OpenAI Decisions API: Cheap, Fast Batch Judgments for Agents and Pipelines
A shocking share of "LLM" work is not writing prose—it is answering thousands of tiny questions: does this lead match our ICP, is this ticket urgent, which category is this SKU? OpenAI's Decisions API, generally available around October 23, 2026, gives that workload a purpose-built endpoint: predicates, choices, and scores, up to 400 questions per request, roughly 10x Responses throughput at about $0.10/1M input tokens.
Three primitives: predicate, choice, score
- Predicates — boolean judgments: "Does this company sell to developers?" Yes or no, at scale.
- Choices — single-select classification: pick the intent, the category, the owner.
- Scores — numeric rating: fit 0–100, urgency 1–5, spam likelihood.
Pack up to 400 of these into one request and gpt-6-luna returns a structured answer set. No chat formatting, no reasoning tokens burned on essays—just the verdicts your pipeline needs, per developers.openai.com.
Decisions vs Responses: the economics
| Dimension | Decisions API | Responses API |
|---|---|---|
| Best at | Batch predicates/choices/scores | Reasoning, generation, tool use |
| Batch size | Up to 400 questions/request | One task per call |
| Throughput | Up to ~10x faster | Baseline |
| Input cost | ~$0.10 / 1M tokens | Model-tier dependent, higher |
| Output | Compact structured verdicts | Free-form content |
Figures reflect launch documentation and coverage—verify current limits on OpenAI's docs before committing architecture.
Where teams actually use it
- ICP filtering — score 50k accounts against your ideal-customer definition before anyone writes an email.
- Support triage — route tickets by urgency and product area on arrival.
- Content moderation — predicate checks over user-generated text and listings.
- Ecommerce enrichment — categorize and attribute SKUs from messy supplier feeds.
- Agent gates — cheap go/no-go checks inside multi-step agent loops before spending reasoning tokens.
Migration sanity checks
Moving a classifier off prompt-hacked Responses onto Decisions is usually a win, but keep three checks: (1) run both in parallel on a labeled sample and compare agreement, (2) watch for boundary drift on scores—calibrate thresholds after migration, and (3) keep a fallback path if a judgment task starts needing explanations, because Decisions optimizes for verdicts, not rationale.
Related on AIXList
- Context Engine — persist what agents (and judgment pipelines) learn.
- Shared agent context workspace — DIY patterns around multi-agent memory.
- Proactive personal agents — where cheap judgment gates matter for overnight work.
Decisions API FAQ
What is the OpenAI Decisions API?
The Decisions API is OpenAI's endpoint for batch judgment tasks—predicates (yes/no), choices (pick one of N), and scores (rate on a scale)—answered for up to 400 questions per request at up to ~10x the throughput of the Responses API, at about $0.10 per 1M input tokens.
Which model powers Decisions?
Decisions is served by gpt-6-luna per OpenAI's developer documentation—a smaller, judgment-tuned model in the GPT-6 family optimized for classification-style outputs rather than open-ended generation.
When should I use Decisions instead of the Responses API?
Use Decisions when you need many small, structured judgments: filtering leads against an ICP, triaging tickets, moderating content, picking categories. Use Responses when you need reasoning, composition, or multi-step tool use.
How many questions fit in one Decisions request?
Up to 400 questions per request at launch. Structure them as predicates, choices, or scores so the model returns a compact, machine-readable answer set.
How much does the Decisions API cost?
Launch documentation listed about $0.10 per 1M input tokens—far below reasoning-tier pricing. Confirm current rates on developers.openai.com before budgeting.
AIXList is an independent AI tools directory. This article summarizes OpenAI's public developer documentation and October 2026 coverage; we are not affiliated with OpenAI. Pricing, limits, and model routing change—verify on developers.openai.com before building.