Choose an AI Decision Model for Your Workflow

OpenRouter list prices checked 2026-10-11, USD per one million tokens: Liquid d1, Microsoft-Decision-1, Clef Omni. These are published channel prices and can change. $0 output tokens does not mean a free workflow.

Separate direct-provider snapshot, checked 2026-10-10: OpenAI /v1/decisions lists $0.10 / 1M input tokens, with regional and long-context multipliers; no output, cache-read or cache-write token charge. This is not an OpenRouter offer. Foundry US/EU Datazone lists $0.042 / 1M input tokens (output N/A); Workers AI Clef lists $0.15 / 1M input tokens, including media input and no output-token charge. These older direct-channel facts were not repriced on 2026-10-11. See the dated primary sources below.

Compare the cost of the same accepted task: input preparation, media tokenization, model requests, retries and human review. Token accounting and API protocols differ across providers. Local hardware, energy, storage and operation are additional costs; local models are excluded from this hosted-price table. No task-level cost was measured.

These AI decision models choose among outcomes you define. A general LLM is still useful when the task needs an explanation, a newly written answer or an open-ended plan. A classification model API can fit a repeated routing step without replacing the rest of your application.

Each row separates the model from its access channel. UNKNOWN / NOT DOCUMENTED means the reviewed contract does not establish a capability. Original Foundry and Workers AI facts remain dated 2026-10-10; the OpenRouter, Liquid and local-model additions were checked 2026-10-11. Source links and dates appear below.

OpenAI: /v1/decisions takes model, input and a questions array. Handle typed answers and refusals explicitly. Its input contract is narrower than Responses: other message roles, function calls, files, audio, tools and item references are not supported. Public HTTP(S) image URLs are supported by the current reference; private or authenticated URLs are not confirmed. The reference was rechecked 2026-10-11; the guide’s earlier inline-only discrepancy was recorded on 2026-10-10. Continue to our Decisions API guide for its focused request walkthrough. Reference.

Microsoft: first provision the model in Foundry. The official text example uses state, a named questions object and type: choice with a criteria map, then validates answers.team.choice. It demonstrates a /providers/microsoft/v1/systemone route; confirm your deployment endpoint and authentication in its quickstart rather than swapping a model name into OpenAI code. We have not tested it. Foundry integration example.

Cloudflare: Workers AI exposes @cf/cloudflare/clef-omni with state and named questions. Its noul label is not OpenAI’s predicate field name. For self-hosting, follow the model card’s loader and output contract. Keep input conversion, answer parsing and error handling specific to each provider. Workers AI example.

Liquid: its own API accepts text/JSON state, named questions and images, with noul, choice and score. This is a separate channel from OpenRouter’s currently documented text route. OpenRouter: use its decision endpoint documentation and per-model input restrictions rather than transplanting a direct-provider request. Phocinae: the local runtime returns booleans, zero-based choice indexes and integer scores; normalize these deliberately in your application. Similar type names do not prove interchangeable schemas. Liquid API, Phocinae contract.

The author’s model card, checked 2026-10-11, reports 144.3M parameters (150M class). Hugging Face labels the model Apache-2.0; the disclosed mmBERT-small base is marked MIT. It targets local text-based typed decisions, not open-ended chat. Its documented phocinae-server runtime loads downloaded weights on your own hardware; we did not install it or download weights.

Author-reported evidence only: the card discloses training overlap involving evaluation rows, fitted language evaluation and calibration limits. It also reports a failed held-out acceptance gate and confusion between similar tool names. Those results cannot establish superiority over commercial models. Check the author’s full methodology before transferring a threshold to your domain.

Use it as a candidate for local triage with an explicit review path, not as the sole safety gate. Hardware, power, maintenance and validation still cost money. It is a separate local option, not a $0-total-cost entry in the cloud API table.

The following routes are editorial judgments under the stated prerequisites. Each decision needs an “uncertain / review” outcome outside automated actions.

Text support tickets. Input: a synthetic ticket → decision: billing, technical, shipping or other → route: OpenAI choice when OpenAI is already approved; Foundry choice when Azure deployment and governance are already in place → prerequisite: labeled ambiguous and multi-issue tickets → avoid migrating platforms solely for a lower token rate; avoid free-form text when your router only accepts four values.

Invoice anomalies. Input: extracted invoice fields → decision: duplicate, arithmetic mismatch, missing field or review → route: a text decision in the existing approved platform, after deterministic arithmetic and duplicate checks → prerequisite: a reliable extraction stage and audit trail → avoid treating a confident label as accounting proof or automatically authorizing payment. Raw invoice photos instead need an explicitly verified image route.

