Celeris-1 Decision
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.2/10
Paid hosted API with a docs site (decisions guide, errors, rate limits, production checklist, llms.txt), a CLI, console and status page. Jev-style systemone request, OpenAI-compatible routes. Activation waitlist, closed weights, SLA only on Enterprise contact.
- Capability7.2/10
Yes/no, choice and score on text, structured data and images, optional one-sentence explanation. AA Decision Index v1.0: 52.3 (8th of 13), fastest at 63 ms. JevBench API composite 62.0 (#17), calibration axis 88. No ECE or method; AA flags overconfidence.
- Adoption6/100
Listed on Artificial Analysis and JevBench, a launch blog (8 Oct) and Jev-SDK compatibility. No OpenRouter or Vercel listing found, no social traction data gathered, no named production users.
Vendor claims
- 80.0% on 'full jev-bench' (22 datasets, 3,210 items) against 79.0% Perplexity, 75.3% Mercury Decide, 73.1% Jev 1.13.0 and 69.0% gpt-6-luna, Brier 0.272 vs 0.380 for Jev, on Celeris's own harness. This is the community dataset Praveenrajus/jev-bench, not Benchmark Heaven's JevBench[Vendor claim — not independently verified]
- 67 ms median response for one question (81 ms for fifty) measured from AWS us-east-1, against 110–147 ms for the other providers[Vendor claim — not independently verified]
- Typed-decisions Brier 0.079, 47% below Jev's, on 400 cases x 5 questions (LocalLLaMA/typed-decisions)[Vendor claim — not independently verified]
publicado (OK do Fabricio 17/10 BRT).
Read this first. Celeris’s headline, “80.0% on jev-bench”, is not Benchmark Heaven’s JevBench. It is a community dataset (
Praveenrajus/jev-bench, 22 configs, 3,210 items) scored on Celeris’s own harness, so it is a vendor number (UnverifiedClaim). On the two independent boards Celeris-1 Decision sits below Jev 1.13, and it does lead on one thing: speed on Artificial Analysis.
Celeris-1 Decision is the decision model of Celeris, a company that serves celeris-1, a diffusion language model tuned for short structured answers. The decision product takes text, structured data or images, answers yes/no, choice and score questions with a probability for each, and can return a one-sentence explanation on request. It is a hosted API only: no open weights, no fine-tuning in this release. It was a watch item since 10 Oct; two independent boards now cover it.
Specs
| Attribute | Value |
|---|---|
| Company | Celeris |
| Launch | 8 Oct 2026 (launch blog dated the same day) |
| Model | celeris-1-decision, a diffusion language model; base and size undisclosed |
| Weights | Closed. “There are no open weights and no fine-tuning in this release” |
| Question types | yes/no (noul), choice, score; text, structured data and images in state; optional explanation via "x_celeris": {"explain": true} |
| Contract | “The request format follows the systemone convention”: clients written for other decision endpoints can point at the Celeris base URL; docs list cURL, the OpenAI SDK and the Jev SDK |
| Context | 131,072 tokens (docs.celeris.ai/models, re-read 17 Oct; “a fixed part is reserved for the answers, so the input limit is slightly lower”) |
| Price | $0.04 per million input tokens, output free (explanations included) |
| Support | Email support; Enterprise tier (SSO, SOC 2, VPC, SLA) by contacting sales |
| Access | Console with an activation waitlist |
| Docs | docs.celeris.ai (Decisions guide at /decisions) |
Independent boards
Artificial Analysis Decision Index v1.0 (read from the AA page, 17 Oct). Celeris-1 Decision scores 52.3, 8th of 13 models. d1-ultra leads at 65.2, then Jev 1.13 (57.5), Jev 1.14 (57.1), GLiDE (53.9), Microsoft-Decision-1 (53.8), Decider v1.1 27B (53.1) and Mercury Decide (52.7).
| AA metric | Celeris-1 Decision | Reference |
|---|---|---|
| Index | 52.3 | Jev 1.13: 57.5 |
| AA-Omniscience accuracy | 37.8 | d1: 41 |
| AA-Omniscience non-hallucination rate | 0.0 | Jev 1.13: 35.2 |
| HLE-MultiChoice | 13.3 | Jev 1.13: 20.9 |
| When2Call | 83.7 (tied 2nd with GLiDE) | d1-ultra 92.6 |
| WildGuardTest | 93.3 | GLiDE 95.1 |
| Time per task | 63 ms (fastest) | Jev 1.14 78 ms, d1-ultra 92 ms |
| Cost per 1M tasks | $31.57 | Jev $41.53 |
A non-hallucination rate of 0 means it never abstained on the knowledge test, the opposite of d1-ultra’s behaviour.
