What is Celeris-1-decision?

A hosted diffusion language model from Celeris built for typed decisions: it takes text, structured data or images, answers yes/no, choice or score questions with a probability per outcome, and returns a one-sentence explanation on request.

How accurate does Celeris claim to be on jev-bench?

The post reports 80.0% on all 22 datasets (3,210 items) against 79.0% for pplx-decider-v1.1-27b, 75.3% for Inception Mercury Decide, 73.1% for Jev 1.13.0 and 69.0% for OpenAI gpt-6-luna.

What are the calibration numbers in this launch post?

On full jev-bench the Brier score is 0.272, which the post says is 28% below Jev's 0.380 and 40% below gpt-6-luna's 0.455; on typed decisions the Brier is 0.079 against Jev's 0.149.

How fast does Celeris say the model is?

Median end-to-end latency measured from an AWS us-east-1 client was 67 ms for a one-question request and 81 ms for fifty questions, versus 110-147 ms and 153-888 ms for the other four providers.

What does Celeris-1-decision cost?

The post says $0.04 per million input tokens with free output, about $8 per million typical short decisions; that matches Perplexity's input rate and undercuts Jev's $0.042 and gpt-6-luna's $0.10.

Can existing Jev clients switch to it?

The request format follows the systemone convention, so clients can point at the Celeris base URL and add an x_celeris explain flag to receive explanations; the release has no open weights and no fine-tuning.

What should readers watch about these claims?

They are Celeris's own vendor benchmarks run through one scorer with pinned Hugging Face datasets; the post itself notes neither provider serves from us-east-1 so latency includes cross-region networking, and typed-decisions gold labels are soft distributions.

Article

CommunityJev Media & Talks

Celeris: Celeris-1-decision claims 80% on jev-bench at 67 ms

Celeris's launch post for Celeris-1-decision, a hosted diffusion language model for typed decisions. It reports 80.0% on the full jev-bench, the lowest Brier scores of five decision APIs tested, a 67 ms median and $0.04 per million input tokens.

See highlights

Open celeris.ai

Format: Article
Reading time: 2 min
Published by: Community
Author: Tom Hamer
Last verified: Oct 2026

Tagscommunity press competition calibration

Works forJev for Decision gating Jev for Classification

Used inJev for Software & DevTools

Highlights

Celeris says its model scores 80.0% on the full jev-bench (22 datasets, 3,210 items) against 73.1% for Jev 1.13.0.
The post reports the lowest Brier score on both suites: 0.272 on full jev-bench and 0.079 on typed decisions, 28% below Jev on jev-bench.
Median end-to-end latency was 67 ms for one question and 81 ms for fifty, measured from an AWS us-east-1 client.
Pricing is $0.04 per million input tokens with output free, about $8 per million short decisions, and explanations are not billed.
The model is a hosted API with no open weights or fine-tuning, following the systemone request convention so existing Jev clients can switch base URLs.

Watch out

Copy reviewed against Celeris's own published launch post (8 October 2026) on 2026-10-17. This is a vendor-authored announcement: the accuracy, Brier, latency and price comparisons against Perplexity, Inception, Jev and OpenAI are company claims run on Celeris's own harness, and Celeris notes neither it nor Jev serves inference from us-east-1, so every latency figure includes a cross-region network path.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

The weekend builds: 838 projects and the claims behind them

Oct 16, 2026

The madewithjev gallery counted 838 builds over Jev's first 28 days. A walk through its sampled X posts: what gets built first, the costs authors report, and why every number needs its source.

Jev Media & Talks#community#use-cases#adoption

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai