Should a support ticket go to billing or technical support? Does a record need human review? Decision models choose from answers you define and give a confidence score for each. Compare what they accept, what they cost and whether you can run them yourself.
We cannot yet name the most accurate or fastest option: the available provider examples use different tests. A high confidence score can still be wrong. Try each model on examples from your own work before letting it approve an action.
Compare models you can use online
An API lets your application send information to a model and receive its answer. A request is one call to that service. Prices use tokens: the small pieces of text the model processes. Images, audio and video may also count toward billed input.
Sort support tickets, review images or score records against choices you define. This is the new v1.1 Gateway offer; its quality has not been independently compared here.
Input limit
262K
Questions per request
Not confirmed
Input price / 1M tokens
$0.02
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, choices and ordered scores
Gateway lists 262,144 context tokens. Its maximum question count is not confirmed. The downloadable version’s supplied code has an 8,192-token limit; that is a different serving setup.
Gateway lists text and image input for classification, routing and scoring. The provider reports improvements on its own decision tests; these are vendor results, not a general-chat score or a measured thinking-effort comparison.
Gateway lists $0.02 per million input tokens and zero output-token price. This is half the older v1 native API input quote, but billing details and quality differ; compare actual usage on the same records.
Choose from allowed answers without generating a paragraph: route a ticket, classify a document or return a probability for each option.
Input limit
131K
Questions per request
Up to 512
Input price / 1M tokens
$0.06
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, multiple-choice probabilities and ordered rating scores. No generated explanation.
Drex v2.0 accepts up to 131,072 tokens of shared information, and up to 262,144 including the longest question. It allows up to 512 questions and 4,194,304 bytes for the serialized request. Requests above a limit are refused; v2.0 no longer hands long requests to v1.5.
Send text state and up to 512 typed questions to /v1/systemone. Drex v2.0 returns probabilities, not a written explanation. Validate those probabilities on examples from your own workflow.
Native v2.0 costs $0.06 per million billed input tokens, with no output, request or question fee. Only successful evaluations are charged. Use the returned usage totals to compare bills.
Choose from your own answer options using text, images, audio or video. Clef Omni returns a confidence score for each answer instead of writing a paragraph.
Input limit
64K
Questions per request
Up to 64
Input price / 1M tokens
$0.15
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, choice probabilities and ordered rubric scores; no generated explanation.
The hosted context is exactly 64,000 tokens, shared by media, questions and text. If media and questions exceed it, the request fails. Otherwise text state is truncated to fit. This is separate from Clef’s 65,536 and Flash’s 24,576 limits.
The hosted schema explicitly accepts up to 64 typed questions, four embedded images, four audio clips and two videos. Audio clips can be 300 seconds / 8 MiB each; videos 60 seconds / 16 MiB each, sampled at 2 frames / second. Audio and video total at most 16 MiB decoded. Remote media URLs are not accepted.
Input costs $0.15 per million billed tokens; answers are free. The provider estimates up to 1,024 tokens per image, about 780 per minute of audio, and up to about 15,400 per minute of video frames, plus audio. These estimates are not a measured price for your task. We could not confirm caching rates or other charges.
Send a document to Nace’s online Drex service and ask it to choose an answer, route a ticket or score a condition. It returns the requested decision instead of a paragraph.
Input limit
131K
Questions per request
Up to 512
Input price / 1M tokens
$0.05
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, multiple-choice probabilities and ordered rating scores. No generated explanation.
State alone can contain 131,072 tokens; state plus the longest question can contain 139,264. Other questions do not add to that row limit. A request allows up to 512 questions and 4,194,304 serialized UTF-8 bytes for state plus questions. Check all boundaries.
Send model drex-v1.5 to POST /v1/systemone at console.nace.ai. The provider documents text state and yes/no, choice and rating questions. It returns probabilities; validate calibration on your own records.
Nace charges $0.05 per million input tokens for v1.5, with no separate charge for answers, requests or questions. Only successful requests (HTTP 200) are billed. Use the input-token count in your response to estimate usage. OpenRouter advertises $0.04, but this model was missing from both catalogs we checked, so we could not confirm access or billing for that offer.
Classify a ticket, choose a workflow step or grade an answer against your rubric. Microsoft’s new text-only decision model returns probabilities over the options you supply, rather than writing a response.
Input limit
33K
Questions per request
Not confirmed
Input price / 1M tokens
$0.042
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Probabilities for supplied yes/no, choice and rating options; no generated explanations
Foundry and Gateway agree on 32,768 input tokens. A maximum number of questions and a cache-read price remain unverified. It does not generate explanations or accept images.
Foundry lists Microsoft-Decision-1 version 1 as Preview. The confirmed Gateway route supports yes/no, choice and ordered scoring, with multiple questions sharing input. It needs a Gateway key without a separate Azure deployment.
Microsoft lists $0.042 per million input tokens and free answers; AI Gateway lists the same rate. Check how your service counts background information shared across questions. This is a token-price comparison, not a measured total bill or cost per correct decision.
Use this hosted model when your task has a fixed set of answers, such as routing a ticket or checking a record. It accepts images directly; it is distinct from the two downloadable d1 checkpoints.
Input limit
Not verified
Questions per request
Not confirmed
Input price / 1M tokens
$0.04
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Choice distributions, ordered Score ratings and yes/no Noul probabilities; no free-form text.
Gateway lists liquid/d1 with 65,536 context and its October 9 announcement explicitly supports images. This is newer than the text-only routed-access note in Liquid’s October 5 launch post. The native maximum context remains unverified; max_tokens=0 describes typed answers rather than zero input capacity.
The native /decisions/v1/systemone endpoint accepts text and images under model d1. Context, maximum questions and calibration on your own labels remain unverified.
Input starts at $0.04 per million tokens. Every question is billed separately, including its text and all images. Images count as 1.5 tokens per 32×32-pixel patch. Add the billed tokens for all questions before using the calculator.
Use GPT-6 Luna to answer yes/no questions, choose from named options or score an answer against your rules. Each answer has a confidence score. Chat benchmark results do not tell us how accurate these decisions will be.
Input limit
Not verified
Questions per request
Not confirmed
Input price / 1M tokens
$0.1
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Typed predicates, choices and rubric scores with probabilities and confidence fields.
Gateway lists 1,050,000 context tokens for its route. Its zero max_tokens is typed-output metadata, not proof of zero capacity or a native API limit.
Use POST /v1/decisions with model gpt-6-luna. Input supports text and images. Native context and question-count caps and an effort contract are not verified in the reviewed guide.
Input starts at $0.10 per million tokens. Answers and storing or reusing cached input have no separate charge. Regional processing and long-context input premiums apply: your selected region or a very long request can increase the price. Other GPT-6 Luna APIs have their own rates.
A decision service from TypeSafe, available through AI Gateway. Gateway announced it on September 16, while its catalog says September 15; neither confirms when the underlying model first launched. Comparable independent accuracy and confidence-score tests are still missing.
Input limit
32K
Questions per request
Not confirmed
Input price / 1M tokens
$0.042
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Typed choices, rubric scores and Boolean probabilities
Gateway declares 32,000 tokens for shared state plus the longest question, and 64,000 total request tokens. This is not a 64K single-state context. The maximum question count is unverified.
The exact evaluation contract accepts text state and typed Choice, Score and Boolean questions. It does not stream free-form output; no independent speed advantage or calibrated confidence is inferred.
AI Gateway quoted $0.042 per million input tokens and free answers when checked on October 4. Its model page rounds that price to $0.04; our calculator uses the exact $0.042 quote. Caching and other fees are not confirmed.
A model for choosing from fixed answers through AI Gateway. Access was announced on October 1. We have not confirmed the underlying model’s launch date or found comparable independent quality tests.
Input limit
8K
Questions per request
Not confirmed
Input price / 1M tokens
$0
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, bounded choices and rubric scores
The exact Gateway API lists 8,192 tokens. A maximum question count is unverified; this is not a guarantee that every state plus question fits.
The hosted evaluation API documents text state and typed yes/no, choice and score questions. No hosted image support or independent calibration sweep is established.
Choose Boundless to use the standard or -free offer at no charge through October 31, 2026. After that, the standard offer becomes paid and the -free offer stops. Later prices, caching rates and the exact cutoff time are not confirmed.
Use text and images to choose from answers you define. The model files are available under Apache 2.0, and Cloudflare offers an online decision service. We have not independently reproduced the provider’s accuracy or speed comparisons.
