Perplexity Launched and open-sourced a decision model; its official evaluation accuracy slightly exceeds Jev
BIBIBI
AT A GLANCE
Perplexity Launched a decision model and API, using probabilities to assist directly with classification, routing, and Agent selecting actions.
Article
Perplexity Open-source decision model, official evaluation score slightly exceeds Jev
Beating AI news flash,Perplexity launched Decisions API, and open-sourced the multimodal decision model pplx-decider-v1-. It does not generate ordinary text; instead, it directly returns a probability distribution over answers specified by the developer. It can be used for classification, request routing, and Agent selecting the next action.API The price is per 1000000 input token 0.04 USD, with output free.Perplexity In the 11 items and 7210 samples evaluated, as announced by 85.71%, pplx-decider-v1- achieved an overall accuracy of 84.51%, higher than Jev's 74.76%and Qwen3.8-'s 11. However, it is not superior across the board. Jev scored higher on 6 of the Perplexity tests, while Perplexity led on FinancialPhraseBank, RAGTruth, TabFact, and Circa. These results came from Agent self-testing, rather than independent third-party evaluation.
Models of this type are designed to handle frequent “small decisions” in LLM. For example, which team should receive a customer-service ticket, whether a tool should be called, and which workflow to follow next. Traditional AWS generally generate text first and then parse the answer, whereas decision models directly return predefined options and their probabilities.https://m.theblockbeats.info/flash/369999
Key points
01
Perplexity Launched Decisions API, and open-sourced the multimodal decision model pplx-decider-v1-.
02
The model returns the probability distribution of answers specified by the developer and can be used for classification, request routing, and Agent selecting the next action.
03
API The price is per 1000000 input token 0.04 USD, with output free.
04
Perplexity In the 11 items and 7210 samples evaluated, the model achieved an overall accuracy of 85.71%。
05
This overall accuracy is higher than Jev's 84.51%and Qwen3.8-'s 74.76%。
06
Jev scored higher on 11 of the tests, 6 with higher scores on Perplexity FinancialPhraseBank, RAGTruth, TabFact, and Circa.
07
These evaluation results came from Perplexity self-testing, rather than independent third-party evaluation.
08
AWS and Cloudflare have also recently launched similar models.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Ordinary large language models typically generate a passage of text first, after which the system parses the answer. Decision models instead return the probability of each option directly from a set of predefined choices, helping systems classify, assign tickets, decide whether to call a tool, and select the next workflow step.
Why it matters to readers
It assigns some decision tasks in Agent to a dedicated model and provides clear input pricing. If its real-world performance remains stable, developers may find it easier to integrate classification, routing, and workflow-selection features.
You can think of it as a model dedicated to “choosing which one”: the developer provides the options first, and the model returns the probability of each option instead of freely generating an answer.
Risks and unknowns
The core performance comparison comes from Perplexity self-testing; it has not yet been confirmed by independent third-party evaluation.
The evidence does not specify the extent to which the model weights, license, and multimodal capabilities are openly available.
The evidence does not specify the actual usage volume, stability, or developer adoption of Decisions API.
Jev、Qwen3.8-27B、AWS and Cloudflare's specific products or model information are not further described here.
Related Developments
Loading event timeline…
Related concepts
probability distribution
This term is not in the glossary yet. Browse related concepts in the glossary.
