OpenAI Decisions API GA with GPT-6 Luna: Fast Bounded Choices at $0.10/M Input Tokens

OpenAI Decisions API GA with GPT-6 Luna: Fast Bounded Choices at $0.10/M Input Tokens

Publication date: 2026-10-26 | Language: English | Audience: platform engineers, agent architects, FinOps leads, product owners, and SRE teams placing classification and routing in the hot path of apps and agents.

Disclosure: This article synthesizes OpenAI’s public Build more responsive apps with the Decisions API announcement (October 22, 2026), OpenAI developer docs for Decisions, and customer quotes published by OpenAI. It is not investment, legal, or security audit advice. Validate current pricing, residency, Zero Data Retention eligibility, and SLA language in your OpenAI contract before production cutover.

Primary sources:

• Build more responsive apps with the Decisions API — OpenAI (October 22, 2026)
• Decisions guide — OpenAI API docs
• OpenAI API changelog — Oct 22, 2026 Decisions GA

For always-on agent shells that consume fast routers, see our OpenAI Dots architecture deep dive and the October 2026 agent wave overview. This piece owns Decisions API product surface, GA deltas, pricing, and hot-path routing patterns only.

Why Decisions matters more than another chat endpoint

Most agent stacks burn frontier tokens on questions that are not open-ended essays. They ask: Is this document a certificate of insurance? Should we continue, retry, or escalate? How severe is this finding on a 1–5 rubric? Individually those calls look cheap. At thousands per minute, latency and output tokens dominate both user experience and the invoice.

OpenAI’s answer, generally available on October 22, 2026, is a dedicated Decisions API on gpt-6-luna: the model focuses on bounded answer types—probability, choice, or score—and OpenAI reports up to ~10× faster decisions than GPT‑6 Luna through the Responses API. Pricing is deliberately asymmetric: $0.10 per million input tokens, $0.01/M for cached input, and no output-token charge. That packaging is a FinOps signal as much as a product one: OpenAI wants structured micro-decisions off the expensive generation path.

This article delivers a fact layer from OpenAI primary sources, systems analysis for agent graphs and voice, 0–3 / 3–12 month forecasts with falsifiers, and role checklists—aligned with our inference economics and model-routing FinOps playbooks.

The October 22, 2026 fact layer (OpenAI)

Product definition

The Decisions API evaluates text, images, and native JSON and returns typed answers. Developers define a question and its allowed answers, then supply context. Supported answer types:

Endpoint shape (docs): POST /v1/decisions with model gpt-6-luna (currently the only supported Decisions model per OpenAI docs).

GA deltas since public beta

OpenAI lists these upgrades since the October 6 beta:

• Native GPT‑Live integration for in-conversation decisions
• Prompt caching for reused context (cached input at $0.01/M)
• Inference residency in the US and EU
• Native JSON inputs (input_json)
• Images inside individual question instructions
• Up to 400 questions per request (up from 200 in beta)
• Continued latency work, including capacity in several Asia regions

Pricing and data controls (as published)

• Input: $0.10 per million tokens
• Cached input: $0.01 per million tokens (see OpenAI pricing guide for 30-minute / 12-hour write details)
• Output tokens: not charged
• Zero Data Retention: available to eligible customers
• Regional processing: US and EU

Treat list prices as a snapshot; always re-check OpenAI’s live pricing page before budgeting.

Customer evidence OpenAI published

OpenAI’s GA post includes named-customer claims (their evaluations, not independent audits):

• Box: ~5× faster route selection on 30 certificate-of-insurance cases vs Responses, same accuracy in their eval
• Datadog: ~40× faster p95 risk scoring of agent-generated remediation commands vs available frontier models in their comparison
• Ramp: 4× faster and ~30% cheaper AI cost attribution vs GPT‑6 Luna Responses path
• Sierra: under 280 ms p95, about 3× faster and ~90% lower cost per call than prior approach
• Speak: GPT‑Live + Decisions cut time-to-next-action from 1.4s → 0.13s (~90%) in their testing

Use these as directional signals for pilot design, not as your SLA.

