Drex Max (Nace AI)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.083663

Cost basis

Nace AI list price USD 0.06 per 1M input tokens x usage.input_tokens (text + image tokens; output not billed)

Measured latency · p50 / p95

0.851 s / 1.842 s

Adjusted latency · p50 / p95

0.851 s / 1.842 s

Latency adjustment

none (hosted API, measured as is)

Public accuracy

74.5%

Sealed accuracy

81.2%

Measurement note

Official ImageJev v0.3 A300+P300 pass 1 on 21 Oct 2026 (21:03-21:18 UTC) through Nace AI's hosted Drex API (drex.nace.ai/v1/systemone), model drex-max, which was reported on all 600 responses (the API documents that a drex-max request is never handed to another model). Request: the item question as state, with the state naming the image by its documented reference "Image 1"; the image as a data URL in Drex's top-level images array (the documented image form for /v1/systemone, inline bytes); one choice question with the same option text and instruction as the Jev-images rows; native option probabilities. Drex has no image-resolution setting; it resizes images to at most 1,638,400 pixels itself. The adapter stopped on any response naming another model, or reporting no image tokens (an image that was not read). 600 of 600 answered (HTTP 200): no refusals, no errors, no item was re-sent by the harness. Images within Drex's documented limits (PNG/JPEG/WebP, at most 16 MiB and 40 megapixels per image, request body at most 4 MiB) were sent byte-identical unless the file exceeded 2.75 MiB, which we kept as our own margin for the body limit. 18 of 600 images were transformed (12 public, 6 sealed): 7 BMP files re-encoded losslessly as PNG (container not accepted by Drex; 6 public, 1 sealed), and 11 PNG files over 2.75 MiB re-encoded as JPEG quality 95 at their original pixel size (6 public, 5 sealed); no image was resized. Other systems received images under their own provider limits, so image treatment differs between rows. Latency is the end-to-end request time from Sandy, one request at a time: median 0.85 s, 95th percentile 1.78 s; 158 of 600 requests took longer than 1.233 s. Exposure was logged before dispatch. Measured for a paid fast-lane evaluation order; published under the order's public-visibility setting.

Published 2026-10-22. Scores and ranks can change in a later release. See the method notes.