Decision model
Perplexity Decider Flash v1.3
1 question per request
- Accuracy, all 2,198 questions
- 49.7%47.6% to 51.8%, 95% interval; 1,093 right
- Answers kept at 99% confidence
- 29 of 2,1980 wrong; share right 88.3% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.025$24.69Basis and price under Cost
- Latency
- 0.033 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-flash-v1.3- Variant
- local
- Parameters
- 4B dense, as the maker states it (counted from the released files: 4.54B).language model (tied embeddings) + vision encoder + 255-way readout; its locate-span adapter and head (0.13B), used only by its locate task, not counted; source: counted from the checkpoint's safetensors, perplexity-ai/pplx-decider-flash-v1.3@d06fb3f, huggingface.co/perplexity-ai/pplx-decider-flash-v1.3; read 9 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Perplexity's released Flash v1.3 checkpoint, distilled from Decider 27B v1.3 on Qwen3.5-4B, run on our hardware through the package it ships.
- Confidence from
- Option score
- Run
- 1123, 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
5424d93c2c6c- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 49.7% | 47.6% to 51.8% | 1,093 of 2,198 |
| Reading | 74.4% | 70.9% to 77.6% | 488 of 656 |
| Grammar | 39.2% | 36.8% to 41.7% | 605 of 1,542 |
Comparison group: βγδζηθ. 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
- 458 of 547, first English run (1126)
- Accuracy
- 83.7%
- Non-hallucination with “I don't know”
- 7.2%: 7 “I don't know” of 97 not right (run 1128)
- Answered with “I don't know”
- 98.7%: 540 of 547
- Error rate with “I don't know”
- 16.7%: 90 of 540 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.02323.17from GPU time
- Latency, median request, first English run
- 0.033 s; 1 request at once
- English run
- 1126
- 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
- 29, of which 0 wrong (share right: 88.3% to 100.0%, 95% interval); 2,169 skipped, the model less than 99% sure
- Calibration gap
- 12.6 points
- AUROC, all questions
- 0.74 0.72 to 0.76
- AUROC, reading
- 0.83 0.80 to 0.86
- AUROC, grammar
- 0.64 0.61 to 0.67
- Brier score
- 0.221
- Accuracy of its surest answers
- most confident fifth: 89.6% right; most confident half: 66.0% right; most confident four fifths: 55.1% right
- Typical score of a right answer
- 79.0%
- Typical score of a wrong answer
- 48.8%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.033 s (this run)
- Latency, 90th percentile
- 0.049 s
- Requests at once
- 1
- Throughput
- 1681.0 questions per minute of request time
- Request time
- 1.3 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
- $0.025$24.69
- Whole run
- $0.054
- 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 | 8.7% | −0.09 points | 8% |
| GPT-6 Astra high · chat | 8.7% | −0.09 points | 8% |
| GPT-6.1 Sol high · chat | 8.7% | −0.09 points | 7% |
| GPT-6 Astra low · chat | 8.8% | −0.09 points | 8% |
| GPT-6 Astra low · chat · 1 per request | 8.7% | −0.09 points | 9% |
| Gemini 3.8 Flash low · chat · 1 per request | 8.8% | −0.10 points | 6% |
| GPT-6.1 Sol low · chat · 1 per request | 8.8% | −0.10 points | 8% |
| GPT-6.1 Sol low · chat | 8.8% | −0.09 points | 7% |
| Claude Opus 5.5 high · chat | 8.8% | −0.09 points | 8% |
| GPT-6.1 Terra high · chat | 8.7% | −0.09 points | 4% |
| Claude Opus 5.5 low · chat · 1 per request | 8.8% | −0.09 points | 9% |
| Gemini 3.8 Flash low · chat | 8.8% | −0.10 points | -0% |
| GPT-6.1 Terra low · chat | 8.9% | −0.10 points | 3% |
| Claude Opus 5.5 low · chat | 8.8% | −0.09 points | 8% |
| Claude Sonnet 5.5 high · chat · 1 per request | 9.3% | −0.09 points | 9% |
| Claude Sonnet 5.5 high · chat | 12.3% | −0.12 points | 12% |