Perplexity Decisions API: $0.04 per million, open weights too

Perplexity launched the Perplexity Decisions API this week. It runs on pplx-decider-v1-27b, a decision model that returns probabilities over answers you define instead of writing text. The price is $0.04 per million input tokens, and output is free. Perplexity also put the weights on Hugging Face under Apache 2.0. So you can rent it by the token or run it yourself. That combination changes the math on every classification call in your stack.

What actually happened

From the Decisions API docs:

• Endpoint: POST https://api.perplexity.ai/v1/decisions. Input can be text, JSON, arrays or images (PNG, JPEG, WebP as base64).
• Three question types: yes/no (a probability from 0 to 1), choice (top option plus the full distribution) and score (a probability-weighted level).
• Batching: 1 to 128 named questions per request.
• Price: $0.04 per million input tokens, free output, no per-request fee.
• Rate limit: 10 requests per second per organization on every plan.

From the model card:

• Base: Qwen3.8-27B, fine-tuned for decisions. Apache 2.0.
• Accuracy: 85.71% averaged over 11 benchmarks, against 84.51% for TypeSafe's Jev and 74.76% for the base Qwen model. It loses to Jev on the hard JevBench public set (70.30% vs 73.27%).
• Hardware to self-host: about 49 GB of GPU memory plus working space.

This is the third decision model we have covered in two days, after Strands Decider 2B and OpenAI's Decisions API. The category is getting crowded, and crowded categories get cheap.

Why the Perplexity Decisions API matters for your business

The price of a label just hit the floor. A support ticket of 500 tokens costs $0.00002 to route. Ten thousand tickets a month is twenty cents. If you still send "refund, sales or spam?" to a frontier chat model and parse the reply, you are paying for prose you throw away.

Hosted and self-hosted, same weights. This is the part we like. Start on the API. If volume grows, or a client will not let customer data leave their network, move the same model to your own GPU. Your prompts and thresholds do not change. That is an exit plan built into the vendor choice.

Watch the rate limit. Ten requests per second is fine for a shop's inbox. It is not fine for scoring a full product catalog in one burst. Use the 128-question batching, or self-host for bulk jobs.

Benchmarks are not your data. 85.71% on public tests still means mistakes. Label 200 of your own tickets, run them through, and set the confidence threshold where wrong answers stop costing you money.

Sources: Perplexity Decisions API docs, Hugging Face: pplx-decider-v1-27b.