Sage 1.4 + images, beta (Levanto Labs)
Closed APIBase model: undisclosedthe operator's hosted API
Image JevBench v0.3.0 composite score
47.922
Rank #3 of 57 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 68.2
- calibration
- 83.5
- speed
- 81.0
- cost
- 43.2
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.078523
- Cost basis
- Levanto list price USD 0.05 per 1M input tokens x billed_input_tokens (text + image tokens) + USD 10 per 1M output tokens x output_tokens
- Measured latency · p50 / p95
- 0.565 s / 1.417 s
- Adjusted latency · p50 / p95
- 0.565 s / 1.417 s
- Latency adjustment
- none (hosted API, measured as is)
- Public accuracy
- 72.8%
- Sealed accuracy
- 83.3%
- Measurement note
- Official ImageJev v0.3 A300+P300 pass 1 on 16 Oct 2026 (16:28-16:39 UTC) through Levanto's Sage /decide endpoint (sage.levanto.ai/decide) with image content, which Levanto documents as beta. Model levanto-sage-v1.4 was reported on all 600 responses; deployment defaults (reasoning off, fast latency mode; every response reported fanout compute mode, none escalated). Request: the image as content with no extra text, one choice question with the item question as instructions and the same option text as the Jev-images rows (Sage-native form, fixed before any scored item was sent). 600 of 600 answered (HTTP 200): no refusals, no errors, no item was re-sent by the harness. On 80 items Sage abstained (no chosen option because its top two options were close); these are scored as the option with the highest returned probability, as /v1/systemone documents for the same backend ("never null"). If the 80 abstentions were scored wrong instead, Sage's Intelligence would be 58.6 (vs 68.2), Capability about 65 (vs 75.8) and composite 44.7 (still #3), and it would rank #5 in the capped Capability headline instead of #1. Images within Sage's limits (PNG/JPEG/WebP, at most 4 MiB, longest edge 8,192 px) were sent unchanged; 7 BMP files were re-encoded losslessly as PNG and 6 images over 4 MiB were downscaled (Lanczos, aspect ratio kept): 13 of 600 images (9 public, 4 sealed). 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.
Published 2026-10-16. Scores and ranks can change in a later release. See the method notes.