Perplexity's Decider v1.3 Hits Top 10 With Open-Weight Agent Decisions

Perplexity has released Decider v1.3 and Decider Flash v1.3, two open-weight models that return structured probabilities for classification, routing, scoring, and extraction. Artificial Analysis placed them at No. 6 and No. 10 on its Decision Index v2.1, with scores of 70 and 64, respectively.

Agent pipelines frequently call generative language models for tasks that produce a label, route, score, or source span. A dedicated decision model can reduce generated tokens, remove generated-text parsing, and expose confidence scores for thresholds, abstention, and escalation.

Four contracts cover common agent decisions

The full Decider v1.3 uses a 27B Qwen3 backbone and accepts multimodal input. Flash distills the system into a 4B model for lower-cost, higher-volume workloads.

The new locate task returns verbatim passages and character offsets, allowing applications to connect an extracted fact or generated claim to its source text. That contract supports citation grounding, document extraction, and evidence collection without a separate span-finding model.

Perplexity reports training sequences of up to 30,720 tokens and says its preprocessing preserves complete inputs. The hosted API advertises a 256K-token context window. Teams sending contexts beyond the training length should test accuracy and calibration across length buckets because accepted request size alone does not establish decision quality.

One marker token drives each decision

Perplexity adapts a generative language-model backbone with a linear decision head. The head reads the hidden state at a special marker token and maps that representation to the requested probabilities, scores, or spans.

The company keeps the trained adapter separate from the backbone after finding that merged weights reduced decision accuracy. This direct readout avoids autoregressive text decoding and its associated formatting failures. Calibration quality remains empirical, so production thresholds still require validation against representative labels.

Both models reach the index top 10

Artificial Analysis reports the following results on Decision Index v2.1:

• Decider v1.3: No. 6 overall with a score of 70.
• Decider Flash v1.3: No. 10 overall with a score of 64, four points above Flash v1.2.
• Component evaluations: Decider v1.3 ranks No. 4 on AA-Paperclip and No. 3 on EvalSafe-O*NET-AA, where it scores 89.
• Largest Flash gains: Perplexity reports double-digit improvements on AA-Symmetry and When2Call.

Perplexity’s internal evaluation covers a broader set of tasks. The company reports 91.6% mean accuracy across 63 query-classification tasks, 6.54 percentage points above its production classifiers, plus 88.16% mean accuracy across eight non-English languages.

Across a separate bank of 72 retrieval and reranking tasks, the larger model records 83.99 nDCG@10. This metric rewards systems that place relevant results near the top of the first 10 positions. Perplexity reports leads over Jev 1.14 of as much as 12.9 points in finance and 9.3 points in law.

Under input perturbations, Decider v1.3 changed its answer on 0.27% of questions, compared with 2.69% to 4.84% for Jev 1.14. It produced no changes across identical repeated inputs in Perplexity’s test, while Jev changed 1.88% of answers. Lower flip rates reduce inconsistent tool selection and route changes in agent systems.

The Artificial Analysis index provides the external cross-model comparison, while the classification, reranking, and stability figures come from Perplexity. Production evaluations should reproduce the relevant task mix, class balance, languages, context lengths, and failure costs.

Hosted input starts at one cent per million tokens

Perplexity lists hosted input pricing at $0.02 per million tokens for Decider v1.3 and $0.01 per million for Flash. Artificial Analysis estimates costs of $68 and $34 per million index tasks using its benchmark prompt profile.

Both models sit on the index’s capability-cost Pareto frontier, meaning no evaluated alternative delivers a higher score at a lower estimated benchmark cost. Actual spending depends on input length, retries, network overhead, and request volume. Self-hosted deployments also incur GPU, serving, monitoring, and engineering costs.

Workload determines the model

Concrete applications include agent tool selection with choice , safety escalation with noul , moderation severity with score , citation extraction with locate , and reranking for retrieval-augmented generation systems.

Production thresholds need local data

1. Choose the task contract. Match the output to the downstream action and avoid converting free-form text into labels after the call.
2. Define options precisely. Test overlapping labels, ambiguous rubrics, class imbalance, and cases where no supplied option fits.
3. Measure local calibration. Use held-out production data to set thresholds for automation, fallback models, and human review.
4. Test long and multimodal inputs. Track accuracy and confidence by token count, document type, language, and input modality.
5. Verify extracted offsets. Check locate results against the original source, including Unicode text and any normalization performed upstream.
6. Benchmark the complete path. Include network latency, batching, retries, concurrency limits, and self-hosting overhead.
7. Pin the deployed version. Re-run calibration and regression suites before changing model weights or API versions.

Open weights and two hosted routes

The 27B weights are available on Hugging Face, and Perplexity also publishes the Flash variant there. Both use the Apache 2.0 license, which permits commercial use, modification, and redistribution subject to its license and notice requirements.

Perplexity serves both models through its Decisions API, while OpenRouter offers a hosted free-tier route. Teams evaluating either service should confirm current rate limits, data-retention terms, regional availability, and free-tier quotas before production deployment.