Skip to content
System One

Compare decision models

A decision model is not the only way to get a label out of text. These pages put Jev beside the alternatives one at a time: other decision models, zero-shot and fine-tuned classifiers, named-entity recognition models, and language models asked for structured output. A few put two other decision models side by side. Four set Clairic, an extraction model, beside Jev, Clef, GLiNER and LLM structured outputs, to show where the second System One class fits. Each one names what the two things actually return, what they cost, and where the comparison stops being fair.

Every decision model at a glance (34)

Built from each model's page, so the numbers here are the numbers there, each with its source. A field a vendor has not published is dated to the day the page was last checked.

Commercial models

Commercial decision models
ModelVendorStatusAccessQuestion typesInput priceContextLatencyLicence
CiceroButler ComputeGenerally availableHosted APIChoice, Score, Noul$0.01/MTok131,072 tokens41 to 41 ms1Apache-2.0
elephant-1BasetenGenerally availableHosted APIChoice, Score, NoulNot published as of 23 Oct 202616,384 tokensNot published as of 23 Oct 2026MIT
mammoth-1BasetenGenerally availableHosted APIChoice, Score, NoulNot published as of 23 Oct 202616,384 tokensNot published as of 23 Oct 2026MIT
d1-ultraLiquid AIGenerally availableHosted APIChoice, Score, Noul$0.042/MTok32,768 tokensNot published as of 23 Oct 2026Not published as of 23 Oct 2026
Span-02RespanGenerally availableHosted APIChoice, Score, Noul$0.01/MTok131,072 tokens33 to 822 ms5Apache-2.0
SubQ SnapSubquadraticEarly accessHosted APIChoice, Score, Noul$0.04/MTok393,216 tokensNot published as of 23 Oct 2026Not published as of 23 Oct 2026
Clef-omniCloudflareGenerally availableHosted APIChoice, Score, Noul$0.15/MTok64,000 tokens130 to 150 ms7Apache-2.0
Microsoft-Decision-1MicrosoftGenerally availableHosted APIChoice, Score, Noul$0.042/MTok32,768 tokens85 to 125 ms8Not published as of 23 Oct 2026
Celeris-1 DecisionCelerisGenerally availableHosted APIChoice, Score, Noul$0.04/MTokNot published as of 23 Oct 202667 to 81 ms9Not published as of 23 Oct 2026
ClefCloudflareGenerally availableHosted APIChoice, Score, Noul$0.24/MTok65,536 tokens38.8 to 209.3 ms10Apache-2.0
pplx-deciderPerplexityGenerally availableHosted APIChoice, Score, Noul$0.02/MTok262,144 tokensNot published as of 23 Oct 2026Apache-2.0
GLiDEFastino LabsGenerally availableHosted APIChoice, Score, Noul$0.04/MTok65,536 tokens360 to 516 ms12Fastino Community License
Mercury DecideInceptionGenerally availableHosted APIChoice, Score, Noul$0.02/MTok65,536 tokensNot published as of 23 Oct 2026Not published as of 23 Oct 2026
d1Liquid AIGenerally availableHosted APIChoice, Score, Noul$0.02/MTok65,536 tokensNot published as of 23 Oct 2026LFM Open License v1.0 (open d1-3B and d1-omni-600M; the hosted d1 has no published licence)
OpenAI Decisions APIOpenAIEarly accessHosted APIChoice, Score, Noul$0.1/MTokNot published as of 23 Oct 2026Not published as of 23 Oct 2026Not published as of 23 Oct 2026
KaiHanzo AIGenerally availableHosted APIChoice, Score, Noul$0.021/MTok1,024 tokensNot published as of 12 Oct 2026Not published as of 12 Oct 2026
DrexNace.AIGenerally availableHosted APIChoice, Score, Noul$0.06/MTok262,144 tokensNot published as of 23 Oct 2026Nace.AI Open RAIL-M
Span-01RespanGenerally availableHosted APINoul$0.02/MTokNot published as of 23 Oct 2026Not published as of 23 Oct 2026Not published as of 23 Oct 2026
Tev1Together AIEarly accessHosted APIChoice$0.042/MTok2,048 tokensNot published as of 23 Oct 2026Not published as of 23 Oct 2026
Decider 1meraGPTGenerally availableHosted APIChoice, Score, Noul$0.03/MTok4,096 tokensNot published as of 12 Oct 2026Not published as of 12 Oct 2026
Solar DecideUpstageEarly accessHosted APIChoice, Score, Noul$0.1/MTok524,288 tokensNot published as of 9 Oct 2026Not published as of 9 Oct 2026
JevTypeSafe AIGenerally availableHosted APIChoice, Score, Noul$0.042/MTok64,000 tokens70 to 500 ms22Not published as of 23 Oct 2026
  1. 1 Cicero: Vendor-run: Butler's model card and models page give a 41 ms median.
  2. 5 Span-02: The Jev Decision Index ran the weights on one RTX PRO 6000 with Respan's own engine: median 33 ms, p95 822 ms, mean 234 ms.
  3. 7 Clef-omni: Vendor-run, from Cloudflare's launch post: about 130 ms at the median for text-only decisions and about 150 ms with images.
  4. 8 Microsoft-Decision-1: Vendor-run.
  5. 9 Celeris-1 Decision: Vendor-run, from the launch post: a median of 67 ms for a one-question request and 81 ms for a fifty-question request, measured end to end from an AWS us-east-1 client.
  6. 10 Clef: Vendor-run, from Cloudflare's launch post, across the 43 evals it ran: median latency 209.3ms for Clef and 38.8ms for Clef-flash, against 524.1ms for Jev, 51.4ms for Kev-9B and 5.8ms for Laya.
  7. 12 GLiDE: Benchmark Heaven measured the hosted API on 2026-10-22 from Germany at a median of 0.360 s and a p95 of 0.516 s per decision.
  8. 22 Jev: 70ms to 500ms end-to-end is TypeSafe's own figure from the launch post, as is the 3 to 329 seconds it compares frontier LLMs against and the resulting 40x to 200x multiple.

