Perplexity launches a Decisions API, and the model behind it is AutoJev-27B

Last updated: October 1, 2026

@perplexitydevs launched the Perplexity Decisions API on 1 October 2026 at 15:24 BRT (~540 likes and ~116k views by the evening). It is “powered by pplx-decider-v1-27b, our multimodal decision model”, fine-tuned from Qwen3.8-27B with a 250k context window, at $0.04 per million input tokens with free output.

What shipped. POST https://api.perplexity.ai/v1/decisions with a System One-shaped body (state, questions; Noul, Choice, Score, images), a quickstart, an OpenAPI spec labelled “gateway-preview” v0.1.0, 10 requests/s per organisation, and a cookbook that triages support tickets and sends escalations to Perplexity’s Agent API. Perplexity’s own timings: a few hundred input tokens under 2 s, ~90k tokens 5 s, ~190k tokens 14 s.

Same weights as AutoJev. The weights went up at 18:24 BRT under Apache-2.0. Every file we compared is byte-identical to denis-pplx/autojev-27b: the same SHA-256 for all 11 safetensors shards, the readout and both configs. The NOTICE calls it the “AutoJev 27B curated SFT checkpoint”. AutoJev-27B is already in the catalog (#4 on the official Decision Index at 56.40), so this is a hosted release of an existing model, not a new one. Perplexity’s post does not mention AutoJev.

The numbers (UnverifiedClaim). 85.71% overall vs Jev 84.51% on an 11-benchmark panel (7,210 rows run through Perplexity’s API); Jev leads on WinoGrande, BBH, JevBench hard, TruthfulQA, Belebele and JudgeBench.

How we list it. The AutoJev page gains a Perplexity section and a hosted price. Maturity goes from 5.1 to 6.6 (hosted API with docs), adoption from 11 to 24, and capability stays at 7.2 (same weights).