A decision API is useful when an application needs a bounded answer: which queue receives a message, whether a document satisfies a criterion or which level of a rubric applies. Paying for a generated explanation may be unnecessary when the next step requires only a typed result and its probability. The challenge is comparing services whose billing units, question grouping and input limits differ.
This comparison uses direct prices where available and labels router-only rates separately. Prices, limits and promotions were checked on November 17, 2026. The calculations use stated token assumptions; they are not measured accuracy, latency or production bills. Start with the shape of your request, then choose a model that meets the decision’s quality and operating requirements.
- 01Normalise the inputA thousand questions can mean a thousand repeated sources or one source shared across many questions. Those produce very different bills.
- 02Name the serving routeDirect provider rates, router promotions and free endpoints are different offers. Preserve that distinction in procurement and monitoring.
- 03Check limits before priceInput modality, grouping rules and effective context capacity can exclude a route before its token rate becomes relevant.
- 04Calibrate the decisionEvaluate probabilities at the thresholds your workflow will use. Similar response schemas do not imply equal accuracy or interchangeable scores.
01 — Decision boundaryUse a decision model for a bounded application choice
A useful question has explicit criteria and a result the application can interpret. For a hypothetical support workflow, that might be whether a message requests a refund or which team should review it. The application still decides what to do with the result. A high probability of a refund request is evidence about the message; it is not permission to issue money or a determination that the customer is entitled to it.
Our decision-model API explainer covers the distinction between probabilities and generated text. If the user needs a readable explanation, you may still need a generation step or a human response. Count that downstream work separately. Moving one classifier call to a cheaper endpoint does not automatically reduce the cost of the complete customer journey.
Define the failure that matters before comparing prices. A routing mistake may create a short delay, while a mistaken automatic rejection may create a much larger service problem. Choose thresholds and escalation rules around those consequences. A model that is cheap per input token can be expensive if it sends many straightforward cases to manual review or confidently mishandles the cases that matter most.
A decision score is an input to your workflow. The application still owns action permissions, exception handling and the consequence of an incorrect result.
02 — Normalised pricesCompare 1,000 separate requests on equal assumed input
For a transparent baseline, assume 1,000 separate requests, each with 2,000 source tokens and 20 additional billed tokens covering the question and required request content. That is 2.02 million input tokens. Multiplying each published input rate by 2.02 produces the table below. This holds assumed billable tokens steady, not tokenisation, accuracy or real request overhead. It excludes caching, grounding, regional premiums, free allowances and other charges.
Direct rates come from Google’s decision pricing, the OpenAI Decisions guide, Microsoft’s model announcement, Perplexity’s quickstart and Workers AI pricing. OpenAI’s $0.10 rate is specifically for GPT-6 Luna through the Decisions endpoint, not every API use of that model.
For HeyGen, the verified rates are the OpenRouter listings for Decision and Decision Turbo. They are not presented as a direct HeyGen rate card. The listings date to November 13; that is a router listing date, not independently verified evidence of the vendor’s original launch date. Turbo is labelled preview, which should be part of a production adoption decision.
| Model | Verified route | Input / 1M | 1,000 requests |
|---|---|---|---|
| Gemini Decision 1 | Google direct | $0.100 | $0.20200 |
| GPT-6 Luna Decisions | OpenAI direct | $0.100 | $0.20200 |
| Microsoft Decision 1 | Microsoft direct | $0.042 | $0.08484 |
| Decider V1.3 | Perplexity direct | $0.020 | $0.04040 |
| Decider Flash V1.3 | Perplexity direct | $0.010 | $0.02020 |
| Clef | Workers AI direct | $0.240 | $0.48480 |
| Clef-flash | Workers AI direct | $0.038 | $0.07676 |
| HeyGen Decision | OpenRouter | $0.140 | $0.28280 |
| HeyGen Decision Turbo | OpenRouter | $0.090 | $0.18180 |
03 — Shared evidenceChange the request shape before changing the model
Google’s Decisions API guide permits up to 1,000 questions within a 262,144-token request budget. Shared input is billed once, while each question adds its instructions, criteria and formatting. The service reports zero output tokens and does not support caching or the Batch API. These rules make a shared-source workload materially different from 1,000 separate calls.
Consider the same hypothetical 2,000-token source with 1,000 questions whose complete additional billed content is assumed to average 20 tokens each. The combined input is 22,000 tokens, costing $0.0022 at Google’s $0.10 rate. The separate-request baseline cost $0.202. This reduction comes from removing repeated source input under Google’s documented billing rules. It does not demonstrate that every real question fits in twenty tokens or that every provider applies the same grouping.
OpenAI permits up to 400 questions per request, so this particular thousand-question workload needs at least three requests there. Our OpenAI Decisions API guide covers that endpoint’s operating details. Decide whether the questions are independent before grouping them: a second question that depends on the first answer may require a subsequent request. Reusing evidence and creating a sequence of decisions are different workflow designs.
| Google-specific example | Assumed billed input | Input cost |
|---|---|---|
| 1,000 separate requests | 1,000 × 2,020 = 2,020,000 | $0.2020 |
| One shared-source request | 2,000 + 1,000 × 20 = 22,000 | $0.0022 |
04 — Grouping behaviourKeep probability comparisons consistent across requests
Perplexity V1.3 models process questions together with shared state billed once per group. When the combined state and questions exceed the context allowance, the service groups questions and bills the state again for each group. Its documentation also notes that questions can influence one another indirectly through shared state. Keep grouping consistent when comparing probabilities or tuning thresholds, rather than assuming regrouping is merely a billing change.
