Agent & decision APIs compared

Independent, source-linked comparison of AI agent and decision APIs: pricing, limits, status, and safety notes.

15 of 15 · Last verified 2026-10-15

OpenAI · Decision model

Dedicated POST /v1/decisions endpoint that returns typed answers (probability, choice, or score) about text and images instead of generated text. Model: gpt-6-luna (only model available).

$0.10 per 1M input tokens; no output, cache-read, or cache-write charges.Verified 2026-10-15 against official page

BetaPartly verified

Microsoft-Decision-1

Microsoft (Foundry) · Decision model

Small decision model in Microsoft Foundry that returns typed numeric decisions (noul / choice / score) about text or JSON state. No free-form output or rationale.

$0.042 per 1M input tokens; output $0 (OpenRouter listing, provider Azure).Verified 2026-10-15 on OpenRouter; Foundry/Azure price page not checked

BetaPartly verified

Vercel AI Gateway — Decisions endpoint

Vercel · Gateway / proxy

OpenAI-compatible POST /decisions endpoint on Vercel AI Gateway that routes to any decision model on the gateway (e.g. openai/gpt-6-luna-decisions, typesafe-ai/jev).

Per-request cost is reported in response metadata; the docs page does not state a markup or price table.Unverified — gateway pricing not on the page checked

BetaPartly verified

Jev (TypeSafe AI)

TypeSafe AI · Decision model

Hosted "System One" decision model from TypeSafe AI: send a state plus typed questions (noul, choice, score) to POST /v1/systemone and get probabilities back. Current versions jev-1.14.0 and jev-1.13.0.

$0.042 per 1M input tokens ($42 per billion) for both Jev 1.14 and 1.13; output tokens free. Long-context mode costs the same per token.Verified 2026-10-15 on docs.typesafe.ai/models

Partly verified

Cloudflare Clef (Clef, Clef-flash, Clef-omni)

Cloudflare (Workers AI) · Decision model

Cloudflare-hosted decision models on Workers AI that take a state and a schema of typed questions (noul, choice, score) and return a probability for every allowed option. Three variants: Clef (27B), Clef-flash (9B), Clef-omni (30B MoE, 3B active).

Clef $0.240, Clef-omni $0.150, Clef-flash $0.038 per 1M input tokens; no output-token charge. Media are converted to input tokens.Verified 2026-10-15 on Cloudflare Workers AI pricing page (last updated Oct 9, 2026) and model pages

Verified

Perplexity Decisions API (pplx-decider)

Perplexity · Decision model

Perplexity API endpoint that answers yes/no, multiple-choice, and scored questions about text and images with probabilities.

$0.02 per 1M input tokens for both models; output tokens free; no per-request fee. State, questions, and images count as input.Verified 2026-10-15 on Perplexity pricing docs

Partly verified

Inception Mercury Decide

Inception · Decision model

Decision model served at POST /v1/decisions (model mercury-decide) that evaluates named questions against a shared state and returns typed answers with probabilities.

Input listed as $0.04 → $0.02 per 1M tokens (shown as a 50%-off promotion); cached input and output $0.Verified 2026-10-15 on Inception models page; promotion end date not stated

Verified

Liquid AI d1

Liquid AI · Decision model

Hosted decision model on the Liquid API (System One shape) returning noul, choice, and score answers with zero generated tokens; supports images. Open-weight variants also documented.

No per-token price found on the d1 docs page.Unverified — not published on the page checked 2026-10-15

Partly verified

Strands Decider 2B (open source)

Strands Agents (AWS) · Open-weight decision model

Small open-source decision model (Qwen3.5-2B-Base + LoRA adapter + option-scoring head) for local routing, tool selection, guardrails, and evals. Apache-2.0, weights on Hugging Face.

Free open weights (Apache-2.0); you pay your own compute.Verified 2026-10-15 on Hugging Face model card

Verified

AWS Bedrock Guardrails

AWS · Policy guardrail

Configurable guardrails: content filters, denied topics, PII handling, contextual grounding, and Automated Reasoning checks; returns block/allow plus assessments. Guardrails are versioned.

Priced per text unit (rates not re-checked).NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified

Azure AI Content Safety

Microsoft · Harm classifier

Harm classifiers with severity levels, Prompt Shields, and groundedness detection. Entra auth; Azure regions; API-versioned.

Priced per text record (rates not re-checked).NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified

Google Vertex AI / Gemini structured output

Google Cloud · Structured output

Constrain Gemini output to an enum or JSON schema and read logprobs; Model Armor for safety screening. Versioned model IDs.

Per-token (rates not re-checked).NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified

Anthropic Claude (tool use / structured outputs)

Anthropic · Structured output

Use tool use / structured outputs to return JSON labels. No native probabilities. Dated model snapshots; available via Anthropic, Bedrock, and Vertex.

Per-token (rates not re-checked).NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified

OpenAI Moderation (omni-moderation)

OpenAI · Harm classifier

Harm classifier returning per-category scores. Dated snapshot.

Previously recorded as free (not re-checked).NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified

Cohere Classify

Cohere · Classifier

Few-shot / fine-tuned classification returning a label plus confidences.

Not known.NOT re-verified on 2026-10-15 — carried from prior research; check before relying on it

Unverified