Perplexity Developers

On internal production classification tasks, both models outperforms our existing classifier and the larger model outperforms Jev 1.14. On one million real Decisions API requests, pplx-decider-v1.3 had the highest agreement with frontier LLM judges of any model test, at 93.98%.

On 17 out-of-domain extraction tasks spanning multiple tasks, both models outperforms general-purpose extraction models of different sizes across all categories.

We're rolling out pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest version of our open-weight decision models. v1.3 adds locate, a new feature that extracts spans based on type or instructions. Now available in our Decisions API, @huggingface, and @OpenRouter. perplexity.ai/hub/blog/perpl…

ultra runs three research agents in parallel and a fourth orchestrates, assigns tasks, and merges the findings. Every child agent runs its own research loop with a separate context, using its own model, tools, and step budget.

medium effort increased its score from 55% to 77% on BrowseComp and DSQA, with almost 90% lower costs. xhigh maintained 76.5% on WANDR, Perplexity's benchmark for long, multi-entity research, with 75% lower costs versus August. ultra holds the highest WANDR score at 85.9%.

Every Agent API effort level now costs less with the same or higher research quality. Effort levels are bundled agent configurations, optimized for specific tasks. A new ultra level is available for the most demanding, long-running requests. pplx.ai/effort-levels

When enabled, Fast Search results join state as evidence before the decision model answers. Pass a URL as the state to ask about a page. Requests are $0.001 per search and $0.02 per 1M tokens. pplx.ai/web-grounded-docs

pplx-decider-v1.2 with web grounding scores 86% on the 100 public PWS-Decisions questions, up from 51% without it. Median latency in the test was 370 ms, with p95 at 685 ms. pplx.ai/web-grounded-blog

The Decisions API can now search the web before it answers. A decision model returns a probability, but only knows what it saw in training, and your state. Add "mode": "web_grounding" and Fast Search runs one query per question, so pplx-decider-v1.2 answers from today's web.

The models are trained differently from other contextual embeddings models. Instead of one gold chunk per query, chunk-level relevance scores are distilled from a query-aware context compression model in Perplexity's answer pipeline. Supporting chunks are no longer negatives.

Introducing pplx-embed-v2-context: open-weight vision-language embedding models in 9B and 0.6B. Both return one int8 vector per document, or one per chunk that sees the whole document. Read the paper, written with @turbopuffer: arxiv.org/abs/2610.38854

After a one-time 5.8 GB model download, the app serves the Decisions API at localhost:8080/v1/decisions. Use it to label files, route messages, check screenshots, or run a pre-commit check, all without data leaving your Mac.

You can now run pplx-decider-flash-v1.2 locally on your Mac, without an API key. Decider for Mac is an app that runs the 4B decision model on device. Ask a question about text or an image and get a probability back, all on your machine. pplx.ai/decider-for-mac