Orqestra Enterprise Benchmark · October 2026
Enterprise benchmark16 October 20262,034 graded tasks12 systems7 dimensions

Same tasks. Same grading. List prices for everyone.

Orqestra measured against the AI gateways enterprises use today (OpenRouter Auto, Portkey and LiteLLM) and against frontier models on their own, across accuracy, cost, routing, adaptive learning, caching, workflow creation, content generation and latency.

85.8%

accuracy, Orqestra (quality). OpenRouter Auto 82.2%; Portkey 78.2%.

−38%

cost per correct answer vs GPT-5.5, at level accuracy.

−40%

cost per correct answer as it learns, from first to last tasks, in each of 3 task orders.

9/16

new workflow requests built and run. Claude Sonnet 5.5 7; GPT-5.5 4.

Fig. 3.1

Accuracy against cost per 1,000 correct answers.

Up and to the left is better. Cost axis is logarithmic. Hover a point.

{{ t.t }}
 {{ t.t }}
{{ p.label }}
← CHEAPERUSD PER 1,000 CORRECT ANSWERS · LIST PRICES
{{ s.k }}

{{ s.t }}

{{ s.d }}

{{ tb.cap }}
{{ c }}
{{ c.t }}
{{ tb.note }}
FAIRNESS

Every system received the same tasks, prompts and token caps, graded by the same code. Costs are list prices for the tokens used.

TEST DATA

Public sets sampled from test splits. Cache workloads and new workflow requests frozen with a hash and timestamp before the measured versions were finished.

STATISTICS

Differences paired on identical tasks with bootstrap 95% intervals. A lead is reported only when the interval excludes zero. Routing replicated in three task orders.

DATASETS

Banking77, LEDGAR and CaseHOLD (LexGLUE), MASSIVE, FinQA; enterprise sets generated with known answers. Harness and raw results in the Orqestra repository.