NewModelHubCalculate Your Cost — Decisions vs Responses
Questions per request does not change the price — cost is per token, not per question — but batching independent questions into one call shares the input context across them. Cache-write tokens are the tokens newly written to cache, not the tokens read back from it.
Your Estimate
Monthly cost, Decisions API: $32.00
Same workload on Responses API (output at $0.50 / 1M): $34.50
Decisions saves $2.50/month of output billing on this workload.
OpenAI Decisions API Rate Card (gpt-6-luna, per 1M tokens)
| Context | Uncached input | Cached input | Cache write 30m | Cache write 12h | Output |
|---|---|---|---|---|---|
| Short (≤ 272K input) | $0.10 | $0.01 | $0.125 | $0.25 | $0 (free) |
| Long (> 272K input) | $0.20 | $0.02 | $0.25 | $0.50 | $0 (free) |
Rates from OpenAI's official Decisions guide (developers.openai.com/api/docs/guides/decisions) and official pricing page (developers.openai.com/api/docs/pricing), verified October 25, 2026 (Asia/Shanghai). Endpoint POST /v1/decisions, model gpt-6-luna only, 1–400 questions per request (up from 200 in beta). Images: inline base64 data URLs or publicly accessible HTTP/HTTPS URLs; file_id inputs are not supported. The guide states cache-read and cache-write rates match the standard gpt-6-luna prompt caching rates. Regional processing premiums and long-context multipliers apply; data-residency endpoints carry a 10% uplift for models released March 5, 2026 and later. The Responses comparison uses the $0.50/1M output rate from the standard gpt-6-luna pricing row; long-context output pricing was not listed in the verified table.
What Drives Your Bill
- Output is free — that is the whole point. Decisions returns bounded answers, not generated text, so you never pay the $0.50/1M output rate the same model charges on the Responses API. Workloads with long answers save the most.
- Cache writes cost more than the input they cache. A 30-minute write ($0.125/1M) costs more than fresh input ($0.10/1M); a 12-hour write ($0.25/1M) costs 2.5×. A cached entry only pays for itself once it is read back: about 2 reuses for 30-minute entries, about 3 for 12-hour ones, at the $0.09/1M saving per cached token.
- Long context doubles input, cached, and write rates. Cross 272K input tokens and every input-side rate doubles. Keep requests under the threshold where you can, or budget the 2× explicitly.
- Batch independent questions into one call. Cost is per token, not per question, and one request holds up to 400 questions — shared context is billed once, not once per question.
- Watch the regional uplift. Data-residency endpoints add 10% for models released March 5, 2026 and later — small per call, visible at a million requests a month.
Related tools
FAQ
What does the OpenAI Decisions API cost?
Decisions bills input only on gpt-6-luna: $0.10 per 1M uncached input tokens and $0.01 per 1M cached input tokens (Short context, up to 272K input tokens). Output tokens are not charged. Cache writes are charged at $0.125 per 1M (30-minute TTL) or $0.25 per 1M (12-hour TTL). Long context over 272K input tokens doubles the input, cached, and write rates.
What is the Decisions API for?
Bounded answers software can act on: predicate questions return a probability (0 to 1) that a condition holds; choice questions return one of your preset values with probabilities and confidence; score questions grade input against a 2-to-10-level rubric and return a probability-weighted average. It is for answers, not open-ended generation.
What are the Decisions API limits?
One request holds 1 to 400 questions (up from 200 in beta). Only gpt-6-luna is supported, via POST /v1/decisions. Images may be inline base64 data URLs or publicly accessible HTTP/HTTPS URLs; file_id inputs are not supported.
How much does Decisions save versus the Responses API?
The same model on the Responses API charges $0.50 per 1M output tokens, while Decisions charges nothing for output. The saving equals your output tokens times $0.50 per 1M — plus the generation step a chat-style call would need to produce a structured answer.
Is this an official OpenAI tool?
No. It is an unofficial, fan-made estimator. OpenAI's invoice and official docs are authoritative. Regional processing premiums and long-context multipliers apply on top of the base rates shown here.
Methodology & sources: rates verified against OpenAI's official Decisions guide and pricing page on October 25, 2026 (Asia/Shanghai) — https://developers.openai.com/api/docs/guides/decisions and https://developers.openai.com/api/docs/pricing. Estimates exclude taxes and reseller markups. Your invoice follows metered usage.