OpenAI Launches the Decisions API Powered by GPT-6 Luna

On October 22, 2026, OpenAI officially launched the Decisions API to the broader public. Simultaneously, the company transitioned this GPT-6 Luna-powered endpoint out of its public beta phase. This newly available service costs a mere $0.10 per million input tokens. As detailed in the official OpenAI Decisions API announcement, the interface delivers strictly delineated and formatted responses. Specifically, it accurately returns probabilities, multiple-choice selections, or tiered scoring evaluations. Consequently, applications can seamlessly classify content, route inquiries, and prioritize critical tasks.

Furthermore, OpenAI highlighted that this API channels the immense capabilities of Luna toward heavily constrained inquiries. This focused approach facilitates near real-time decision-making from text, imagery, and structured data. Impressively, this endpoint generates answers up to ten times faster than querying GPT-6 Luna through the standard Responses API.

The Evolution of GPT-6 Luna and Sol

Previously, on September 22, 2026, the tech giant unveiled GPT-6 Luna alongside GPT-6 Sol. At that time, they priced Luna at $0.10 per million input tokens and $0.50 per million output tokens. Remarkably, the company slashed the API pricing for both models by 50% compared to the promotional rates of GPT-5.6. They explicitly described this dynamic duo as a monumental leap in cost efficiency. Meanwhile, the organization continues to champion GPT-6 Astra as its unparalleled flagship model across all performance metrics. Within the API ecosystem, developers can access these two efficient iterations under the standard designations gpt-6-sol and gpt-6-luna.

How the Decisions API Operates

Developers seamlessly send requests directly to the POST /v1/decisions endpoint. According to the developer documentation, gpt-6-luna currently stands as the sole supported model. Each API request elegantly merges shared input data with a set ranging from one to 400 distinct questions. This shared data can easily encompass a text string or an ordered array containing messages, tool calls, and tool outputs. Ultimately, the response delivers an answers array, meticulously repeating the developer-assigned nomenclature for every single question.

Supported Question Types

The API perfectly accommodates three specific question formats. First, predicate questions yield a precise probability falling between zero and one. This indicates whether a specific condition holds true, such as accurately detecting visible damage in a product photograph. Second, choice questions select an optimal path from a predefined array of developer-crafted options. During this process, the system returns the corresponding probability of each option alongside a holistic confidence metric.

Finally, scoring questions logically evaluate the provided input against a sequential scale of two to ten criteria levels. Subsequently, the engine rapidly computes a weighted average probability across these respective level indices. Therefore, the final calculated score might elegantly rest between two distinct integer levels.

Data Input Modalities and Limitations

Engineers can effortlessly inject structured data directly through the input_json body within user messages. Moreover, they can supply visual assets as embedded base64 data URLs or publicly accessible HTTP/HTTPS links. Because the system universally accepts conversation histories as input, the API can brilliantly evaluate requests contextually. This holistic perspective encompasses function calls, customized tool invocations, and their corresponding results. However, the documentation explicitly states that the API completely ignores files, audio clips, and item references. Additionally, this endpoint strictly rejects any file_id inputs.

For complex decisions that rely heavily on preceding answers, the official guidelines instruct developers to dispatch entirely separate requests. The documentation also provides a strategic technical recommendation for modern application builders. When a project requires a bespoke JSON object or a model-driven tool call, developers should deliberately utilize Structured Outputs. Alternatively, they can leverage function calling within the Responses API instead of the three rigidly defined answer types.

Enhancements Since the Public Beta

Following the initial public testing phase, OpenAI integrated several remarkably powerful new features. They proudly added native GPT-Live integration and prompt caching for effortless context reuse. Additionally, they introduced highly requested specialized data residency options for the United States and the European Union. Simultaneously, they significantly bolstered inference computing capacity across several key Asian regions.

Now, skilled programmers can seamlessly transmit native JSON inputs without hassle. They can seamlessly embed images directly into their foundational question instructions. Furthermore, they can submit an impressive maximum of 400 questions per request, doubling the previous restrictive limitation.

Advanced Features and Organization Tools

The robust roster of recent documentation updates highlights several other crucial capabilities. It now fully includes HTTP and HTTPS-hosted image data processing. The system possesses the profound ability to dynamically assess conversation histories alongside tool calls and their subsequent outputs. Moreover, the platform introduced comprehensive organizational usage reporting alongside highly robust Private Link support.

The specialized GET /v1/organization/usage/decisions usage endpoint absolutely necessitates an administrator API key endowed with api.usage.read permissions. Upon authentication, it promptly reveals detailed token consumption metrics and total request tallies. Administrators can conveniently filter or logically group this vital data by project, API key, or AI model.

Prompt Caching and GPT-Live Integration

Thanks to prompt caching, repetitive server requests can intelligently recycle a matching prefix of the shared input data. This prefix must contain a minimum of 1,024 tokens to ultimately qualify for a substantially discounted cache-read billing rate. Naturally, the actual question content, including strictly question-specific imagery, falls completely outside this reusable prefix boundary. By default, the API establishes an implicit memory cache breakpoint at the very end of the shared input data. Alternatively, developers can actively partition their reusable data utilizing a unique caller-provided prompt_cache_key. They may also confidently transition to explicitly marked breakpoints for superior granular control.

With the seamless GPT-Live integration, modern voice applications can dynamically execute crucial decisions upon receiving input or output transcripts. Meanwhile, the Live system continuously emits distinct decision events for the core application to process organically. An insightful example configuration from OpenAI strategically resides within session.triggers. This clever setup dynamically alters an avatar’s facial expressions based entirely on conversational transcript updates.

The company enthusiastically described a theoretical customer support agent utilizing this exact methodology. This agent autonomously decides whether to immediately ask a clarifying question or gracefully route a highly complex request. It can also instantly choose to escalate a critical issue directly to human management.

Real-World Application and Early Reviews

Internal OpenAI teams have actively integrated this highly powerful API into their everyday product and business workflows. The “dots” team leverages the API to dramatically shrink response times and significantly boost overall user interactivity. Separately, the Personal Finance division employs the API to rapidly categorize intricate transaction descriptions. Furthermore, the robust Communications team utilizes the technology to systematically categorize news articles by overarching global themes. They also actively score whether a particular story effectively reinforces or contradicts prevailing mainstream narratives. Consequently, this sophisticated process clearly identifies which content deeply requires rigorous human scrutiny.

External Client Success Stories

Several prominent clients cited in the official announcement reported extraordinary speed enhancements and profound cost reductions during independent evaluations. Biswaroop Palit, Vice President of Platform Product Management at Box, shared a deeply compelling operational anecdote. He noted that the Decisions API works brilliantly in tandem with Box Extract. This combination effectively accelerated routing selection by roughly a factor of five across thirty distinct insurance certificate cases. It achieved this remarkable operational speed while maintaining strict accuracy parity with the standard Responses API. Naturally, the system intelligently forwards any lower-confidence edge cases directly to a seasoned human reviewer.

AJ Stuyvenberg, a Senior Engineer at Datadog, actively highlighted another truly fascinating technical implementation. He stated that the API successfully scores the inherent operational risk of agent-generated rshell commands right before their execution. Impressively, it achieved a p95 processing latency forty times faster than other leading state-of-the-art models without sacrificing operational accuracy. René Sultan, a distinguished Research Engineer at Ramp, enthusiastically added further glowing praise. He declared that the API correctly isolates vital AI costs tied to corresponding workloads. It achieves this monumental task four times faster and 30% cheaper than a standalone GPT-6 Luna implementation, completely preserving original accuracy.

Ajeet Grewal, an esteemed Engineer at Sierra, detailed how the API optimally selects the absolute best next action during direct customer interactions. This occurs specifically within several intricate Sierra orchestration nodes, actively achieving a blistering p95 latency strictly under 280 milliseconds. This highly innovative approach operates approximately three times faster than their legacy methodology. It also brilliantly slashes per-call computing costs by 90% while steadfastly maintaining superb customer outcomes. Finally, Andrew Hsu, Chief Technology Officer at Speak, effusively praised the profound impact of the new GPT-Live integration. During their rigorous internal testing, the API decisively slashed next-action execution times by an astonishing 90%. The processing latency plummeted incredibly rapidly from 1.4 seconds to a mere 0.13 seconds.

Pricing Structures and Data Control Mechanisms

Financially, the robust Decisions API costs a highly competitive $0.10 per million standard input tokens. Additionally, intelligently cached input tokens cost a mere $0.01 per million. Crucially, OpenAI exclusively charges for input tokens while providing completely free output tokens. However, standard regional processing surcharges and specialized pricing multipliers for extensive context inputs definitely still apply. The respective cache read and write fees precisely mirror the standard rates associated with the gpt-6-luna prompt caching feature.

Furthermore, the expansive platform fully supports localized regional processing within the United States and the European Economic Area, including Switzerland. Dedicated enterprise data residency services remain universally available across all officially supported global API regions. Qualifying corporate customers can easily activate robust Zero Data Retention policies alongside powerful HIPAA compliance frameworks. Programmers can directly access the powerful API through the dedicated Decisions endpoint. Finally, they can actively leverage the interactive Decisions Playground situated directly on the OpenAI platform. There, developers can thoroughly test diverse operational questions against their own proprietary text or imagery before writing a single line of code.