JevBench ranks Sage 1.3.0 first of 27 hosted systems at composite 74.0

Opened October 11, 2026, the JevBench API board ranks 27 hosted systems on release v1.6.1. The page also names board revision v1.7.44. Sage 1.3.0, from Levanto Labs, is first at composite 74.0. Jev 1.13.0 is fourth at 71.5. The row calls Sage a closed API and leaves the base undisclosed.

The composite row prints Intelligence 66, Calibration 92, Speed 94, and Cost 58, with an estimated $0.025 per 1,000 decisions. The capability card prints 78.6, from Intelligence 65.6 and Calibration 91.6, and a median of 0.15 seconds. d1 is second at 73.0. Mercury Decide is third at 72.4. Microsoft-Decision-1 is sixth at 69.1. We did not rerun the suite, and we did not call Sage.

Florian S posted those ranks on October 10 at 20:09 UTC. The text says Sage leads the API list and Jev is fourth. The four attachments are charts. The figures in the paragraph above are the page, not a transcription of the stills. The share image on this article is the first of those stills.

Levanto’s pricing page, opened the same day, lists five plans. Free and Developer charge $0.05 per million input tokens. Starter is $0.046. Pro is $0.042. Growth is $0.038, and that row names a 128K context. Every plan lists output at $10 per million tokens and web search at $0.10 each. The usage notes say output tokens are the reasoning Sage does when it thinks a question through, and that answers are free. When the included allowance runs out, calls return HTTP 402. The kinds named on every plan are Yes/No, Choice, Scale, Sort, and Tags, plus batching, reasoning, images, and web search.

The same pricing page prints a different score. It labels the model Sage v1.1 and the set JevBench v1.2 public, 213 items: Sage 200, Jev 1.13 186. A second line, multistep_decisions, is 93 against 71 of 100. Those two lines are not the composite 74.0, and they are not the capability 78.6.

On October 9, Big Iron Chris posted SageCraft. The attachment is a video of about 511 seconds. SageCraft shows Sage a screenshot and a short list of moves, then turns the pick into keyboard and mouse input. The post’s log is 32 sessions, 5 hours 29 minutes, and 4,573 decisions. The dwarf died at level 2, then reached level 5 after more than 128 kills, at 2 cents a kill. The last 18 minutes, late level 4 to level 5, ran with no manual input. The reply links levanto.ai/sagecraft and levantolabs/sagecraft. We did not open those pages. The 4,573 decisions are a play log. They are not a row on the API board.

Other hosted calls of this shape are on decision models like Jev. The October 11 board reading is also on the JevBench page.