Input limit
25K
Questions per request
Up to 64
Input price / 1M tokens
$0.038
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, bounded choices and rubric scores
Hosted context is 24,576 tokens, including media and questions. Media that exceeds the window makes a request fail; otherwise text state is truncated to fit the remaining space. Inspect retained context before acting on a decision.
The model card declares text, JSON, images and video input. The hosted schema explicitly documents 1–64 named questions and up to four embedded images; hosted video request support is not established by the model-card claim. No adjustable effort sweep is verified.
Cloudflare lists $0.038 per million input tokens and free answers. Images add to your input usage. We could not confirm caching rates or other charges; check the usage reported by your service.
Use text and images to choose from answers you define. The model files are available under Apache 2.0, and Cloudflare offers an online decision service. We have not independently reproduced the provider’s accuracy or speed comparisons.
Input limit
66K
Questions per request
Up to 64
Input price / 1M tokens
$0.24
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, bounded choices and rubric scores
Hosted context is 65,536 tokens, including media and questions. Media that exceeds the window makes a request fail; otherwise text state is truncated to fit the remaining space. Inspect retained context before acting on a decision.
The model card declares text, JSON, images and video input. The hosted schema explicitly documents 1–64 named questions and up to four embedded images; hosted video request support is not established by the model-card claim. No adjustable effort sweep is verified.
Cloudflare lists $0.24 per million input tokens and free answers. Images add to your input usage. We could not confirm caching rates or other charges; check the usage reported by your service.
A decision model with downloadable Apache 2.0 model files and a dedicated online API. It accepts text and images and gives confidence scores for fixed answers. The provider’s tests are not comparable independent results, and its launch date is not confirmed here.
Input limit
262K
Questions per request
Up to 128
Input price / 1M tokens
$0.04
Output price / 1M tokens
$0
Input limits and supported answers
Answer format: Yes/no probabilities, bounded choices and rubric scores
The Decisions API accepts fewer than 262,144 total input tokens, including state, images and all questions; this is an input cap, not a chat prompt-plus-output context allowance.
The dedicated Decisions API accepts text, JSON and images and supports 1–128 named questions in a request. It returns typed probabilities; no chat tool-calling interface or adjustable reasoning-effort sweep is established.
Perplexity charges $0.04 per million input tokens, using the count returned with each answer. Answers are free, with no extra fee per request. Background information, all questions and images count as input. Caching is not confirmed.
Enter how many input tokens you are charged for per request, then your monthly number of requests. Include the questions and any image processing charges expressed as tokens. Use your provider’s usage report when you have one: the same task can be billed differently by different services.
This estimates input charges only. Extra charges for your processing region or very long requests, output charges, retries, taxes and other fees are excluded; credits are not deducted.
Perplexity Decider v1.1 27B · Gateway$2
Base monthly input charge at $0.02 per million billed tokens.
Estimate uses the Gateway input-token quote. Output tokens are free; additional per-request fees and caching terms are unverified.
Drex v2.0 · native API$6
Base monthly input charge at $0.06 per million billed tokens.
Native v2.0 charges $0.06 per million billed input tokens. Requests over its input limits are refused instead of being answered by v1.5.
Cloudflare Clef Omni$15
Base monthly input charge at $0.15 per million billed tokens.
Answers are free. Include questions, images, audio and video in your billed input. Video usage depends on length and resolution.
Drex v1.5 · native API$5
Base monthly input charge at $0.05 per million billed tokens.
Answers are free, with no separate request or question fee. Credits are not deducted from this estimate. If an old model ID redirects to a newer version, you pay the rate for the version that answers.
Microsoft Decision-1$4.2
Base monthly input charge at $0.042 per million billed tokens.
Answers are free. Check how your chosen service bills the information shared across multiple questions.
Liquid d1$4
Base monthly input charge at $0.04 per million billed tokens.
Each question is charged for its text and images again. No answer tokens are generated, so there is no output-token charge; the input still costs money.
GPT-6 Luna Decisions$10
Base monthly input charge at $0.1 per million billed tokens.
This shows the starting input price. Your region or a very long request can cost extra. Answers and storing or reusing cached input have no separate charge.
Jev$4.2
Base monthly input charge at $0.042 per million billed tokens.
AI Gateway lists free answers. Caching and other fees are not confirmed.
Laya$0
Base monthly input charge at $0 per million billed tokens.
