OpenAI Decisions API: 10x faster classification at $0.10 per million tokens

What changed

OpenAI announced the Decisions API, a new tool focused on fast classification tasks. The service is in public beta now, with general availability coming soon. It runs on gpt-6-luna only and costs $0.10 per million input tokens. Output tokens are free.

The Decisions API classifies text, images, or both. It handles three types of responses: yes/no probabilities, picks from predefined categories, and scale-based ratings. Use cases include damage detection in photos, automatic customer inquiry routing, and document classification, according to OpenAI.

The company says it built the API to run about ten times faster than the Responses API. It supports zero-data retention and meets HIPAA requirements for use in the US and Europe.

OpenAI also restructured its paid API tiers from five down to three: Build, Launch, and Grow. Organizations move up automatically when their total credit purchases hit the next threshold. Monthly usage limits sit at $500, $5,000, and $200,000.

Why it matters

The Decisions API addresses a real constraint in production workflows. Many applications need fast binary decisions or category assignments but do not require the full reasoning power of a general-purpose large language model. The 10x speed improvement over Responses API lets you handle higher volumes with lower latency.

For teams processing high-volume classification tasks, the pricing matters too. At $0.10 per million input tokens, the model is designed to be cost-efficient compared to running the same logic through standard API calls. Free output tokens remove a second cost layer.

The three response formats cover common production patterns: damage or compliance checks using yes/no probabilities, ticket routing or tagging using categories, and quality scoring or severity levels using scales. This scope should fit existing workflows in customer support, content moderation, damage claims, and document intake.

Data retention and HIPAA compliance open the door to regulated industries. Healthcare organizations, financial services firms, and other compliance-sensitive sectors can now use OpenAI’s classification speed without building custom infrastructure.

The tier restructuring also signals a shift in how OpenAI packages API access. Consolidating from five to three tiers simplifies onboarding and may encourage teams to stay within OpenAI’s ecosystem rather than switching providers as they scale.

What to test

Before moving Decisions API into production, validate OpenAI’s core claims against your own use cases.

Speed benchmarks

The company says the Decisions API is about ten times faster than Responses API. Measure the actual latency in your environment with real inputs and load patterns. Test both text and image inputs if both are relevant to your workflow. Latency under load matters more than a single-request average.

Accuracy on your data

Classification quality depends on domain fit. Test the API on a representative sample of your actual inputs. For yes/no decisions, check both precision and recall across your dataset. For category picks, measure confusion between similar classes. For scaled ratings, validate that boundary cases (ambiguous items) return reasonable confidence scores.

Compare results to your current classification logic, whether that is rule-based, human-reviewed, or another model. A 10x speed gain is worthless if accuracy drops enough to break downstream processes.

Confidence scoring

The API returns probabilities alongside classifications. Validate that confidence scores correlate with actual accuracy. A 90% confidence result should be wrong roughly 10% of the time across your dataset. Use confidence thresholds to flag low-confidence calls for human review if your workflow requires it.

Predefined categories constraint

The API picks from categories you define in advance. Ensure this fits your use case. If you need open-ended classification or discovery of novel categories, the Decisions API is the wrong fit. Test how the API handles inputs that do not fit neatly into your predefined set.

Cost modeling

At $0.10 per million input tokens, calculate the actual cost per classification in your workflow. Factor in token overhead from formatting and context. Compare to your current approach. Also model the impact of free output tokens if your workflow relies on detailed explanations alongside classifications.

Data retention and compliance

Verify that zero-data retention and HIPAA compliance actually meet your organization’s requirements. Test the setup in your compliance environment. If you are in a regulated industry, have legal review the terms before committing.

Tier transitions

OpenAI’s new tier thresholds ($500, $5,000, $200,000 monthly) may change your cost structure or support options as you scale. Model your expected growth and confirm the tier transitions align with your budget and operational needs.

The conclusion

The Decisions API addresses a real gap: fast, cost-efficient classification for production workflows. The 10x speed improvement over Responses API and the $0.10 per million token pricing are credible for classification-only tasks. HIPAA compliance and zero-data retention expand the addressable market to regulated industries.

The constraint is scope. The API is purpose-built for predefined classification tasks. If you need open-ended reasoning, fine-grained explanations, or novel category discovery, the Decisions API is not the right tool. The trade-off between speed and flexibility is real.

Before adopting, test the API on a representative dataset from your domain. Validate that latency and accuracy improvements justify the switch from your current approach. Measure confidence scores to understand when the API’s classifications are trustworthy and when they need human review. Model costs across your expected volume.

Watch for general availability timing and whether OpenAI expands the model options beyond gpt-6-luna. Competing decision models may also emerge. The 10x speed advantage is meaningful today, but market movement is fast.