Decision model
Perplexity Decider 27B v1.3
1 question per request
- Accuracy, all 2,198 questions
- 68.4%66.5% to 70.3%, 95% interval; 1,504 right
- Answers kept at 99% confidence
- 276 of 2,1980 wrong; share right 98.6% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.081$81.18Basis and price under Cost
- Latency
- 0.110 smedian request, 1 question per request
Perplexity Decider 27B v1.2 · API and Perplexity Decider 27B v1.2 · API · with web, with Perplexity's web search off and on, are compared on Web grounding.
How we called it
- Kind
- Decision model
- Access
- Perplexity Decider server, self-hosted
- Paid as
- Our own hardware
- Model ID
pplx-decider-v1.3- Variant
- local
- Parameters
- 27B dense, as the maker states it (counted from the released files: 26.09B).language model (no LM head) + vision encoder + 255-way readout; its locate-span adapter and head (0.47B), used only by its locate task, not counted; source: counted from the checkpoint's safetensors, perplexity-ai/pplx-decider-v1.3@9c2ebb2, huggingface.co/perplexity-ai/pplx-decider-v1.3; read 9 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Perplexity's released v1.3 checkpoint, the next version of Decider 27B v1.2 on Qwen3.8-27B, run on our hardware through the package it ships.
- Confidence from
- Option score
- Run
- 1134, 14 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
015175ac1bde- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 68.4% | 66.5% to 70.3% | 1,504 of 2,198 |
| Reading | 89.0% | 86.4% to 91.2% | 584 of 656 |
| Grammar | 59.7% | 57.2% to 62.1% | 920 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
- 490 of 547, first English run (1137)
- Accuracy
- 89.6%
- Non-hallucination with “I don't know”
- 8.8%: 5 “I don't know” of 57 not right (run 1139)
- Answered with “I don't know”
- 99.1%: 542 of 547
- Error rate with “I don't know”
- 9.6%: 52 of 542 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.08382.63from GPU time
- Latency, median request, first English run
- 0.115 s; 1 request at once
- English run
- 1137
- Machine, English runs
- RTX PRO 6000
- Date
- 14 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 276, of which 0 wrong (share right: 98.6% to 100.0%, 95% interval); 1,922 skipped, the model less than 99% sure
- Calibration gap
- 5.0 points
- AUROC, all questions
- 0.82 0.81 to 0.84
- AUROC, reading
- 0.91 0.89 to 0.94
- AUROC, grammar
- 0.76 0.74 to 0.78
- Brier score
- 0.159
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 91.2% right; most confident four fifths: 76.8% right
- Typical score of a right answer
- 90.4%
- Typical score of a wrong answer
- 53.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.110 s (this run)
- Latency, 90th percentile
- 0.145 s
- Requests at once
- 1
- Throughput
- 511.2 questions per minute of request time
- Request time
- 4.3 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
- $0.081$81.18
- Whole run
- $0.18
- Basis
- Estimated from GPU time: the run's request time x the GPUs its server held x Runpod's on-demand price for an RTX PRO 6000, $2.49 per GPU-hour (read 5 November 2026). A cloud-equivalent estimate, not measured spending.
- Price
- $2.49 per GPU-hour, 1 GPU, for the run's request time (source, 2026-11-05)
- Input tokens per question
- 372
- Output tokens per question
- 0
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 | 23.6% | −0.13 points | 20% |
| GPT-6 Astra high · chat | 22.4% | −0.11 points | 21% |
| GPT-6.1 Sol high · chat | 26.0% | −0.13 points | 21% |
| GPT-6 Astra low · chat | 22.4% | −0.11 points | 21% |
| GPT-6 Astra low · chat · 1 per request | 22.3% | −0.10 points | 22% |
| Gemini 3.8 Flash low · chat · 1 per request | 26.5% | −0.13 points | 18% |
| GPT-6.1 Sol low · chat · 1 per request | 27.3% | −0.17 points | 26% |
| GPT-6.1 Sol low · chat | 24.1% | −0.12 points | 17% |
| Claude Opus 5.5 high · chat | 23.6% | −0.13 points | 22% |
| GPT-6.1 Terra high · chat | 22.3% | −0.10 points | 7% |
| Claude Opus 5.5 low · chat · 1 per request | 24.3% | −0.16 points | 24% |
| Gemini 3.8 Flash low · chat | 25.5% | −0.19 points | -5% |
| GPT-6.1 Terra low · chat | 31.4% | −0.16 points | 12% |
| Claude Opus 5.5 low · chat | 32.7% | −0.26 points | 31% |
| Claude Sonnet 5.5 high · chat · 1 per request | 33.8% | −0.26 points | 33% |
| Claude Sonnet 5.5 high · chat | 49.2% | −0.05 points | 47% |