Celeris-1 Decision

Celeris-1 Decision is a System One model from Celeris, a lab that builds diffusion language models for low latency. Tom Hamer announced it on 8 October 2026. You send text, structured data or images and a set of typed questions, and it returns a probability for each outcome. It is a hosted API with no open weights and no fine-tuning in this release.

The main endpoint is POST https://inference.celeris.ai/celeris-1-decision/v1/systemone. It implements TypeSafe’s System One API, so code written for Jev with the Jev SDK runs against it once you change the base URL, key and model name. A second endpoint, POST /v1/decisions, follows the OpenAI Decisions API format. Both serve the same model at the same price. A Noul returns the probability of yes. A Choice returns a probability per option, and a Score returns a probability per level plus the probability-weighted level. Setting x_celeris.explain adds one sentence per question saying why the model chose its answer. That makes the request slower and cuts the cap from 512 questions to 64.

Celeris’s docs use ticket routing as their example, which maps to support inbox triage and intent and model routing. Its launch post also checks an expense claim against a photo of a receipt, with up to 8 images per request. The launch post reports a median of 67 ms for one question and 81 ms for fifty, from an AWS us-east-1 client. Celeris ran those measurements itself, so treat them as vendor-run. The same post reports results from Celeris’s own decision-benching harness (commit 0cfd276), which called each provider’s endpoint and scored every answer with one scorer. On the 22 datasets and 3,210 items of jev-bench, Celeris-1 Decision scores 0.800 accuracy and 0.272 Brier, against 0.731 and 0.380 for Jev 1.13.0 and 0.790 and 0.288 for pplx-decider-v1.1-27b. A Brier score is the squared distance between the predicted probabilities and the right answer, so lower is better. These figures are vendor-run. The calibration claim rests only on Brier scores that Celeris computed itself. The post also reports typed-decisions results. That benchmark is the one the Decider 1 page describes as built by meraGPT, so this page leaves those figures out. BenchLM shows the jev-bench numbers and labels them provider-reported, with no overall rank. Benchmark Heaven’s JevBench does not list the model as of 9 October 2026.

It writes no free text. The only prose is the optional one-sentence explanation. There are no weights to self-host. The context window and the numeric rate limit are not published as of 9 October 2026. Images must be base64 data URLs, with no external links.

It is available now through a Celeris API key, with prepaid credit bought by card in the Console. TypeSafe’s typesafe-sdk for Python works with base_url set to https://inference.celeris.ai/celeris-1-decision. Credit expires 30 days after it is granted, and an empty balance returns 402.

Question types Choice Score Noul Max Choice options512 Score levelsUp to 64 Questions per call512 Total contextNot documented State budgetNot documented Rate limitLimits are per workspace and per model, and all API keys in a workspace share them. Going over returns HTTP 429 with a Retry-After header, as rate_limit_exceeded or service_busy. The numeric limit is not published as of 9 October 2026, and Celeris tells customers to contact it for a committed or higher rate. EndpointPOST https://inference.celeris.ai/celeris-1-decision/v1/systemone SDKsPython: typesafe-sdk The context window is not published as of 9 October 2026. The docs limit a request to 512 questions, or 64 when x_celeris.explain is true, 512 options per Choice, 64 levels per Score, 8 images of up to 25 megapixels each, 100,000 JSON values and an 8 MiB body. Images are base64 data URLs, count as input tokens and are read at the resolution set by x_celeris.max_soft_tokens. External image URLs are not supported. Each response is one JSON body with no streaming.

celeris-1-decision, 8 Oct 2026, The only version. The launch post is dated 8 October 2026. The model id in the request body must match the model segment of the URL path. The docs also serve it at POST /v1/decisions in the OpenAI Decisions API format, with the OpenAI Python SDK 3.26.0 or later. Release notes

What people use Celeris-1 Decision for, one page per pattern. Workflow controlStarter

Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket. Choice Score Noul Workflow controlIntermediate

One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue. Choice Score VendorCeleris TypeCommercial, hosted API StatusGenerally available Announced8 Oct 2026 AccessOpened 8 Oct 2026 Input price$0.04 per 1M tokens Output priceFree Latency67 to 81 ms (vendor claim) ContextNot published Rate limitLimits are per workspace and per model, and all API keys in a workspace share them. Going over returns HTTP 429 with a Retry-After header, as rate_limit_exceeded or service_busy. The numeric limit is not published as of 9 October 2026, and Celeris tells customers to contact it for a committed or higher rate. API docsdocs.celeris.ai

01Celeris-1-decision launch post (Tom Hamer, 8 October 2026) celeris.ai 02Decisions (Celeris docs) docs.celeris.ai 03Pricing (Celeris docs) docs.celeris.ai 04Rate limits (Celeris docs) docs.celeris.ai 05Models (Celeris docs) docs.celeris.ai 06About Celeris celeris.ai 07Celeris terms of service celeris.ai 08Celeris-1 Decision (BenchLM) benchlm.ai 09JevBench by Benchmark Heaven benchmarkheaven.com 10typesafe-sdk on PyPI pypi.org