| --- |
| pretty_name: PWS-Decisions Benchmark |
| license: other |
| license_name: pws-decisions-evaluation-no-redistribution |
| license_link: LICENSE |
| language: |
| - en |
| tags: |
| - benchmark |
| - decision-models |
| - decision-api |
| - web-grounding |
| - agent-evaluation |
| - multiple-choice |
| - leaderboard |
| - synthetic |
| size_categories: |
| - n<1K |
| task_categories: |
| - text-classification |
| extra_gated_prompt: "By clicking Agree, you accept the PWS-Decisions Dataset Evaluation Terms in LICENSE, including no redistribution, and agree to share your Hugging Face username and email with Parallel for dataset access records. Access is granted immediately." |
| extra_gated_heading: "Access PWS-Decisions" |
| extra_gated_button_content: "Agree and access dataset" |
| --- |
| # PWS-Decisions Benchmark |
|
|
| **A web-grounded decision model benchmark for AI agents.** |
|
|
| Agents routinely need to make small decisions with real consequences: whether a product supports an integration, a release fixes a vulnerability, a service is available in a region, or an organization meets a qualification rule. The answer often depends on a specific fact on the web—and on when that fact was true. |
|
|
| PWS-Decisions evaluates decision APIs on these externally grounded questions. The goal is to give developers and agents a common, reusable benchmark for choosing a decision API for their workflows. |
|
|
| ## What the benchmark measures |
|
|
| Each question describes a situation and asks for one answer from a fixed set of choices. The benchmark emphasizes dated facts, product requirements, and exceptions that broad familiarity with a topic can miss. |
|
|
| The full challenge set contains **300 questions**. This release provides **100 questions with reference answers** for reproducible evaluation; the remaining 200 stay private. Scoring is exact-match accuracy, with no LLM judge required. The set was curated for difficulty rather than to estimate average traffic accuracy. |
|
|
| ## Decision model leaderboard |
|
|
| Mean accuracy ± sample standard deviation across three runs. Whole-set accuracy includes all 300 questions, including the public 100. Evaluated October 29, 2026 (Pacific). |
|
|
| | Decision API | Public Set Accuracy (100) | Whole Set Accuracy (300) | |
| |---|---:|---:| |
| | **Parallel Decisions API** | **78.33% ± 0.58** | **76.22% ± 0.19** | |
| | Perplexity Decisions | 51.00% ± 0.00 | 50.89% ± 0.38 | |
| | Baseten Mammoth | 37.00% ± 0.00 | 45.00% ± 0.00 | |
| | OpenAI Decisions API | 42.00% ± 0.00 | 41.67% ± 0.00 | |
| | TypeSafe Jev 1.14 | 33.33% ± 0.58 | 33.44% ± 0.19 | |
|
|
| Uniform random guessing scores **44.67% on the public set** and **42.42% on the whole set** because questions have different numbers of choices. See [evaluation conventions](docs/methodology.md). |
|
|
| ## Latency and price for web-grounded decisions |
|
|
| Compare Parallel Decisions API with OpenAI Luna and Anthropic Haiku research harnesses using native web search. These are separate configurations from the decision API leaderboard above. Measurements cover complete answers across three runs of the whole set. |
|
|
| | Configuration | Evaluation date (Pacific) | p50 / p95 latency | USD / 1,000 questions | |
| |---|---|---:|---:| |
| | **Parallel Decisions API** | **2026-10-29** | **0.56 / 0.71 s** | **$1.00** | |
| | OpenAI Luna + native web search | 2026-10-29 | 7.42 / 19.17 s | $16.52 | |
| | Anthropic Haiku 5.5 + native web search | 2026-10-29 | 6.38 / 10.95 s | $27.43–$28.79 | |
|
|
| The search harnesses scored 97.33% (Luna) and 96.56% (Haiku) on the whole set. |
|
|
| ## Use PWS-Decisions in your evaluation |
|
|
| Use the released questions to compare decision APIs, evaluate web-grounded classification, or measure the accuracy, latency, and cost of an agent's decision step. Keep the question wording, choices, and date constraints intact so results remain comparable. |
|
|
| Sign in to Hugging Face and accept the access form to download immediately. Parallel receives your account username and email; no manual approval is required. The dataset includes the questions, reference answers, and baseline predictions. The companion GitHub repository provides the runner, API examples, and detailed results. [GitHub repository](https://github.com/parallel-web/pws-decisions-benchmark). |
|
|
| ## Related decision model benchmarks |
|
|
| PWS-Decisions adds an externally grounded evaluation to the wider ecosystem of decision model benchmarks and leaderboards, including [JevBench](https://github.com/fstandhartinger/jevbench), [LeanMCP's typed-decision benchmark](https://leanmcp.github.io/jevbench/), and the [Artificial Analysis Decision Index](https://artificialanalysis.ai/methodology/decision-index). Developers comparing TypeSafe Jev, Parallel Decisions API, OpenAI Decisions API, Perplexity Decisions, and Baseten Mammoth can use these complementary evaluations to assess the tasks that matter to their agents. |
|
|
| ## Dataset terms |
|
|
| The dataset is available for benchmarking and evaluation under the [PWS-Decisions Dataset Evaluation Terms](LICENSE). You may publish aggregate results and send question inputs to the APIs you evaluate. Do not redistribute questions, reference answers, or per-question baseline predictions; direct other users to this dataset page for their own access. The companion runner and examples are MIT-licensed. |
|
|
