What it is
pplx-decider is Perplexity's family of decision models. It launched with the Decisions API (POST /v1/decisions) in October 2026 at $0.04 per million input tokens and has been updated twice since: v1.1, which cut the price to $0.02, and v1.2 with a smaller pplx-decider-flash-v1.2, published on 15 October. The 27B model is fine-tuned from Qwen3.8-27B and the 4B Flash from Qwen3.5-4B. Weights are on Hugging Face under Apache 2.0.
It answers three question types (choice, a yes/no type called noul, and score) about text, JSON or images up to 1.6 megapixels, with 1 to 255 options per question. Version 1.2 reads the state and all questions in one sequence, so several questions cost little more than one. The model accepts up to 262,144 tokens, but the card says training used sequences of about 32K and recommends staying there.
On Perplexity's API, v1.2 costs $0.02 per million input tokens and Flash $0.01, with free output and no per-request fee. Requests to the older v1 and v1.1 model names are now served by v1.2.
Our review
Perplexity's Decider is the clearest sign that the open side has overtaken Jev on typed decisions. On the community Decision Index 0.4.1, which now weights only private tests (60% same-skill, 40% new-domain), v1.2 ranks fourth of 229 entries at 67.98, against 60.24 and 41st place for Jev 1.14. The index treats scores within about 0.9 points as tied, and fourth to eighth place sit inside 0.5 of each other, so the ranking matters less than the gap of nearly eight points to Jev.
The card is candid about how it got there. Official benchmark training splits were added to the data, decontaminated against the test sets, which is why the private-test result carries more weight than Perplexity's own 68.09 on the public suite. Probabilities remain slightly over-confident (expected calibration error about 0.05), and the card publishes the temperature of 1.382 that corrects it.
Running it yourself is a serious GPU job, not a laptop one. The BF16 weights are about 52 GB, a 243K-token request peaked at 107 GiB on a GB200, and correct results need the bundled pplx_decider package because the model uses non-causal attention and a 255-row readout head; ordinary causal inference in a stock server will not reproduce it. On one GB200 the card reports a 96 ms median request and about 20 requests a second.
The joint-question design has a measurable edge and a measurable cost. Grouping facets of one decision added 11.3 points on SATA-Bench, while grouping competing candidates cost up to 4.8 on ToolRet, so rerankers should still send one candidate per request. Flash, at 4B, falls to 71st with 54.29, a bigger drop than its half price implies.
Where it earns its place
A trust-and-safety engineer screening a million user posts a day
Label each post against a fixed policy list with probabilities, not prose
One choice question per post through the Decisions API at about 600 input tokens: 30 million posts a month is 18 billion tokens, or $360 on v1.2 and $180 on Flash. Output is free, and per-option probabilities make threshold tuning straightforward.
A search engineer reranking 50 candidates per query
Score relevance of retrieved passages before generation
Send each candidate as its own request: the card shows grouping competing candidates in one sequence lowers accuracy by up to 4.8 points. At $0.02 per million tokens, 50 candidates of 400 tokens cost $0.0004 per query.
A compliance lead at a European insurer that cannot send claim photos to a US API
Triage damage photos and forms on in-house hardware
Run the Apache 2.0 weights on a GB200 or similar GPU with more than 52 GB free, using the pplx_decider package. No per-token bill and no data leaves the building; the cost is the GPU and the engineering to keep pace with releases.
Before you commit
- Setup
- API: POST /v1/decisions with model pplx-decider-v1.2 or pplx-decider-flash-v1.2. Self-host: pip-install the bundled pplx_decider package and load with Decider.from_pretrained; call warmup() first, since the fast path takes about a minute to compile.
- Integrations
- Perplexity API with per-request token counts in usage.input_tokens; Hugging Face weights. Stock vLLM or SGLang serving is not documented and causal inference does not reproduce the checkpoint.
- Data and privacy
- Perplexity says API customer content is not used to train generative models; retention varies by model and zero data retention is available only by written agreement. SOC 2 Type II. Weights are Apache 2.0, with no usage restrictions.
- Cost at scale
- 1 million decisions at 600 tokens each costs $12 on v1.2 and $6 on Flash, against about $25 on Jev. Self-hosted, one GB200 handles about 20 requests a second on the fast path (25 with optional FP8), roughly 1.7 million a day at full load.
- Limits
- 1 to 255 options per question; images capped at 1.6 MP (about 1,600 positions each). Trained to about 32K tokens; over-long requests are split, so questions stop being read together. Decisions API rate limits are not published.
What's good
- Fourth of 229 on the private-test Decision Index, nearly eight points above Jev 1.14
- Apache 2.0 weights for the 27B and 4B models
- Reads images up to 1.6 MP
- $0.02 per 1M input tokens; output free
What's not
- Self-hosting needs about 52 GB for weights and Perplexity's own inference package
- Trained on sequences of about 32K; longer inputs only spot-checked
- Old model names now route to v1.2 without a version pin
- Flash scores far lower, 71st on the same index
Best for
Teams that want strong typed decisions over text and images, either through a cheap API or on their own GPUs.
Alternatives
Perplexity Decider (pplx-decider): common questions
- Is Perplexity Decider (pplx-decider) worth it?
- A decision model you can download, read images with and run for under half the price of Jev, scoring well above Jev on tests nobody can train on. We rate it 4.5 out of 5.
- How much does Perplexity Decider (pplx-decider) cost?
- API: pplx-decider-v1.2 $0.02 and Flash $0.01 per 1M input tokens; output free, no per-request fee. Weights free under Apache 2.0.
- What are the downsides of Perplexity Decider (pplx-decider)?
- Self-hosting needs about 52 GB for weights and Perplexity's own inference package; trained on sequences of about 32K; longer inputs only spot-checked; old model names now route to v1.2 without a version pin; flash scores far lower, 71st on the same index.
- Who is Perplexity Decider (pplx-decider) best for?
- Teams that want strong typed decisions over text and images, either through a cheap API or on their own GPUs.
- How much does Perplexity Decider (pplx-decider) cost for a team or in production?
- 1 million decisions at 600 tokens each costs $12 on v1.2 and $6 on Flash, against about $25 on Jev. Self-hosted, one GB200 handles about 20 requests a second on the fast path (25 with optional FP8), roughly 1.7 million a day at full load.
- Is Perplexity Decider (pplx-decider) safe for company data?
- Perplexity says API customer content is not used to train generative models; retention varies by model and zero data retention is available only by written agreement. SOC 2 Type II. Weights are Apache 2.0, with no usage restrictions.
- What are the best Perplexity Decider (pplx-decider) alternatives?
- The closest alternatives we have reviewed are d3 (vLLM Semantic Router), Jev, Clef, GLiNER2.5-Decide and Laya.
Answers checked Oct 22, 2026.
Make Perplexity Decider (pplx-decider)? Show this rating on your site
