Perplexity's Open Decider v1.3 Costs a Fifth of Gemini Decision 1 per Task on Artificial Analysis. It Scores 70, Exactly What v1.2 Scored, and It Got Slower.
Perplexity released pplx-decider-v1.3, a 27-billion-parameter classifier built on Qwen3.8-27B, and pplx-decider-flash-v1.3 at 4 billion, on November 12 under Apache 2.0 on Hugging Face, on OpenRouter, and in its Decisions API at $0.02 and $0.01 per million input tokens with free output. New in v1.3 is locate, which returns verbatim text spans with character offsets for entity recognition, PII detection, and extractive question answering. Artificial Analysis scores v1.3 at 70 on Decision Index v2.1, sixth of 21, at $68 per million tasks and 289 milliseconds, against Gemini Decision 1 at 66 and $345. Decider v1.2 also scored 70 at $68, in 229 milliseconds. Perplexity's reported gain, 91.6% mean accuracy across 63 query-classification tasks, is on its own tasks with references relabeled by a frontier LLM. We read the docs, the model card, and AA's pages.
By T.W. Ghost
Perplexity released pplx-decider-v1.3 and pplx-decider-flash-v1.3 today, November 12, 2026, one day after Google's Gemini Decision 1. Both are open weights under Apache 2.0 on Hugging Face, both are on OpenRouter, and both serve through Perplexity's Decisions API at $0.02 and $0.01 per million input tokens with output free, a 256,000-token context, up to 255 options per question, and up to 10 score levels. The model card describes v1.3 as "a multimodal classification model on Qwen3.8-27B," about 49 GiB in BF16; Flash is 4 billion parameters. New in this version is locate, which returns verbatim text spans with character offsets for named-entity recognition, PII detection, and extractive question answering.
Artificial Analysis scores it on Decision Index v2.1 at 70, sixth of 21 decision models, at $68 per million tasks and 289 milliseconds per task. Gemini Decision 1 sits at 66 and $345, eleventh of 21. That is the headline, and it is true. The catch is the row above it on AA's own pages: Decider v1.2 also scored 70, also at $68, and did it in 229 milliseconds. The improvement Perplexity reports, 91.6% mean accuracy across 63 query-classification tasks and 6.54 points over its production classifiers, is measured on Perplexity's tasks with references relabeled by a frontier LLM. We read the Perplexity post and community note, the Decisions API docs, the Hugging Face card, and AA's model pages. Here is what held up.
What You Are Calling
| pplx-decider-v1.3 | pplx-decider-flash-v1.3 | |
|---|---|---|
| Size; base | 27B, on Qwen3.8-27B; about 49 GiB BF16 | 4B |
| License; where | Apache 2.0 on Hugging Face (repo created Nov 9, not gated); OpenRouter; Decisions API | same |
| Price | $0.02 per 1M input; output free | $0.01 per 1M input; output free |
| Limits | 256K context; up to 255 options; up to 10 score levels | same |
| Locate | Text only; max 128 questions; not with web grounding; trained to 8K tokens, evaluated to 32K; mixing locate and decisions costs extra forward passes | |
| AA Decision Index v2.1 | 70; 6th of 21; 289 ms; $68 per 1M tasks | 64; 12th of 21; 178 ms; $34 per 1M tasks |
| Card discrepancy | Card text says "The repository is private"; the Hugging Face API reports it public |
The Board, by Who Reported It
| Claim | Number | Who |
|---|---|---|
| Decider v1.3 | 70; 289 ms; $68 per 1M tasks | Independent (Artificial Analysis) |
| Decider Flash v1.3 | 64; 178 ms; $34 | Independent (AA) |
| Decider v1.2 | 70, 7th of 21; 229 ms; $68 | Independent (AA) |
| Gemini Decision 1; board median | 66, 11th of 21, 211 ms, $345; median 66, 206 ms, $137 | Independent (AA) |
| 91.6% mean accuracy on 63 query-classification tasks; 6.54 points over production classifiers | Vendor (references relabeled by a frontier LLM; via Unite.AI and the Perplexity blog) | |
| Free on OpenRouter until Nov 19 | Unverified (X summary only) |
Three honest observations:
- Flat on the independent board, and slower. The same 70, the same $68, and 289 milliseconds where v1.2 took 229. Whatever v1.3 improved, AA's six evaluations did not register it.
- The gain is vendor-graded. Sixty-three tasks, Perplexity's own, with reference labels rewritten by a frontier LLM. Real for Perplexity's product; not a board.
- Open, and a fine-tune. The download is Alibaba's Qwen3.8-27B base with Perplexity's heads and adapters, under Apache 2.0. That is the cheapest way into this race, and it is also a dependency on someone else's base model.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Nov 12, 2026; Decisions API, Hugging Face, OpenRouter | Verified (vendor) |
| Apache 2.0; repo created Nov 9; not gated | Verified (Hugging Face API) |
| Built on Qwen3.8-27B; about 49 GiB BF16 | Verified (model card) |
| Card says private; API says public | Verified discrepancy |
| Prices; context; options; score levels | Verified (docs) |
| Locate limits | Verified (docs; model card) |
| AA rows for v1.3, Flash v1.3, v1.2, Gemini Decision 1, and the median | Verified (AA pages) |
| 91.6%; 6.54 points | Vendor |
| OpenRouter free until Nov 19 | Unverified |
Who Should Use It
Use Decider v1.3 if you want the cheapest row near the top of AA's decision board that you can also download, with extraction through locate for text under 32,000 tokens.
Skip the upgrade story; if you run v1.2, the independent score is the same and the latency is worse, and skip locate for anything that needs web grounding or more than 128 questions.
The frontier is elsewhere: on the same board Drex Max v1.1 and GLiDE v1.2 are at 73, and the most calibrated thing AA has measured is still a general LLM at max effort. Our model comparison has the current standings side by side.
Sources
- Perplexity: Decider
- Hugging Face: perplexity-ai/pplx-decider-v1.3
- Perplexity docs: Decisions API quickstart
- Perplexity community: rolling out pplx-decider-v1.3 and pplx-decider-flash-v1.3
- Artificial Analysis: pplx-decider-v1.3
- Artificial Analysis: pplx-decider-flash-v1.3
- Artificial Analysis: pplx-decider-v1.2
- Artificial Analysis: Gemini Decision 1
- Unite.AI: Perplexity releases Decider v1.3 models with text span extraction
*Not sure whether your classifier should be a fine-tune you download or an API you call? Take the free 2-minute quiz and get matched. Then read about the model Google priced at five times the cost per task: Gemini Decision 1.*