Perplexity Decider v1.1 27B: Cost, Benchmarks and API

👁 115 reads

Decider v1.1 27B is Perplexity’s decision-making model, with open weights under the Apache-2.0 license. In the Artificial Analysis Decision Index dated October 13, 2026, it scored 53.1 points, placing 4th out of 11, but it was the cheapest model in the ranking: a million tasks cost $16.50, 2.5 times less than the leader, Jev 1.13. In my test in Polish, it classified 44 of 47 submissions correctly and had the best urgency score. I cover the category in my guide to AI decision-making models.

What is Decider?

Perplexity, known for its AI search engine, released its decision-making model on October 1, 2026 (version v1), followed a few days later by version v1.1. The model does not write text: it receives a state and questions with a fixed list of answers, then returns a selected option and the probability of each one. According to the model card, version 1.1 has the same base as v1, and the improvements mainly come from two changes:

• The causal mask was removed from the full-attention layers, so the model can “see” the entire input at once in both directions.
• The model was trained on more data, including material from the tasksource collection.

Perplexity reports an increase from 56.4 to 61.56 points on the community Decision Index on Hugging Face, more than 3.5 points above Jev. Decider ranks behind Jev in the independent Artificial Analysis ranking - these are two different test sets.

Artificial Analysis benchmark results

• Top-tier classification. Its WildGuardTest score of 93.8 is higher than Jev’s and Claude Haiku 5.5’s; only GLiDE scored better (95.1).
• Lowest-cost classification. According to AA, a million tasks from WildGuardTest alone cost $12.10 (Claude Haiku 5.5 low: $91.50).
• Second in reasoning among decision-making models, with a score of 15.4, behind Jev.
• Weak at saying “I don’t know” - its score of 13.3 means it is more likely to guess when answering factual questions.

My test in Polish (October 13, 2026)

This was the same test set used for 8 other decision-making models and 2 LLMs: 50 Polish-language customer submissions to a store, with a choice of category from 7 options and an urgency rating, plus 10 conversations in which an assistant had to choose whether to call a tool, ask a follow-up question, answer directly or refuse. Decider was accessed through OpenRouter. The displayed examples and results refer to a test conducted in Polish, not an English-language test.

• Urgency: 48 out of 50 - tied with GPT-6 Luna Decisions for the best result in the test, ahead of Claude Haiku 5.5 (46).
• Confidence as a filter: At a confidence score of at least 0.9, it classified 39 of 41 submissions correctly.
• Cost: A million tasks like these would cost about $7.80, compared with $96.70 for Claude Haiku 5.5 low - 12 times more.
• Latency: Median of 417 ms from Poland through OpenRouter (Claude Haiku 5.5: 1689 ms).

API and limits

The limit of 10 requests per second works out to about 864,000 requests per day. For higher traffic, you need to combine questions in a single request or use OpenRouter or your own server.

Run it yourself

The perplexity-ai/pplx-decider-v1.1-27b weights are available on Hugging Face under the Apache-2.0 license, which allows commercial use without a revenue threshold. According to the model card, you need a CUDA GPU with about 49 GiB of space for the weights, plus working memory - an 80 GB-class GPU or two GPUs. The included helper code for local use supports text and images for one decision within an 8192-token limit. The model will not fit on an RTX 4090 with 24 GB without quantization, which Perplexity has not released.

When should you choose Decider?

• Choose it for: High-volume, low-budget classification and moderation; businesses that want Apache-2.0 open weights and to run the model themselves; decisions involving images.
• Probably not for: Factual questions where the model should admit it does not know (Microsoft-Decision-1 or Jev may be better); very high traffic through the Perplexity API (the limit is 10 requests per second).

Frequently asked questions

How much does Perplexity Decider cost?

$0.02 per million input tokens, with free output (Perplexity and OpenRouter, as of October 13, 2026). Artificial Analysis puts the cost of a million tasks at $16.50, the lowest in the ranking.

Is Decider open source?

Its weights are open under the Apache-2.0 license on Hugging Face. Running it requires about 49 GiB of GPU memory.

Is Decider connected to Perplexity’s search engine?

Not directly. It is a separate model for making decisions in applications, available through an API. Read my guide to Perplexity for more about the search engine.

Sources

• Perplexity: Decisions API (checked October 13, 2026)
• Hugging Face: perplexity-ai/pplx-decider-v1.1-27b
• OpenRouter: perplexity/pplx-decider-v1.1-27b
• Artificial Analysis Decision Index v1.0 (October 13, 2026)
• My test: October 13, 2026, 60 tasks through OpenRouter from Poland.

// read morebasal-omni: Remek Kinas’s Model Tops JevBench’s Open ModelsBenchmarksDecision Index: Jev 1.13 Tops Benchmark for Decision AIBenchmarksMercury Decide vs Celeris-1: Speed and BenchmarksBenchmarks

Kinetyra