.8 questions per minute of request time
- Request time
- 20.6 minutes, over the requests whose answers were kept
Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.
Cost
- Cost per 1,000 questionsper 1M questions
- ≤ $1.034$1,034
- Whole run
- ≤ $2.27
- Basis
- Estimated from published prices and recorded token usage
- Price
- $0.02 input, output free per 1M tokens, plus $0.001 per request (one web search) (source, 2026-11-05)
- Input tokens per question
- 1,710
- Output tokens per question
- 1
≤ Upper-bound estimate: one $0.001 web search counted per request; Perplexity bills only completed searches.
Going first in front of an LLM
Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.
| LLM behind it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flash high · chat | 16.8% | −0.09 points | -26% |
| GPT-6 Astra high · chat | 18.6% | −0.13 points | 2% |
| GPT-6.1 Sol high · chat | 21.9% | −0.17 points | -42% |
| GPT-6 Astra low · chat | 18.7% | −0.14 points | -0% |
| GPT-6 Astra low · chat · 1 per request | 16.8% | −0.09 points | 13% |
| Gemini 3.8 Flash low · chat · 1 per request | 16.8% | −0.09 points | -91% |
| GPT-6.1 Sol low · chat · 1 per request | 20.0% | −0.16 points | 1% |
| GPT-6.1 Sol low · chat | 18.7% | −0.14 points | -73% |
| Claude Opus 5.5 high · chat | 17.0% | −0.10 points | 1% |
| GPT-6.1 Terra high · chat | 19.1% | −0.15 points | -182% |
| Claude Opus 5.5 low · chat · 1 per request | 21.9% | −0.16 points | 13% |
| Gemini 3.8 Flash low · chat | 17.0% | −0.10 points | -368% |
| GPT-6.1 Terra low · chat | 29.0% | −0.19 points | -214% |
| Claude Opus 5.5 low · chat | 25.2% | −0.19 points | 4% |
| Claude Sonnet 5.5 high · chat · 1 per request | 29.8% | −0.20 points | 16% |
| Claude Sonnet 5.5 high · chat | 50.5% | −0.10 points | 22% |