Compare · reference
Jev vs d1-ultra
Reviewed November 10, 2026. Prices are USD per million input tokens unless stated otherwise — published rates, not a quote or a measurement of this host.
Jev and d1-ultra answer typed questions about supplied state. Compare the model, the access channel, and the workload separately. jevtypesafe.org is an independent Jev host, not TypeSafe AI or Liquid AI, and does not document a d1-ultra endpoint.
Capabilities depend on the model and host
Liquid open-weight d1 variants with video, audio, or extra types are different models. Official Jev long-context headers do not establish matching features on this host.
| Capability | TypeSafe Jev official | Jev on this host | Liquid d1-ultra official |
|---|---|---|---|
| Typed questions | Choice, Score, Noul; Box on 1.14 | Choice, Score, Noul; Image/Box beta, not recommended for production | Choice, Score, Noul |
| Input | Text and images | Text; images are beta | Text and images |
| Model selector | jev-1.14.0; jev-latest points to it | jev or omitted; upstream pinning not established here | d1-ultra in published API example |
| Endpoint | https://api.typesafe.ai/v1/systemone | https://jevtypesafe.org/api/v1/decide | https://api.liquid.ai/decisions/v1/systemone |
| Local weights | Not offered in reviewed official guide | No local deployment documented | Ultra classified as hosted, not open-weight download |
| Streaming or tool execution | Not established by reviewed System One contract | No support established by reviewed host contract | Not established by reviewed ultra sources |
Three price channels, not two interchangeable plans
The two official list rates share the same numeric input price. That does not establish equal request cost. Host retail rates are a separate service price.
| Channel | Input | Output | Conditions |
|---|---|---|---|
| TypeSafe official Jev 1.14 or 1.13 | $0.042/M | Free | Official model guide; separate official account and key |
| jevtypesafe.org hosted Jev | $0.175–$0.42/M | Free under host metering | Prepaid retail packs; $5 minimum listed; rates depend on pack size |
| Liquid official d1-ultra | $0.042/M; cached input also $0.042/M | $0.00/M | Current pricing page; each question billed as its own prompt |
Read the host pricing page before buying capacity.
Multiple questions change the calculation
Liquid states that each d1-ultra question is billed as its own prompt, including its text and all images. A shared state sent once over HTTP can therefore be counted in more than one billed prompt.
Illustrative text-only math: 1,000 billed state tokens and three 100-token questions → 3 × (1,000 + 100) = 3,300 input tokens → $0.0001386 at $0.042/M. Assumed counts, not a live receipt.
TypeSafe documents ingesting state once for parallel questions. For an exact Jev bill, use returned usage rather than assuming identical tokenization. On this host, credit multiplier is 1 and balance follows usage.billed_tokens. Do not use openai-decisions 3× multiplier as a Jev or d1-ultra price.
Published latency is not a head-to-head result
There is no common workload, region, concurrency, percentile, and error policy across these rows. Do not sort them into a speed ranking.
| Evidence | Reported latency | What it does and does not show |
|---|---|---|
| Sardine, November 2 | Around 100 ms per Jev call, under 1,000 input tokens | A particular fraud workflow, not this host or a D1 comparison |
| arXiv v2, Jev 1.13 | Mean 0.36 s client-side, including network, at 32 concurrent requests | Published experiment on 37 datasets, not Jev 1.14 or this host |
| Encima, October 8 | Jev mean 0.41 s; Decisions API mean 0.52 s | Separate mixed-task comparison, not a d1-ultra test |
| Liquid d1-ultra release and reviewed docs | No comparable numeric latency figure established | Single forward pass is an architecture claim, not end-to-end p95 latency |
Vendor evaluations and your evaluation
Liquid's October 15 release reports 112 unique evaluations against d1 and Jev 1.14 and claims better accuracy and calibration — a vendor claim, not an independent benchmark by this site.
Replay a fixed, permitted dataset against the versions you intend to use. Record templates, input sizes, number of questions, image dimensions, usage, errors, concurrency, and client region. Measure task quality and calibration alongside p50 and p95 latency.
Use the judgment-and-code method to keep policy outside either model. Explore Jev patterns in the playground — not a D1 trial or model-switching UI.
Sources
Referencia de origen:
This is an unofficial community resource. We are not affiliated with TypeSafe AI or Liquid AI.