Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Inputs can contain text, images, or both. Answers use predicate, choice, or score.
- Prepare model, input, and questions separately. Responses return an answers array preserving each question's name.
- Independent questions can share one input. Questions depending on earlier answers need separate requests.
- OpenAI's approximately 10-times speed comparison is a vendor claim. Confirm latency and classification quality with your own samples.
- This is a public beta. Check pricing, regional-processing surcharges, and project limits separately; beta does not mean free or unlimited.
What Decisions API does, and when to choose it
Support systems may first classify a message as payment, shipping, or technical support; agents may route the next step to a specific workflow. Decisions API expresses these as explicit questions and returns program-readable answers. The guide accepts text, images, or both, either as a string or a user message containing text and images.
It suits classification, routing, and prioritization. For extracting custom fields or generating explanations, the guide recommends Responses API Structured Outputs. For proposed tool calls with parameters, use function calling. Decide which result you need before selecting an interface.
Three answers: Probability, categories, and ordered scores
predicate evaluates a condition and returns probability from 0 to 1, such as whether a product photo shows damage. choice selects a value from supplied choices, suitable for unordered categories such as support queues or product types. Give each option a clear description to reduce overlapping labels.
score uses your ordered levels. Indices start at 0, and the result is a probability-weighted average of level indices, so it may fall between levels. Both choice and score return option distributions and separate confidence. Use choice for a single category rather than treating a continuous score as a category ID.
Start with the model, endpoint, and SDK
Organize questions and samples in the official Decisions Playground, then configure server-side calls. The dedicated endpoint is https://api.openai.com/v1/decisions using POST, Authorization: Bearer with an API key, and Content-Type: application/json. Set model to gpt-6-luna; the guide lists no other supported model at this check.
The body contains model, input, and questions. Each question needs a unique name, type, and instructions; choice adds choices and score adds levels. Official examples require OpenAI Python SDK 3.26.0 or JavaScript SDK 7.30.0 or later. If an older SDK lacks decisions, check version and endpoint instead of changing only the model name.
Organizing a minimal support-routing request
A redacted message can establish the smallest request. This illustrative request follows documented fields and has not been executed: {"model":"gpt-6-luna","input":"The order is paid, but I cannot find tracking information.","questions":[{"type":"choice","name":"support_queue","instructions":"Select the support queue best suited to this message.","choices":[{"value":"payment","description":"Payment failure, duplicate charges, or billing issues."},{"value":"shipping","description":"Dispatch, delivery, or tracking questions."},{"value":"other","description":"Insufficient information or outside the categories above."}]}]}.
Have support staff inspect results before automatically transferring cases. Prepare positive examples, counterexamples, and insufficient-information samples for each queue, then check whether payment and shipping are separated according to your business definitions. These queues illustrate integration; they are neither OpenAI's ecommerce standard nor this site's measured results.
Product images and multiple questions
The official image example combines input_text and input_image in a user message's content, passing a Base64 data URL through image_url. Check the current format; an image-path string does not mean the image was uploaded. Define the object and exclusions clearly, such as distinguishing product damage from packaging shadows.
Independent damage and category questions can share a questions array and input while using different answer types. If the repair-category question depends on confirmed damage, obtain the first answer before requesting the next. Multiple questions in one request are not a sequential workflow.
Read answers and provide a path for uncertainty
The response's answers array preserves question names. Locate results by name and type, then read probability, choice, or score. SDK examples separately handle refusal. For refusals, missing expected answers, or failed requests, use human review or the existing workflow instead of interpreting an empty result as approval.
Probability and confidence are not business accuracy. Use labeled business samples to determine thresholds and the costs of false positives and negatives. Compare with human labels, record category errors, escalation rate, and latency, then decide which results may route automatically. Refunds and penalties need independent business rules and permissions.
Confirmed costs and beta conditions
As of October 7, 2026, the guide prices gpt-6-luna input on /v1/decisions at $0.10 per million tokens. Only input tokens are charged; cache reads, cache writes, and output tokens are not. Regional-processing surcharges and long-context input multipliers still apply. This endpoint's rate does not automatically apply to other interfaces using the same model.
The service is public beta. General availability is expected in the coming weeks without a fixed date. Eligible customers can use ZDR and HIPAA support; residency and regional processing cover the US and Europe's EEA and Switzerland, subject to separate eligibility and agreements. This guide has not verified a personal project's rate limits, balance, or plan, nor made paid calls. Budget from the project console and applicable rates.
Measure speed and diagnose networking from your own records
The guide describes Decisions API as about 10 times faster than Responses API; the developer announcement compares with GPT-6 Luna called through Responses API. This is a vendor comparison, not a guaranteed improvement for every message, image, or region. Record response time, quality, and input usage for identical samples. This page provides no independent performance test or savings guarantee.
When no HTTP response arrives, investigate DNS, TLS, timeouts, and the client's network path. When a response arrives, inspect its actual error and authentication, model, and quota settings first. Fixed egress can reproduce network conditions but does not add API entitlements or make outputs trustworthy. Question design, access, and connectivity require separate checks.
Sources
Frequently Asked Questions
Is Decisions API generally available?
As of October 7, 2026 it is public beta, announced October 6. The guide expects general availability in the coming weeks without a definite date.
Can I use GPT-6 Astra or another model?
The current guide lists only gpt-6-luna. Wait for official interface updates rather than importing the ordinary chat model list.
Which type fits categories, damage checks, and support scores?
Use choice for unordered categories, predicate for visible damage, and score for explicitly ordered severity levels. Define labels and criteria for your business.
Can one request check damage before choosing a repair category?
Independent questions can share an input. Dependent questions need separate requests, with the application choosing the next question from the first answer.
Does high confidence authorize a refund?
No. Confidence is neither business accuracy nor action permission. Determine thresholds and fallbacks with labeled samples; refunds still need independent rules and authorization.
Is beta free, and is 10-times speed guaranteed?
The guide specifies input-token charges and surcharge conditions, so it is not free or unlimited. Approximately 10 times is OpenAI's comparison; this page made no calls, billing measurements, or independent performance tests.