Benchmarks · Published results, path stated per row

Published benchmarks, and how they were measured.

Matched or beat the strongest frontier baseline on all five published benchmarks, with 30–140× lower serving cost. This page is the record behind that line: for each benchmark, the dataset, the metric, the frontier model actually run on the same items, the cost basis on both sides, and what the numbers do not show.

Path stated per row · frontier models run on the same items · quality normalized to the strongest frontier baseline

The five published benchmarks

The path tested is stated per row: the intent row ran the specialist directly from the benchmark harness; embedding and reranking called the direct endpoints; ICD-10 and invoice ran the product path through auto via the eval API. Quality is shown relative to the strongest frontier baseline run on that benchmark (= 100); the underlying metric is named per row. The cost ratio is that baseline's measured cost divided by Pareta's for the same items.

Task Dataset Metric Pareta scorefrontier = 100 Frontier compared Cost ratio Path tested Date / version
Intent classification Banking77 test split macro-F1 113 Claude Opus 4.7zero-shot · adaptive thinking 88× lower Specialist-path (benchmark harness) May 2026classification sweep v1
Text embedding CUAD retrieval queries nDCG@10 109 OpenAI text-embedding-3-large 32× lower Direct endpoint (/v1/embeddings) 2026-07-10
ICD-10 coding De-identified inpatient discharge summaries from a credentialed clinical dataset, ICD-10-CM micro-F1 106 GPT-5.5 138× lower Product path (auto, eval API) 2026-07-06
Invoice extraction katanaml real invoices field-level F1 103 Claude Opus 4.7vision input 142× lower Product path (auto, eval API) July 2026
Document reranking CUAD retrieval queries nDCG@10 101 Gemini 3.5 Flashlistwise LLM reranker 31× lower Direct endpoint (/v1/rerank) 2026-07-09

How to read the table

One convention for every row, so the five can sit on one axis without averaging unlike metrics.

NormalizationThe strongest zero-shot frontier baseline run on that benchmark is set to 100. The Pareta score is Pareta's raw metric divided by that baseline's raw metric, multiplied by 100 and rounded to the nearest whole number. 100 is parity; 113 means Pareta's macro-F1 was 13% higher than the baseline's on the same 3,080 items.
The raw metric is named per rowScores are never averaged across tasks. Each row's metric is the one the task's grading contract scores — macro-F1, nDCG@10, micro-F1 or field-level F1 — and the same scoring function is applied to every contender on that benchmark.
Cost ratioThe frontier baseline's measured cost for the items divided by Pareta's, rounded to the nearest whole multiple. "88× lower" means Pareta's serving cost was one eighty-eighth of the baseline's. Across the five rows the measured ratios run from 31× to 142×; the published line rounds this to 30–140×.
Matched or beatAll five Pareta scores are at or above 100 on the point estimate. On the smaller sets the confidence intervals of the Pareta score and the frontier score overlap; read those rows as parity on quality, with the cost difference as the robust finding.

Methodology

What was run, against what, and how each number on the table was produced.