Open-weights models you run yourself (listed, not ranked)

Open-weights decision models
ModelVendorStatusAccessQuestion typesInput priceContextLatencyLicence
H2O-Lightning-31BH2O.aiGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API131,072 tokens46 to 57 ms1H2O Model Non-Commercial Licence v1.1
Jeffan-31b-1710Trong VuGenerally availableOpen weights, self-hostedChoiceNo hosted API8,192 tokens43 to 121 ms2Apache-2.0
Quyet-1.0-LargeQuyetGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API8,000 tokens105.4 to 115.8 ms3Apache-2.0
Torchcast Decision 27BTorchcast AIGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API262,144 tokens33 to 49 ms4CC-BY-NC-4.0
H2O-Lightning-4BH2O.aiGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API40,960 tokens28.8 to 30.2 ms5Apache-2.0
Strands-Decider 2BStrands LabsGenerally availableOpen weights, self-hostedNoul, Choice, ScoreNo hosted API4,096 tokensNot published as of 23 Oct 2026Apache-2.0
GLiNER2.5-DecideFastino LabsGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted APINot published as of 23 Oct 202638 to 167 ms7Apache-2.0
CLMContrastive-LMGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API2,048 tokensNot published as of 23 Oct 2026Apache-2.0
Surogate Rune 26B-A4BSurogateGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted APINot published as of 23 Oct 2026Not published as of 23 Oct 2026Apache-2.0
LayaConvai InnovationsGenerally availableOpen weights, self-hostedChoice, Score, NoulNo hosted API512 tokens32.8 to 39.5 ms10Apache-2.0
VonwfzyxDeprecatedOpen weights, self-hostedChoice, Noul, ScoreNo hosted API8,192 tokensNot published as of 23 Oct 2026Apache-2.0
KevJared PalmerGenerally availableOpen weights, also hosted on OpenRouterChoice, Score, Noul$0.042/MTok8,192 tokens41.5 to 145.2 ms12Apache-2.0
  1. 1 H2O-Lightning-31B: H2O reports a median of 46 ms per request, one at a time, for its JevBench runbook on one RTX PRO 6000 (vendor-run).
  2. 2 Jeffan-31b-1710: The author reports a median of 43 ms, a mean of 103 ms and a p95 of 407 ms per request on one B200, running the public suite one request at a time (author-run).
  3. 3 Quyet-1.0-Large: Two independent medians for one question.
  4. 4 Torchcast Decision 27B: Torchcast's own measurements on one H100 80GB in bf16, end to end over HTTP, mean of 100 requests: 33.0 ms for one yes/no question and 49.0 ms for a request with three questions.
  5. 5 H2O-Lightning-4B: H2O's own median for all 231 public JevBench items, one request at a time over a persistent local connection: 29.8ms on one H100 80GB and 30.2ms on an RTX PRO 4500 Blackwell.
  6. 7 GLiNER2.5-Decide: Founder George Maloney's launch post gives p50 latency for short, single-document requests: 167ms on a 48-vCPU Intel Xeon Platinum 8581C CPU, and 38 to 47ms on GPU (T4, L4, V100, A100).
  7. 10 Laya: Convai's own measurements on one Tesla T4 GPU for a single question: 32.8ms on laya-multilingual and 39.5ms on the English checkpoint.
  8. 12 Kev: Self-reported by the author.

Head to head

Two models or tools on one page, with the spec table and a short written verdict. Most pair two decision models. The Clairic pages set Cloudflare's extraction model beside a decision model, beside GLiNER, and beside LLM structured outputs.

Jev vs other approaches

Jev against the other ways to get a label out of text: fine-tuned classifiers, named-entity recognition models, and language models asked for structured output.

Ranked

Every model a production app can call by API, with an independent score, ranked by a published rubric: quality and calibration from the System One Score, access, price, question types and tooling. The System One Score combines three independent boards, the Artificial Analysis Decision Index, the Jev Decision Index and JevBench. Open weights you run yourself are listed apart by System One Score, not ranked.