Image inspection. Input: an authorized product image → decision: damaged packaging / intact / unclear → route: OpenAI images for its hosted typed contract, Liquid’s own image API, or Clef-omni where its media pipeline or self-hosting is required → prerequisite: representative resolution and lighting examples, authorized image access → avoid assuming Microsoft’s Qwen base proves Foundry image support; do not expose a private image publicly merely to make a URL work.

Audio complaints. Input: a synthetic complaint recording → decision: billing dispute / service outage / other → route: Clef-omni’s documented Workers AI audio path (not its OpenRouter route) → prerequisite: duration, language and consent rules checked against the hosted contract → avoid native-audio claims for OpenAI Decisions or Microsoft; a separate transcription step is a different pipeline and cost.

Video events. Input: an authorized short test clip → decision: event present / absent / ambiguous → route: Clef-omni on Workers AI → prerequisite: sampling, audio, clip count and token-window constraints fit the task → avoid inferring exact event timing or a safety guarantee from a clip-level decision probability. Use human review for consequential actions.

Several questions about one input. Input: one incident description → decisions: urgency, team and impact score → route: OpenAI’s independent questions or Clef’s typed-question schema; Foundry only after confirming the complete deployment contract → prerequisite: questions do not secretly depend on a previous answer → avoid assuming shared input means a bulk asynchronous API, or that independent outputs create a reliable causal chain.

Very low latency / high throughput. Input: repeated short classification items → decision: route or reject → route: shortlist the service closest to your application and permitted region → prerequisite: measure p50/p95 latency, throttles, retries and accepted decisions per second under the same load → avoid choosing from a vendor’s headline speedup measured against a different workload.

Private deployment. Input: controlled internal documents → decision: one of a fixed policy set → route: Phocinae for a small local text-decision candidate, or Clef open weights for a larger multimodal runtime, if your team can validate and secure the infrastructure → prerequisite: capacity, license review, access controls and a labeled test set → avoid confusing Foundry model deployment with possession of downloadable weights. Do not promise local OpenAI Decisions weights.

Existing ecosystem. Input: your application’s current approved data path → decision: which adapter to maintain → route: Foundry for an established Microsoft deployment, Workers AI for a Cloudflare application, OpenAI Decisions for an existing OpenAI integration → prerequisite: the required modality and contract are verified → avoid ecosystem convenience when a missing input type or data policy makes that route unsuitable.

Hosted text routing on OpenRouter. Input: a short labeled ticket → decision: one allowed queue or review → route: compare Liquid d1, Microsoft-Decision-1 and Clef Omni on the same text task → prerequisite: per-channel contract, error handling and budget verified → avoid selecting solely from the lowest listed token price or assuming Liquid images and Clef audio/video work on this channel.

The official model card reports testing on one H200, with about 64 GB GPU memory for the BF16 backbone. That is the reported test environment, not proof that every deployment must use an H200, and not a total-memory guarantee for every batch or media workload. It does not establish effortless operation on a consumer GPU.

If a provider publishes an applicable quantized release or a deployment framework supports it, verify that exact artifact and runtime, memory use, media preprocessing and answer parity before relying on it. This page did not find or test a specific smaller-GPU configuration. No weights were downloaded or inference run.

Microsoft Foundry deployment configures hosted access; it is not an open-weight grant. OpenAI’s Decisions endpoint provides hosted model access, not a locally downloadable Luna release. A private hosting requirement therefore changes the shortlist before any accuracy comparison.

Microsoft publishes its own evaluation discussion; Cloudflare publishes model-card evaluation tables with suite and cohort context. Those are vendor-reported results, not measurements by AgentSkillsHub. This page does not reproduce an accuracy leaderboard because no common, independently reproduced shared experiment across these models was performed. A shared dataset name alone does not establish identical revisions, prompts, preprocessing, serving conditions or calibration. Microsoft evaluation, Cloudflare evaluation.

A future task test should keep the same labeled inputs and decision definitions, report abstentions/refusals and failed requests, and separate quality from p50/p95 latency, throughput and price per completed decision. Calibration asks whether probability ranges correspond to observed frequencies on representative data; it is not the same as accuracy. A score is not approval to execute an action.

Start with synthetic or properly authorized inputs. Choose thresholds from a held-out set, perform deterministic validation, and require human approval for high-risk effects. No threshold here is claimed to be validated. Review our evidence methodology before interpreting any future result.

Write down the allowed inputs, output labels, failure/refusal handling, data location and maximum acceptable review burden. Eliminate routes whose documented inputs or deployment boundary do not fit. Prepare a small labeled set that includes unclear cases, then seek approval for a bounded test with its own cost limit.

For OpenAI-specific integration details, use the Decisions guide. For other architectural choices, return to the Compare library. This comparison remains documentation-based; it does not certify vendor access, model quality or production suitability.