Temporary Boundless promotion; later rates and fees are not confirmed.
Cloudflare Clef Flash$3.8
Base monthly input charge at $0.038 per million billed tokens.
Answers are free. Images are charged at the input rate. Caching rates and other fees are not confirmed.
Cloudflare Clef$24
Base monthly input charge at $0.24 per million billed tokens.
Answers are free. Images are charged at the input rate. Caching rates and other fees are not confirmed.
Perplexity Decider v1 27B$4
Base monthly input charge at $0.04 per million billed tokens.
The provider says answers are free, with no additional fee per request.
Check how many decisions are correct and how often you need to retry before choosing by price. Also check each input limit: a low estimate does not mean your request will fit.
Could you run a decision model yourself?
Perplexity, FrontiersMind, Nace, Liquid AI and FINAL-Bench offer downloadable models that you can run on your own hardware. You need enough memory, the correct software and a way to check answer quality. We do not have a hosted price for these downloads; compare hardware and running costs instead.
Check the license for the trained model files, often called weights, as well as the software used to run them. Drex DLM allows noncommercial use; Drex v1.5 and v2.0 have restrictions tied to revenue, funding and competing products. The software’s Apache license does not override these rules. You may also need to change the software settings to use the model’s full input capacity.
Choose labels or yes/no answers for records in English, Hindi, Tamil and Kannada. This downloadable model returns probabilities and can serve a ticket-routing workflow on your own hardware.
Shared information plus one question must fit within 8,192 tokens. Each question reads the shared information separately; answers do not change one another.
Apache 2.0 weights. The model card recommends about 55 GB of free GPU memory, typically an 80 GB card, or several GPUs. Install the provider’s software and test confidence thresholds on your own examples before automating decisions.
Run the v1.1 decision model on your own GPU to assign labels and probabilities to text or images. This download has different limits from the hosted Gateway offer.
The supplied prepare function limits text and all images for one decision to 8,192 tokens. Training and evaluation used images up to 512×512; the serving default allows up to 2048×2048 and uses more memory.
Weights are Apache 2.0. The card asks for Python 3.12+, CUDA and roughly 49 GiB for weights plus working memory. Its private-repository wording conflicts with publicly readable files; confirm download access. Keep the supplied attention and calibration settings to preserve the tested behavior.
Run typed decisions on your own hardware. This 35B model uses about 3B parameters per token; it returns option probabilities rather than prose.
The documented runner default is 16,384 tokens, with an explicitly enabled maximum of 131,072. In llama-server, context is split across slots: the example -np 2 / -c 32768 gives 16,384 per slot. Native API row and question limits do not establish local runner limits.
Weights use the modified Nace.AI Open RAIL-M license, while runtime code is Apache 2.0. The weight license restricts competing products and generally requires a separate license for non-research/non-personal use above $1 million prior-year gross revenue or total funding. Read the exact license and its other use restrictions before deployment.
Run this decision model on your own hardware. Give it background information and questions with fixed answer options; it returns confidence scores instead of prose. It is based on Efficient-DLM-8B.
The exact card documents 32,768 maximum tokens, enabled explicitly; runners default to 16,384. Align the context, batch and microbatch caps. The provider has checked Q8_0 at 16,384; its full-window equivalence is not established by that test.
Weights are CC BY-NC 4.0, which does not grant ordinary commercial use. Runtime code is separately Apache 2.0. The provider reports tests on Apple M5 Max/M5 Pro and an H100; CPU-only inference is untested. These are vendor compatibility checks, not independently measured performance.
Run the downloadable 9B Drex model on your own hardware. It uses the same fixed-answer format as the online service, but has separate hardware, license and setup requirements.
The documented runner default is 16,384 tokens, with an explicitly enabled maximum of 131,072. In llama-server, context is split across slots: the example -np 2 / -c 32768 gives 16,384 per slot. Native API row and question limits do not establish local runner limits.
Weights use the modified Nace.AI Open RAIL-M license, while runtime code is Apache 2.0. The weight license restricts competing products and generally requires a separate license for non-research/non-personal use above $1 million prior-year gross revenue or total funding. Read the exact license and its other use restrictions before deployment.
This second Darwin release adds training for choosing actions and workflow steps. Test it separately from version 1, including whether its confidence scores match how often it is right.
