Strategy

The Decision Layer Gets a Price List: Four SKUs in One Week, from $0.10 per Million Tokens

By IdeaBosqueLast updated: November 14, 2026

Key takeaways

Four decision-model SKUs shipped in a single week (Nov 11–13, 2026) from Google, Perplexity, HeyGen, and Hark — a category that ten days earlier consisted of 21 models from 13 makers listed behind OpenRouter's Decisions API without a single first-party price anchor.
Gemini Decision 1 prices bounded choices at $0.10 per 1M input tokens with $0 output billed, scoring up to 1,000 yes/no, choice, and rubric questions in one pass at 150–250 ms end-to-end — nearly 5x cheaper across Google's shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
Perplexity Decider v1.3 (26B, Qwen3.8-27B backbone) and Flash v1.3 (5B) are the category's first shipped weights, released under Apache-2.0 on Hugging Face — the self-host path now exists for the layer that routes, watches, and grades inside an agent loop.
The honesty markers come from the vendors themselves: HeyGen's docs describe confidence as "a property of the returned distribution, not a calibrated probability," and Perplexity tells users to tune decision thresholds on their own labeled traffic rather than trust the shipped temperature.
The policy is not on the price list: which decisions may defer to a classifier, and what happens when the cheap watcher is wrong, remain architecture decisions — the five-decision framework's third and second rows, not the vendor's.

This article builds on Five Decisions That Determine Whether Your AI Agent Ships, which ordered the architectural decisions behind production agents. Its third decision — which model does what — gained a catalog on October 16, when OpenRouter listed 21 decision models from 13 makers behind one Decisions API. In the week of November 11–13 the catalog crossed a line that catalog presence does not cross: three vendors shipped first-party priced SKUs and a fourth productized the handoff next to them. The decide-then-act layer is no longer a pattern to assemble from parts. It is a purchase.

What a decision model is, in one paragraph

Perplexity's definition is the cleanest: a decision model is a general-purpose classifier. You give it an input — a state, plus optional images — and a set of named questions, each with its possible answers, and it returns a probability distribution for each question in a single forward pass. No token-by-token answer generation, no free-form text, no 30-second reasoning chain deciding whether a support ticket goes to billing or integration. The question types are the same across vendors: yes/no (a probability of true), multiple choice (2–255 options with a normalized probability for each), and ordered rubric scoring. Perplexity credits the lineage directly: Decider is inspired by TypeSafe AI's Jev, the non-LLM System One model we analyzed in TypeSafe Jev and the Third Layer of the Agent Model Stack. When the second entrant in the category names the first observer's pattern, the category has a shared idea, not just a shared marketing word.

The four SKUs

The week's supply, compared on the dimensions a buyer prices:

SKU	Maker, date	Access	Scale	Speed and cost	Honesty marker
Gemini Decision 1 + Decisions API	Google, Nov 11	API (OpenRouter, LangChain, LiteLLM, Vercel integrations)	Up to 1,000 questions per call over text, images, audio, video, PDFs (256K-token input)	150–250 ms end-to-end; flat 1→100 questions; $0.10/1M input, $0 output	~5x cheaper claim is against Google's own shared benchmark suite
Decider v1.3 / Flash v1.3	Perplexity, Nov 12	Apache-2.0 weights, MaaS API, OpenRouter	26B (on the Qwen3.8-27B backbone) and a 5B distillation; 1–255 options per question	Fast path 1.7–1.9x faster (separate mode), 9–18x (isolated mode)	Threshold honesty: tune on your own traffic; weights repo is access-gated despite the Apache license
HeyGen Decision 1 (+ Turbo preview)	HeyGen, Nov 13	Synchronous API	1–400 named questions; up to 4 images + 1 video (Turbo preview: 100 images, 10 videos)	Synchronous response, `output_tokens: 0` always; no price published on the docs page	Docs: confidence is "a property of the returned distribution, not a calibrated probability"
Hark Handoff²	Hark, Nov 12	Hark Pro (computer-use agent)	82% on an internal 478-task, 238-site browsing benchmark, up from 73%	46% faster model, ~2x end-to-end	Vendor-reported internal benchmark, not an open eval

The depth of the category shows in what each vendor bet on. Google bet on breadth and batching: one multimodal input, up to 1,000 questions, latency nearly flat from 1 to 100 questions per request, input billed once regardless of question count. Perplexity bet on ownership: shipped weights mean the routing, moderation, and grading layer can live on hardware the buyer controls, with a new locate type that extracts verbatim spans with character offsets — classification and extraction in one request. HeyGen bet on the media path: text, JSON, four images, and one video of sampled frames in one synchronous call, with a Turbo preview pushing to 100 images and 10 videos. Hark bet on the adjacent primitive: handoff — the moment a browser agent's work becomes a human's work — productized and measured 2x faster against its own v1. Google's internal deployments hint at where this lands in production: Gemini Notebook uses the Decisions API as a real-time model router, reporting up to 40% lower latency on everyday queries, and Google Search uses it as a safety layer to detect adversarial behavior against MCP tools.