JevBench, hosted API board (Benchmark Heaven, board v1.7.46, checked 17 Oct): composite 62.0, #17 of 40; Intelligence 59, Calibration 88, Speed 87, Cost 48; Capability 73.6 (official capability rank 10 among Jev-class); $0.0525 per 1,000 decisions (estimate from the $0.04/M list price, output free); p50 0.43 s end to end. That is slower than AA’s 63 ms: the board times from its own host, AA from its own, and the board’s figure includes the network.
The two boards are separate and are not mixed with each other or with the Decision Index 0.3.
What Celeris claims (UnverifiedClaim)
- 80.0% accuracy, Brier 0.272 on “full jev-bench” against 73.1% / 0.380 for Jev 1.13.0, on a dataset and harness Celeris chose. Neither independent board reproduces a lead over Jev: JevBench composite 62.0 vs 71.5, AA 52.3 vs 57.5. The page does say “Jev still leads on knowledge and reasoning”.
- 67 ms median from us-east-1, “60% faster than Jev”. AA’s independent 63 ms (vs Jev 1.13 127 ms) is consistent with a large speed lead; the JevBench board’s 0.43 s is not. Report both.
- Brier 0.079 on typed decisions (a community soft-label set). No ECE and no calibration method for production use.
Fit / anti-fit
Fit when you need: the lowest per-task latency on the AA board, yes/no, choice and score on text and images, a one-sentence explanation next to each decision, and an endpoint that a Jev client can point at.
Anti-fit when you need: the best accuracy (Jev, d1-ultra, Sage and kettei all score higher on the independent boards), abstention on questions it doesn’t know (non-hallucination 0), open weights or fine-tuning, or a self-serve SLA.
Limits
- Closed weights; base model, size and training data undisclosed. No published ECE.
- AA finds every decision model overconfident overall, mainly on the knowledge and reasoning tasks.
- The 131,072-token context (docs.celeris.ai/models) and the Jev-compatible
POST /v1/systemoneroute (docs.celeris.ai/decisions) were re-read on 17 Oct; the launch-page figures are from 10 Oct. The API was not called for this page.
Why it is in the catalog
A dedicated decision model with a typed API returning per-option probabilities, public pricing and docs, and independent results on two boards. Moved from the 10 Oct watch row to a full page now that independent numbers exist.
Score working (17 Oct 2026)
- Maturity = availability 6 × 0.30 + docs 7 × 0.25 + integrations 4 × 0.25 + support 3 × 0.20 = 1.80 + 1.75 + 1.00 + 0.60 = 5.15 → 5.2
- Capability = types 9 × 0.30 + calibration 5 × 0.25 + latency 8 × 0.25 + accuracy 6 × 0.20 = 2.70 + 1.25 + 2.00 + 1.20 = 7.15 → 7.2
- Adoption = engagement 5 × 0.35 + ecosystem 12 × 0.35 + production 0 × 0.30 = 1.75 + 4.20 + 0.00 = 5.95 → 6
Availability 6: paid hosted API with an activation waitlist (as d1-ultra). Docs 7: full docs site with a decisions guide, errors, limits and a production checklist. Integrations 4: Jev-style systemone request (3) plus the CLI and OpenAI-style routes. Support 3: email support, SLA only through sales.
Types 9: three kinds, images and explanations, option caps not checked. Calibration 5: Brier figures from the vendor’s own harness and an independent calibration axis of 88 (below Jev’s 91), no ECE or method. Latency 8: AA’s independent 63 ms, no p99, tempered by the board’s 0.43 s. Accuracy 6: independent evidence exists but is mid-field (AA 8th of 13, JevBench #17 of 40) and the vendor’s lead over Jev is not reproduced; below the 8 given to boards where the model leads.