Celeris opens celeris-1-decision, a hosted typed-decision API at $0.04/M

What did Celeris ship, and when?

The dated announcement is the 8 October 2026 product page signed Tom Hamer. It calls celeris-1-decision “a diffusion language model built for typed decisions.” A request sends text, structured data or images plus named questions; the reply is a probability for each allowed outcome and, if asked, a one-sentence explanation. The “Smartest” and “60% faster than Jev” wording is Celeris’s headline, not an independent ranking.

Docs list three hosted models: `celeris-1` for general work, `celeris-1-magnus` for slower, more capable answers, and `celeris-1-decision` for yes/no, multiple-choice and scoring. The decision model is a separate base URL. The models page says it “does not serve the chat, Responses, or models endpoints.”

Typed decisions are already a product category this week. OpenAI’s public beta Decisions API uses gpt-6-luna on POST /v1/decisions. Liquid AI’s open d1 weights are downloadable classifiers, not this host. Celeris is a third vendor in that shape: hosted, no weights, two wire formats.

How do the two endpoints work?

The API reference names POST /v1/decisions, using the OpenAI Decisions API format with an `input` and an ordered `questions` list, and POST /v1/systemone, using the System One convention with a `state` object and named typed questions. Both exist only under `https://inference.celeris.ai/celeris-1-decision/v1`. The model field in the body must match the path segment.

The launch-page example is a receipt image plus a claim JSON, with three `noul` (yes/no) questions: whether the total matches, whether it is travel, and whether the receipt looks altered. The printed response, which the page says is “verbatim from the live endpoint on Oct 8, 2026,” returns 0.94 / 0.96 / 0.04 and three one-sentence explanations. That example is Celeris’s illustration, not a test we ran.

A probability is still a model estimate. Callers who threshold it for payouts or routing need their own labels, the same way any structured output path needs a schema and a fallback. The launch page says explanations can be checked by a person, thresholded by code and audited later. That is a product claim about the response shape, not a compliance certification.

What do the vendor benches show?

Celeris says all five providers received the same questions through their own decision endpoints and were scored by one scorer. The harness note on the page is “decision-benching harness 0cfd276 · one scorer for all providers · October 2026.” gpt-6-luna questions were translated into OpenAI’s predicate, choice and score format. Datasets are pinned to Hugging Face revisions: Praveenrajus/jev-bench (22 configs, 3,210 items) and LocalLLaMA/typed-decisions (400 cases × 5 questions).

On full jev-bench the page lists celeris-1-decision at 0.800 accuracy and Brier 0.272, Perplexity pplx-decider-v1.1-27b at 0.790 / 0.288, Inception Mercury Decide at 0.753 / 0.372, Jev 1.13.0 at 0.731 / 0.380 and OpenAI gpt-6-luna at 0.690 / 0.455. On typed decisions it lists 0.765 / 0.079 against Perplexity 0.738 / 0.173. Lower Brier is better. The page says it wins 17 of 22 jev-bench datasets against Jev and gpt-6-luna.

Latency was measured from an AWS us-east-1 c7i.large. Median end-to-end is 67 ms for one question and 81 ms for fifty. The other four providers are listed between 110 ms and 147 ms for one question. Celeris notes that neither it nor Jev serves inference from us-east-1, so every figure includes a cross-region hop. We did not repeat the client.

How do you access it, and what does it cost?

Authentication is a bearer key (`ck_...`). The models page says the global base URL picks a region; a US East regional host is listed for `celeris-1`, not specifically for the decision model. Requests must keep the path model and the body model in agreement or they 404 / fail validation.

Pricing docs charge `celeris-1-decision` US$0.04 per million input tokens and the same for cached input, with output free. The launch page translates a typical 200-token one-question request to about $0.000008, or about $8 per million short decisions, and the 818-token receipt example to about $0.00003. Those arithmetic lines are Celeris’s. Credits are prepaid in the console, expire 30 days after they are granted, and return `402 insufficient_quota` at zero. We did not open a billing account.

What is not established?

We did not send a request, reproduce jev-bench, or compare calibration on a private label set. The 80.0% and 67 ms figures are Celeris’s, from one harness and one us-east-1 client. The “60% faster than Jev” headline is not the same number as the 67 ms versus 135 ms pair in the table; readers should use the table, not the adjective.

There are no open weights and no fine-tune in this release. A System One-compatible client is not proof that every Jev or Mercury caller will accept Celeris explanations without code changes. Image input in the example is an inline data URL, not a hosted file id.

This is an evidence review of pages opened on 9 October. It is not a first-hand latency test and not a recommendation to replace an existing Decisions or d1 path.