← All evidence

Seven synthetic agent workloads: Jev vs Qwen 3.8 27B vs Needle 3

AIndependent · Reproduction artifacts availableAgent workloadsSeven synthetic workloads (author's own) · n = 4242026-09-17

Drive seven simulated application workloads — tickets, routing, driving, guardrails, approvals, scoring and home commands — validating every output before applying a simulated action.

Accuracy
Every recorded metric
ModelAccuracyRecallFPRAUROCECELatencyCost
Jev
via TypeSafe native API
176 ms / 336 ms
from browser client
$0.000025 / call
derived from the run's stated total of $0.011919 across 480 dispatched requests; the source reports totals, not per-call figures
Qwen 3.8 27B
via Cerebras, schema-constrained output
215 ms / 452 ms
from browser client
$0.00065 / call
derived from the run's stated total of $0.310581 across 476 dispatched requests
Needle 3
via local runtime on Apple M4 Pro
225 ms / 724 ms
from local Apple M4 Pro
$0 / call
run locally; the source explicitly excludes hardware and electricity
Protocol

Qwen 3.8 27B served by Cerebras with schema-constrained structured output; Jev through its native API with batched Choice/Noul questions; Needle 3 as a local tool-calling runtime accessed through the Hugging Face CLI. Every path validates its output before applying a simulated action. Fixture agreement is exact match against locally expected outputs. Qwen and Jev were measured from a browser client on 2026-09-17; Needle was measured separately on an Apple M4 Pro on 2026-09-18.

Dataset
Seven synthetic workloads (author's own)
n = 424
Tickets 100, guardrails 100, approvals 100, scoring 100, home 24, plus routing and driving trajectories that are not scored as fixtures
Limitations (8)
  • The author states plainly that this is 'not a controlled speed ranking': Qwen and Jev were measured from a browser client while Needle ran locally on different hardware, on a different day. The latency figures are not comparable across those two groups.
  • No confidence intervals. The author notes that 'repeated synthetic cases are not an independent population sample'.
  • The workloads are synthetic and authored by the benchmarker, so label quality cannot be independently checked and nothing here is drawn from a public dataset.
  • Fixture agreement is exact match against locally expected outputs. The author states it is 'not evidence of calibration or general security guarantees'.
  • There is no single accuracy for this run and we do not compute one: the five scored workloads are different tasks and averaging them would invent a number nobody measured.
  • Cost and latency are aggregate across the whole suite and cannot be attributed to any one workload.
  • Per-call costs are derived by JevBench from the run's stated totals and dispatch counts, not reported per call.
  • The repository does not disclose the author's affiliation or who funded API costs.
Note on this run
The clearest split result in the dataset. Jev leads on scoring (100% vs 93%) and trails badly on home commands (62.5% vs 100%), against the same baseline in the same suite. Cost differs by roughly 26x in Jev's favour across the whole run.

What this benchmark does not prove

  • 01That Jev is more accurate than Qwen 3.8 27B, or less. It wins two of the five scored workloads, ties two and loses one badly; there is no aggregate and we do not compute one.
  • 02That Jev is faster than Qwen. The 176 ms against 215 ms gap is inside what a browser-client measurement can shift, and the author explicitly disclaims a speed ranking.
  • 03That Needle 3 is a weak model. It ran on different hardware through a different harness and failed 117 dispatches; this measures the combination, not the model.
  • 04That the 26x cost gap transfers to your workload. It follows from this suite's token shapes, and Jev spent 8x more output tokens than Qwen here — which costs nothing only while Jev's output stays free.
Source and artifacts

typesafe-ai-benchmark — Jev vs Qwen 3.8 27B vs Needle 3 across seven synthetic agent workloads iammrduncan / Hackers in the Loop, verified 2026-09-19

Reproduction artifacts: github.com/iammrduncan/typesafe-ai-benchmark

Independent: yes · Vendor-affiliated: no · Independently reproduced: not yet

Author states 'This is an independent benchmark, not a TypeSafe implementation or parity claim' and warns the timing figures are 'not a controlled speed ranking'. Affiliation and funding are not disclosed.