The ordinary Perplexity decision path should not be confused with its specialised limits: Locate has a 128-question cap and text-only state, while web grounding allows up to 100 questions. Grounding also adds search or fetch charges and additional evidence tokens. If the application already supplies authoritative source material, determine whether external retrieval is necessary before adding that cost and another source of variation.
For any provider, store enough request context to reproduce a decision during evaluation: model version, question wording, options, grouping and the evidence supplied. A threshold calibrated on one arrangement may behave differently after an alias update or a change in surrounding questions. Treat those changes as reasons to recheck the workflow, even if the response schema is unchanged and the new model is advertised as a drop-in replacement.
Keep the evidence, question wording, grouping and model version fixed when comparing scores. Change one deliberately and recheck the acceptance thresholds.
05 — Input constraintsSelect the route that can actually read the evidence
Gemini Decision accepts text, images, audio, video and PDF evidence, with text question instructions. OpenAI Decisions supports text, images and structured conversation or tool evidence, but its guide excludes direct audio and file inputs. Perplexity’s current Decider models accept text and images. HeyGen’s router listings include text, images and video, with context windows of 131,072 for Decision and 262,144 for Turbo. These are different input contracts, not interchangeable labels for multimodal support.
Cloudflare’s hosted Clef-flash context is 24,576 tokens following its October update, while Clef is listed at 65,536. There is a further route-specific caution: OpenRouter’s current Clef pages warn that the Workers AI route truncates long text state to roughly the first 2,000 tokens, with images treated separately. Check the actual endpoint behaviour and provider documentation before assuming the larger model context label guarantees that every part of a long document is read.
Construct a small evidence-coverage check for your chosen route. Put a relevant fact near the beginning and another near the end of a representative input, then verify that the application’s evaluation can distinguish the cases. Use appropriate labelled examples rather than assuming that successful HTTP responses establish complete reading. If preprocessing removes pages, samples video frames or converts audio to text, that preprocessing is part of the system being evaluated.
06 — Route economicsTreat free endpoints and promotions as separate offers
OpenRouter’s Clef listing shows a $0.192 input rate through a 20% DeepInfra promotion, while its Cloudflare provider row remains $0.24. Its Clef-flash listing shows a $0.019 DeepInfra promotional rate, while Cloudflare remains $0.038. Those router offers do not establish a reduction in the direct Workers AI rate. Record the provider route as well as the model name in a cost comparison.
The router also lists Decider V1.3 free and Decider Flash V1.3 free, with zero token prices and rate limits. Perplexity’s direct API is paid. A free endpoint can be useful for a permitted evaluation, but it should not be described as a permanent, unlimited production entitlement. Check availability, limits and the terms that apply to the data you intend to send.
Build the business case around a route the application can actually use. A promotional headline is less valuable if its capacity, region or provider choice does not fit the workload. OpenAI’s direct Decisions pricing also requires attention to cache writes, regional processing and long-context premiums. The AI cost-optimisation playbook explains why routing, caching and task design belong in the same ledger as the advertised base rate.
Start with the provider rate
Record the endpoint, input unit and applicable premiums before considering another serving route.
Name the selected provider
Keep promotions, capacity and route restrictions attached to the quoted price.
Measure the complete bill
Include repeated evidence, retrieval and fallback work when estimating operating cost.
07 — Portable interfaceUse common tooling without assuming common behaviour
Hugging Face’s November 10 decision-model announcement describes support through DecisionPipeline, AutoModelForDecision, DecisionConfig and a compatible serving interface. This makes the category easier to integrate and evaluate in the open-model ecosystem. It does not establish that every compatible model supports the same modalities, request limits, probability calibration or deployment hardware.
If local deployment is attractive, compare the complete operating requirement with the hosted route. You need the model artifact, its licence, a supported runtime, sufficient hardware and a maintained service. A model with no per-request vendor bill still consumes compute and engineering time. Conversely, a hosted endpoint may simplify operations while introducing rate, data-handling or availability constraints that matter to your application.
Keep an adapter boundary between your application’s decision contract and the provider-specific request. That can make a future comparison easier, but resist flattening away important differences. A service that returns a probability distribution, a selected category or an expected rubric score may need distinct interpretation. Preserve the evidence and schema needed to understand the answer rather than converting every result into a generic confidence number.
08 — Selection trialChoose on decision quality at the operating threshold
Build a labelled evaluation set that includes the ordinary cases, ambiguous cases and costly mistakes in your workflow. Define which results can proceed automatically, which need another check and which go to a person. Compare models at those thresholds, including false positives, false negatives and the volume of manual review. A single overall accuracy figure can hide a poor result on the category that drives operational cost.
Keep the selection trial proportionate. Start with the input routes that meet your modality and context needs, then compare direct cost and realistic grouping. Evaluate latency under your intended request shape rather than borrowing a vendor’s measurement from another model or workload. Include the cost of fallbacks and rejected attempts when calculating the expense of producing an accepted decision. No inference benchmark was run for this article.
Our AI transformation services connect model selection to a defined business process. For decision APIs, the right choice is the route that reads the required evidence, produces usable probabilities and meets the workflow’s error and cost limits. The cheapest token rate is a useful candidate, not a conclusion. Request design and a well-calibrated escalation path can matter more than a small difference in the rate card.
Price the decision you will actually run
A cost-per-thousand comparison is useful only when the input assumptions and request grouping are visible. Start with direct rates, label router offers and account for evidence reuse, limits and additional features. Then test the models at the thresholds your application needs. A decision API earns its place by producing dependable, actionable results within the complete workflow’s operating budget.