01
Path tested: per row
The path tested is stated per row. Intent classification ran the specialist directly from the benchmark harness, on the same serving stack and configuration auto uses for that task in production, without end-to-end routing, verification, or frontier escalation. Text embedding and document reranking called the direct endpoints /v1/embeddings and /v1/rerank, which are the production path for those workloads. ICD-10 coding and invoice extraction ran the product path — auto through the eval API — including routing and verification. None of the rows include route-mix or escalation statistics, so no row measures the blend of specialist and frontier serving that a live workload sees.
02
Frontier baselines actually run
The frontier models were run on the same items, through their vendor APIs, and scored with the same grading contract. No frontier score on this page is taken from a vendor's published leaderboard. Each row names the baseline the comparison is made against; the strongest baseline per row is GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash or OpenAI text-embedding-3-large; Claude Opus 4.8 and Claude Sonnet 4.6 were also run and scored below the row's baseline.
03
Frontier cost
The baseline's actual usage tokens on those items — input and output, including reasoning tokens where the vendor bills them — priced at the vendor's list price at the time of the run. On live Pareta requests the X-Pareta-Frontier-Would-Have-Cost header uses the same tokens × list-price arithmetic but is an estimate, because the frontier model is not run; in these benchmarks it was.
04
Pareta cost
Actual metered billing for the same calls: the specialist token rates for tasks served through auto, and the per-unit rates for /v1/embeddings and /v1/rerank, as listed on the pricing page. Where a run predates the published token rates, the recorded cost derivation for that run is used instead; every figure carries a recorded derivation.
05
Configuration
Specialists run with the configuration auto applies to the task in production: the task's prompt and output contract, constrained decoding where the contract defines it (response_format with json_schema), and, for classification, the retrieved few-shot examples the serving path injects. Frontier baselines run zero-shot with the same label schema or output contract through the vendor API. For the intent row the benchmark harness's own 8-shot retrieval was used; the production path retrieves shots with Pareta's own embedder and scored 93.6 macro-F1 on the same 3,080 items in August 2026.
06
Decoding
Pareta specialists run at temperature 0. Frontier baselines run at temperature 0 except where the vendor API requires otherwise: the Claude Opus 4.7 legs run with adaptive thinking, which the Anthropic API only permits at temperature 1. Embedding and reranking involve no sampling: the embedder returns a vector, and the reranker returns a relevance score for each (query, document) pair.
07
Scoring, by metric
Every contender on a benchmark is scored by the same function against the same gold. The functions are the task grading contracts the evaluation API applies to customer data.
MetricRowsHow it is scored
macro-F1Intent classificationOne label per item, compared exactly to the gold label; an output that is not one of the labels counts as wrong. F1 is computed per class across the 77 intents and averaged without weighting, so rare intents count as much as common ones.
nDCG@10Text embedding · Document rerankingEach query's candidate pool is ranked — by cosine similarity of the embeddings, or by the reranker's relevance score — and scored against graded human relevance judgments derived from CUAD's expert clause annotations: 2 for a chunk that contains at least 90% of the annotated clause span, 1 for a substantial partial overlap, 0 otherwise. nDCG@10 is averaged over queries.
micro-F1ICD-10 codingThe predicted set of ICD-10-CM codes for each note is compared to the gold set. True positives, false positives and false negatives are pooled across all 150 notes before F1 is computed, so frequent codes weigh more than rare ones. Codes are compared without the decimal point.
Field-level F1Invoice extractionEach header field is compared to gold after normalization. Line items are matched to gold rows without regard to order and then scored per sub-field; missing and hallucinated rows count against the score.

What the benchmarks do not show

These results answer one question — how the specialist performs on the task — and leave others open.

Route mix and escalation rateThese runs do not measure how often auto serves a request with a frontier model on a live workload, or the blended cost that results. Not yet published. Every response does carry the exact amount billed (X-Pareta-Billed) and an estimated frontier comparison (X-Pareta-Frontier-Would-Have-Cost); per-request serving path, escalation, and the serving revision ride as X-Pareta-Served-By, X-Pareta-Escalated, and X-Pareta-Route-Version.
End-to-end latencyNot measured here, and per-workload p50/p95 latency is published on the production page. What can be stated today: per-request timeouts apply (45 s default; 120 s ICD-10 coding; 180 s long-document extraction) — past the specialist deadline the request is escalated to a frontier model, and a request no path can serve returns an explicit error — never a silent empty answer — and is not billed. A cold specialist can add up to 60 s of boot time before that deadline applies.
Verification and escalationThe quality gate auto applies in production — result checks and frontier escalation when a specialist misses the bar — is exercised only by the two product-path rows (ICD-10 coding, invoice extraction); the harness and direct-endpoint rows do not exercise it, and no row reports route mix or escalation rate.
Your distributionThe datasets are public benchmarks; CUAD in particular is public and may appear in frontier models' training data. The only number that settles a workload is one measured on its own production distribution, which is what the evaluation API below is for.

Reproduce it on your data

The grading contracts that scored these benchmarks are the ones the evaluation API applies to your data. Upload representative items from the workload you plan to deploy — an evaluation set is 5–50 items — name the frontier models to compare, and Pareta scores every contender, auto and each frontier baseline, with the same scorer on the same items, before you move live traffic.

01
Upload
5–50 items with their expected outputs, under the output contract you will use in production (json_schema for structured tasks).
02
Compare
Pareta runs auto and the frontier baselines you name on the same items. Evaluation calls are metered like any other request: specialist rates where a Pareta specialist answers, the provider's list price for each frontier call.
03
Read
Per contender: quality with a confidence interval and measured cost — the same two columns as the table above, on your distribution.
$30 in credit · no card required · OpenAI-compatible

Change history

Changes to the published set are recorded here.