Decision Index 0.3 goes live, and Perplexity's pplx-decider v1.1 takes #1 at half v1's API price
Two releases landed on 6 October 2026 and belong together.
Decision Index 0.3
The Jev Decision Index on Hugging Face published edition 0.3 (board data regenerated 7 Oct; 115 entries when we read it on 9 Oct). What changed:
- Full score = 20% public + 50% private tests of the same skills + 30% private tasks from new domains. The private parts are mapped onto the public scale using untuned stock models. “Most of the score comes from tests nobody can train on, so a public-only advantage counts for little.”
- Ties: entries within 0.9 points of the first in their group share a rank.
- Public index rebuilt (GSM8K rebuilt, ForecastBench retired), a vision board added, and superseded versions removed (pplx-decider v1 → v1.1, RSI-Jev v4.0 → v5.0, among others).
| Rank | Entry | Full score |
|---|---|---|
| 1 | pplx-decider v1.1 27B | 62.75 |
| 2 (tie) | Fastino GLiDE no-thinking (28B) | 60.21 |
| 2 (tie) | Jev | 60.11 |
| 2 (tie) | Torchcast Decision 27B (public #1, 65.10) | 59.91 |
| 6 (tie) | deck31b (frozen Gemma-4-31B) | 59.03 |
| 6 (tie) | Kev 27B | 58.78 |
| 6 (tie) | Quyet-1.0-Large | 58.71 |
| 11 (tie) | Blink v0.3 26B-A4B | 57.76 |
| 11 (tie) | Decider chat · Gemma-4-31B | 57.62 |
| 11 (tie) | Surogate Rune 26B-A4B v3 | 57.43 |
| 14 (tie) | GEV-26B-Decide | 56.17 |
Scores are not comparable with Decision Index 0.2.1. A big public-vs-full gap is the thing to watch: Torchcast Decision 27B has the best public score (65.10) but drops to 59.91 on the full score.
pplx-decider v1.1
Perplexity published pplx-decider-v1.1-27b (Apache-2.0) and announced it at 18:11 BRT (forum post). It keeps the Qwen3.8-27B backbone but has new weights: full-attention layers run noncausal, and it was trained on more data, much of it from tasksource. It is now the default model of Perplexity’s Decisions API at $0.02 per 1M input tokens (pricing), half of v1’s launch price; v1 stays selectable at the same price.
The card’s 61.56 is Perplexity’s own run of Decision Index 0.2.1 (UnverifiedClaim). The official 0.3 board scores it 62.75, #1 on its own. On JevBench v1.6.1 it has the highest open-weights Intelligence (75.7), but self-hosting a 27B model costs 6.7× Jev there, so it sits outside the Jev-class caps.
How we list it. v1 was byte-identical to AutoJev-27B and never had its own page. v1.1 is a new training run, so it gets one: pplx-decider v1.1, 6.3 / 8.1 / 16. The AutoJev page gets a note and the $0.02 price.