Celeris-1 Decision (Celeris)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.021674

Cost basis

Celeris list price USD 0.04 per 1M input tokens x usage.input_tokens (text + image tokens; output free)

Measured latency · p50 / p95

0.889 s / 1.530 s

Adjusted latency · p50 / p95

0.889 s / 1.530 s

Latency adjustment

none (hosted API, measured as is)

Public accuracy

57.0%

Sealed accuracy

63.2%

Measurement note

Official ImageJev v0.3 A300+P300 pass 1 on 16 Oct 2026 (20:12-20:26 UTC) through Celeris's System One API (inference.celeris.ai/celeris-1-decision/v1/systemone; Celeris routes by the model name in the URL path), model celeris-1-decision, which was reported on all 600 responses. Request: the item question as state, which names the image; the image as a data URL in x_celeris.images (Celeris's documented image form for /v1/systemone); one choice question with the same option text as the Jev-images rows; native option probabilities. Image resolution: max_soft_tokens 280, Celeris's documented default, fixed before any scored item was sent and sent explicitly on every request (about 310 image tokens per image). Celeris recommends 1120 for OCR, documents and small text; that setting was not used, and it could score higher on text-heavy items. The adapter stopped on any response naming another model or a fallback, or with too few input tokens for an image to have been read. 600 of 600 answered (HTTP 200): no refusals, no errors, no item was re-sent by the harness. Images within Celeris's documented limits (PNG/JPEG, at most 25 megapixels, request body at most 8 MiB; the server resizes to its own working resolution) were sent byte-identical. 11 of 600 images were transformed: 7 BMP files re-encoded losslessly as PNG (6 public, 1 sealed), and 4 PNG files over 5.5 MiB re-encoded to fit the body limit (2 public as JPEG quality 95, 2 sealed as smaller PNG); 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 0.89 s, 95th percentile 1.53 s; 131 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 2026-10-16. Scores and ranks can change in a later release. See the method notes.