Perplexity drops dirt-cheap open decision models

Perplexity has open-sourced two multimodal decision models, pplx-decider-flash-v1.2 and pplx-decider-v1.2, built for lightweight routing and agentic tasks at rock-bottom prices.

The pair comes in 4B and 27B parameter sizes, available both as open weights on Hugging Face and through the Perplexity Decisions API. Calling the API costs just 1 cent per million tokens for the 4B flash variant, and 2 cents for the 27B model. CEO Aravind Srinivas noted that future plans include input caching and serving optimizations to drive those costs down even further.

Faster inference and benchmark rankings

Both models post higher scores across text and vision benchmarks. The key architectural perk is throughput: the models can run all questions within a single request in one pass, driving inference speeds up to 5x faster.

On the Hugging Face Decision Index, the 4B flash edition ranks number one among all 4B models, while the 27B edition sits at second place overall. You can grab the weights directly on Hugging Face to self-host, or plug into the hosted API.

Why it matters

If you build agent loops or complex decision trees, this release is worth testing immediately. Routing models get hit constantly, so paying 1 to 2 cents per million tokens while cutting round-trips can drastically slash both API bills and latency.

We think Perplexity made a smart move by commoditizing the decision layer. Providing competitive open weights alongside dirt-cheap inference makes high-frequency multi-step agent architectures practical for solo developers and small teams without enterprise budgets.