Systems analysis: where Decisions sits in an agent graph

Hot-path vs cold-path

A practical split for late-2026 stacks:

1. Hot path (Decisions): classify, route, gate, score risk, pick next tool, choose avatar reaction, triage inbox themes.
2. Warm path (mid-tier Responses / Claude Sonnet / Gemini): draft, summarize, short tool loops.
3. Cold path (frontier / long-horizon agents): multi-hour research, large code migrations, ambiguous strategy.

OpenAI’s own Dots team reportedly uses Decisions to cut interactivity latency—see our Dots deep dive. That is the pattern: bounded brain in front of generative brain.

Voice and live sessions

GA adds native GPT‑Live hooks: transcript deltas can trigger Decisions (e.g., choose an avatar expression from {neutral, happy, sad, concerned}). For support and tutoring products, this is the difference between “smart after the turn” and “smart during the turn.”

Failure modes to design for

• Over-bounded questions: forcing a choice set that cannot express “insufficient context” creates false confidence—always include clarify / abstain options when stakes are high.
• Batching abuse: 400 questions/request is a throughput gift and a prompt-injection surface—validate question templates server-side.
• Cache stale context: prompt caching saves money; it can also serve outdated policy text—version your shared context blobs.
• Residency mismatch: US/EU residency does not automatically satisfy every industry control framework—map to your DPA.

Forecast scenarios (with falsifiers)

0–3 months (through January 2027)

Base case: Decisions becomes the default OpenAI SKU for classification and agent routing in new builds; many teams dual-write Responses → Decisions for A/B latency.

Falsifier: OpenAI adds competing “fast Luna” Responses modes that erase the 10× gap, or GA reliability issues push enterprises back to self-hosted classifiers.

3–12 months

Base case: Multi-vendor “decision APIs” (OpenAI + Anthropic classifiers + open routers) appear in FinOps dashboards as a first-class cost line; voice agents standardize on sub-300 ms routing.

Falsifier: Output-free pricing proves unsustainable and OpenAI reintroduces output fees that collapse the ROI story.

Role checklists

Platform / agent architects

• Inventory every Responses call that returns an enum, boolean, or score.
• Prototype the top 5 on Decisions with identical golden sets.
• Add abstain / escalate choices to every high-stakes rubric.
• Wire GPT‑Live triggers only after offline transcript replay tests.

FinOps

• Model cost at $0.10/M input with cache hit-rate scenarios (10% / 50% / 80%).
• Compare against current Responses spend on the same traffic.
• Cap Decisions QPS per product so a runaway agent cannot burn the monthly commitment.

Security / compliance

• Confirm Zero Data Retention eligibility before sending regulated payloads.
• Prefer US/EU residency endpoints when contracts require them.
• Log question templates and choice sets as versioned config (audit trail).

Product

• Start with one recurring decision users feel as lag (route, escalate, next lesson).
• Measure p95 decision latency and wrong-route rate, not just token cost.

Risks and misconceptions

• Misconception: “Decisions replaces GPT‑6 for everything.” It does not—open-ended generation still belongs on Responses / Agents paths.
• Misconception: “No output charge means free.” Input tokens, cache writes, and request volume still scale with traffic.
• Risk: Treating vendor customer quotes as your latency budget without measuring your own payload sizes and region.

Bottom line

The October 22, 2026 Decisions API GA is OpenAI’s clearest bet that agents need a specialized micro-decision lane, not just a smarter chat model. If your graph still calls a full Responses completion to pick among three routes, you are leaving both latency and margin on the table—subject to your own evals on accuracy, residency, and cache design.

Related reading on WordOK: OpenAI Dots · GPT-6 Intelligent UI / agent wave · Inference cost caps and model routing