The configuration lists 262,144 internal positions, but a validated maximum for the decision interface is unverified. The v2 benchmark used an explicit 32,768 limit for one task; that is a test setting, not a universal endpoint allowance.
The card declares Apache 2.0; bundled autojev code carries its MIT notice. The provider describes about 54 GB of BF16 weights, before runtime memory. v1’s card corrects mixed-run numbers. v2 discloses training on other splits of tasks that also appear in S1MB; we have not audited that leaderboard or imported its scores.
A downloadable text model that scores your fixed answer choices in one processing pass per question.
The configuration lists 262,144 internal positions, but a validated maximum for the decision interface is unverified. The v2 benchmark used an explicit 32,768 limit for one task; that is a test setting, not a universal endpoint allowance.
The card declares Apache 2.0; bundled autojev code carries its MIT notice. The provider describes about 54 GB of BF16 weights, before runtime memory. v1’s card corrects mixed-run numbers. v2 discloses training on other splits of tasks that also appear in S1MB; we have not audited that leaderboard or imported its scores.
An experimental local decision checkpoint for text with images or audio. It can route spoken requests without generating a transcript or free-form answer.
The request limit is 16,384 combined text/image/audio positions. With images, state and question text are cut to 896 tokens. Audio clips are cut at 30 seconds, and training covers English speaker–assistant requests.
Weights use the LFM Open License v1.0, with a commercial-use revenue limitation; review the exact checkpoint license before deployment. Provider benchmark and device timings are not independent measurements. Hosted token prices are unknown; local inference has hardware and operating costs.
A local decision checkpoint for text and images. It returns fixed answers in one pass rather than writing a response.
The model card documents 32,768 context tokens for d1-3B.
Weights use the LFM Open License v1.0, with a commercial-use revenue limitation; review the exact checkpoint license before deployment. Provider benchmark and device timings are not independent measurements. Hosted token prices are unknown; local inference has hardware and operating costs.
Test the decisions you will actually act on
Give each model the same records, questions and answer choices, with known correct answers. Count wrong approvals, cases that should have gone to a person, and unanswered questions. Measure response time and cost too. Check confidence scores on separate examples: when a model says 90% confident, is it actually right about nine times out of ten? Record how many examples you tested; a small test leaves more uncertainty.
For example, a support-ticket test can ask for the destination team, urgency and whether the record contains enough information. Record when the model is confidently wrong and what a retry costs. Extra questions and images can add input charges. Liquid d1 bills each question as its own prompt, including its images; multiply the billed prompt by the question count before using the budget estimate.
What happens when the model is unsure?
You can send uncertain answers to another model or a person. AI Gateway supports confidence rules that call a second model, but both calls are billed. Choose the threshold using examples with known answers: a confident mistake will not be caught just by raising the cutoff. How Gateway’s fallback rules work ↗
Will your information fit?
The same model name can come with different limits online and on your own hardware. For example, Perplexity Decider v1.1 on Gateway lists much more input space than its downloadable software. Choose the service first, then check its limits.
Drex v2.0 now accepts up to 131,072 tokens of shared information, or 262,144 including the longest question. Longer requests are refused; they are no longer handed to v1.5. The whole request must also fit its 4 MB size limit. Running the downloadable v2.0 yourself has a separate limit of up to 131,072 tokens.
Jev on AI Gateway allows 32,000 tokens for the information shared across questions plus the longest question, and 64,000 tokens across the whole request. Its maximum number of questions is not confirmed. The exact input quote is $0.042 per million, with zero output-token price. Laya on AI Gateway lists an 8,192-token limit; its maximum number of questions is not confirmed. The free quote requires Boundless and ends after October 31, 2026. The older Perplexity v1 native API accepts fewer than 262,144 total input tokens and up to 128 questions. The new v1.1 Gateway offer lists a 262,144-token context, but its question limit is not confirmed. Its local download accepts only 8,192 tokens per decision. Clef documents 65,536 tokens, Flash 24,576 and Omni 64,000, each with up to 64 questions. Media and questions share the available space. Some background text can be cut off to make everything fit. A larger limit does not guarantee that the model will use every detail correctly. Clef and Flash model descriptions mention video, but their hosted API instructions explicitly document images. Omni’s API also documents audio and video. Check the service you will actually use.
A downloadable model still needs hardware, electricity, software and maintenance. Include those costs when comparing it with an online service.
AI Agent Store