Introducing Celeris-1.1 Decision: the fastest decision model, at about 40 ms

Celeris-1.1 Decision is the fastest hosted decision model, and smarter than TypeSafe Jev 1.14.0, Perplexity Decider v1.2, Inception Mercury Decide and OpenAI gpt-6-luna, beating all four across multiple standard benchmark datasets, such as the Artificial Analysis Decision Index, jev-bench and typed-decisions. It answers in a median of about 40 ms: 2.5× faster than Jev and 5–7× faster than the others. With this release the price drops tenfold, to $0.004 per million input tokens, and output is still free.

It is built for the decisions software makes many times a second: route this ticket, call this tool, flag this message, approve this invoice. It reads text, structured data or images, answers yes/no, multiple-choice and score questions with a probability for every option, and gives a one-sentence reason when asked. A typical decision costs less than a thousandth of a cent.

Existing integrations need one change: set "model": "celeris-1.1-decision". The endpoint, request format and response format are the same as Celeris-1 Decision.

Accuracy

Every provider received the same questions through its own decision API, and one scorer graded all of them.

On jev-bench, Celeris-1.1 Decision leads Perplexity by 0.4 points, Mercury Decide by 4.8, Jev by 7.6 and gpt-6-luna by 11.8. On typed-decisions it leads Perplexity by 0.9 points, Jev by 2.7, Mercury Decide by 5.2 and gpt-6-luna by 5.9. Its lead over Perplexity is small on both benchmarks; its lead over the other three is large on both.

On typed-decisions, the largest gain over Jev is on agent-trace observability, where Celeris-1.1 Decision scores 71.2% against Jev’s 61.6%. It also leads Jev on every question type: yes/no (84.0% vs 81.0%), choice (71.5% vs 70.3%) and score (73.9% vs 70.3%).

Calibration

A decision API returns probabilities that code thresholds and acts on, so their quality matters as much as the top answer. Typed-decisions gold labels are full probability distributions, so it measures this directly. Lower is better in every column.

Celeris-1.1 Decision’s distributions sit closest to the gold: its KL from gold is a sixth of Jev’s and an eighth of gpt-6-luna’s, and its typed-decisions Brier score is the lowest of the five. Perplexity has the lowest ECE and a jev-bench Brier score 0.003 below ours.

Latency

We measured latency from a GCP e2-standard-4 VM in us-central1-a, the region most cloud workloads run in. Each provider was measured on its own, one request at a time over a kept-alive HTTPS connection, on 150 real items from each benchmark, after three warm-up requests. Figures are end to end from the client, including the network.

Celeris-1.1 Decision answers in a median of 36 ms on jev-bench and 44 ms on typed-decisions, with a p95 within 15 ms of the median. Our server reports about 10 ms of processing per request; most of the rest is the network. From us-central1, the round trip to our endpoint is about 24 ms, against about 11 ms to the edge servers in front of the other four providers, so the gap comes from model time, not a shorter network path. For a decision made inline, such as routing a request, gating a tool call or screening a message, that time is added to every request.

Deployment

Celeris-1.1 Decision is a hosted API at the same endpoint as Celeris-1 Decision; the model field selects the version. The request format follows the systemone convention, so clients written for other decision endpoints can point at the Celeris base URL. Add "x_celeris": {"explain": true} to get a one-sentence explanation with each answer.

POST https://inference.celeris.ai/celeris-1-decision/v1/systemone

{
    "model": "celeris-1.1-decision",
    "state": { "message": "Can you cancel my order and refund the card I used?" },
    "questions": {
        "refund":  { "type": "noul", "instructions": "Asks for a refund." },
        "billing": { "type": "noul", "instructions": "Belongs with billing." }
    },
    "x_celeris": { "explain": true }
}

See Celeris Decision for the full request and response format, and pricing for billing details.

Capabilities, context and price

Celeris-1.1 Decision is multimodal: a request can carry text, structured JSON and images together, and questions can refer to each image by name. It is also the only hosted decision API of the five that explains its answers, returning a one-sentence reason for each decision when asked.

Context length

Price

Celeris-1.1 Decision costs a fifth of Perplexity and Mercury Decide, a tenth of Jev and a twenty-fifth of gpt-6-luna. Output, including explanations, is not billed. A typical one-question request with a short text input is about 200 input tokens, which comes to $0.0000008 per decision, or about $0.80 per million decisions. Images count as input tokens, and the resolution setting on each request controls how many.

Methodology

Models. celeris-1.1-decision; Perplexity pplx-decider-v1.2; TypeSafe jev-1.14.0; Inception mercury-decide; OpenAI gpt-6-luna on its Decisions API. Each was called through its provider’s own decision endpoint. The System One providers received each case as published; for OpenAI, questions were translated to its predicate, choice and score format, with option descriptions passed through.

Test sets. Both are public and pinned to a fixed Hugging Face revision.

Metrics. Accuracy counts the highest-probability option as the answer. On typed-decisions, accuracy, KL, Brier and ECE follow the dataset’s own scorer: one request per case with all five questions, compared against the gold distribution. A missing or unparseable answer scores as wrong. All accuracy runs completed with no request errors.

Latency. Measured separately from the accuracy runs, as described above, from GCP us-central1-a on November 5, 2026.

Questions people ask

What is Celeris-1.1 Decision? The second version of Celeris’s hosted decision model. It returns a probability for each typed question you ask about text, structured data or an image, and a one-sentence explanation on request.

How accurate is it? 80.2% on jev-bench and 76.2% on typed-decisions, the highest of the five hosted decision APIs we tested on both.

How fast is it? A median of 36–44 ms end to end from GCP us-central1, about 2.5× faster than Jev 1.14.0 and 5–7× faster than Mercury Decide, gpt-6-luna and Perplexity.

What does it cost? $0.004 per million input tokens, about $0.80 per million short decisions. Output tokens, including explanations, are free.

How do I upgrade from Celeris-1 Decision? Set "model": "celeris-1.1-decision". The endpoint, request format and response format are unchanged.