Four decision-model SKUs compared on access, scale, speed, and the honesty marker each vendor publishes against itself. The pattern a buyer should hold onto: every speed and price row is the vendor's own measurement.

Why bounded choice is priced like this

The economics are structural, not promotional. A generative call pays for output tokens, and a routing or moderation decision pays token by token to produce the least useful artifact in the pipeline: free-form prose that ends in one of three words. A decision model replaces that with a classification read — a trained linear head over predefined option tokens — so the Decisions API bills $0 for output, bills shared input once per request no matter how many questions ride on it, and skips variable thinking-token charges entirely. Perplexity's architecture notes explain the second reason: the 1.3 models read bidirectionally and answer each question from a shared state in one pass, and the HF card documents a fast path that packs many sequences per forward pass, 1.7–1.9x faster in separate mode and 9–18x faster in isolated mode. For the self-hosted buyer, the arithmetic is concrete: 49 GiB of BF16 weights fits the memory envelope of one 80 GB GPU, and the Ornn Compute Price Index settled an H100 SXM at $3.29 per GPU-hour on November 13 — a routing, grading, and moderation layer on an hourly rate that a single frontier-model generation call was already exceeding every day. That is the price collapse this week is really about: the five vectors of inference cost management extended to the decisions around the generative calls.

What the price list does not include

Four things the SKUs do not sell, each visible in the vendors' own documentation.

Confidence is not calibration. HeyGen's docs are the most explicit about this in the industry: confidence is "a property of the returned distribution, not a calibrated probability." Google ships a 0–1 confidence score on every answer; that tells you where the distributor ends, not where your correctness threshold sits. Thresholds are your data's problem. Perplexity's Getting Started section states it plainly — label frequencies and the costs of false positives versus false negatives vary across applications, so tune thresholds on a small, representative labeled development set. The same honesty appears on the HF card: "a single temperature cannot be right for every task, and per-task thresholds are usually better." Questions can't depend on each other. HeyGen's API documentation notes that one question's answer does not become input to another in the same request — dependent decisions need sequential calls, which is where latency stops looking free. Perplexity documents the same boundary from the other side: in the default separate mode questions never see each other, and the shared-state isolated mode trades that independence for cost, shifting 1.5–2.2% of top answers. Latency and context claims have scopes. Google's 150–250 ms holds across contexts up to 16K tokens — the 256K-token multimodal input is a different envelope — and Hark's 2x is a vendor-reported figure against its own v1. None of these caveats argues against the category. They argue for piloting with your own labeled traffic and treating confidence values as untuned promises until your thresholds say otherwise — the same honesty discipline we apply to GPU price prints, where the index itself repriced between reads.

What stays yours: the policy layer

A decision model can absorb your classification, routing, watching, and grading calls. It cannot absorb the two decisions that make them safe. First: which decisions may defer to a classifier at all — a support-ticket queue choice can route itself; a supplier price write-back to NetSuite cannot. Second: what happens when the cheap watcher is wrong — Google's own design answer is routing ambiguous cases on confidence to a fallback or human reviewer, but which cases are ambiguous, and how often the human reviews, is your escalation policy. In the five-decision framework this is Decision 3 extended downward: routing, watching, and kill-switch evaluation are becoming purchasable components instead of custom code, while the loop layer that enforces judgment in the runtime keeps the approvals it always owned. What the price list changes is the cost of the watch layer. What it does not change is who signs for it.

A representative build

A mid-market industrial distributor running NetSuite, BigCommerce, and three supplier catalogs wanted its quoting agent to work around the clock without a human reading every draft. The decision layer structured the split: a frontier model plans the RFQ response and drafts the quote; a decision-model tier watches everything else. Every inbound RFQ gets graded for completeness and risk in one request before a generative call is spent; supplier responses get classified against the quote's line items; each drafted quote gets a confidence-scored policy check, with anything under threshold routed to a human desk instead of NetSuite. The watch layer costs $0.10 per million tokens on Google's priced SKU — the kind of bounded-choice bill that replaces roughly five Flash-Lite generation calls per decision by Google's own benchmark comparison — and runs at a latency the human reviewer never waits for. Quote cycle time dropped from three days to under four hours with a human approving every write; the outcome is the desk's capacity, not its headcount.

That is the pattern: buy the bounded choices, price the confidence honestly, keep the policy.