kettei-fat (Sqwish)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.010929

Cost basis

Sqwish list price USD 0.06 per 1M input tokens x usage.input_tokens (text + 64 per image; no output tokens)

Measured latency · p50 / p95

0.418 s / 0.863 s

Adjusted latency · p50 / p95

0.418 s / 0.863 s

Latency adjustment

none (hosted API, measured as is)

Public accuracy

73.6%

Sealed accuracy

82.1%

Measurement note

Official ImageJev v0.3 A300+P300 pass 1 on 16 Oct 2026 (17:24-17:34 UTC) through Sqwish's OneHot API (api.sqwish.ai/v1/systemone), model kettei-fat (the largest kettei size; Sqwish docs: kettei-fat v1.0, released 14 Oct 2026). kettei-fat was reported on all 600 responses (no version field returned). Request: the item question and image as state (the image as a Sqwish image asset, TypeSafe-compatible form that Sqwish documents for /v1/systemone), one choice question with the same option text as the Jev-images rows, native option probabilities. /v1/systemone never falls back to another model (the named model answers or the request fails); the adapter also stopped on any response naming another model or a fallback, or reporting an unread image. 600 of 600 answered (HTTP 200): no refusals, no errors, no item was re-sent by the harness. Images within Sqwish's documented limits (JPEG/PNG/WebP still images, at most 20 MiB, 64 megapixels, aspect ratio 200:1; the server resizes to 65,536-2,097,152 pixels itself) were sent byte-identical; 7 BMP files were re-encoded losslessly as PNG (6 public, 1 sealed), so 7 of 600 images were transformed; no image was resized client-side. 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. Exposure was logged before dispatch. Measured for a paid fast-lane evaluation order; published under the order's public-visibility setting. Cost counts the 64 tokens Sqwish bills per image; on the about 1,400 tokens the model reads per item (text plus about 1,310 image tokens) the composite would be 45.4 (kettei-fat, #5) and 50.2 (kettei-chub, #3); Capability does not depend on cost. Sqwish's pricing page says you pay for what OneHot reads, while its images page bills each image as 64 input tokens; cost follows the images page, and the order's key was charged USD 0, so no real charge confirms either reading.

Published 2026-10-16. Scores and ranks can change in a later release. See the method notes.