Now available: GPT-6 Luna Decisions
Try the modelDecision models · Reviewed October 7, 2026
Give software a bounded decision it can act on.
OpenAI released Decisions API in beta on October 6. GPT-6 Luna Decisions is now available here through OpenRouter, using the same state and typed questions as Jev. The native OpenAI protocol differs; this comparison explains the integration actually offered on this platform.
This is an independent platform with no affiliation to OpenAI or TypeSafe. GPT-6 Luna Decisions is served through OpenRouter using this platform’s account, API key and credits.
Verified integration and open questions
Live requests verified text, JSON objects, arrays and image inputs through OpenRouter. These are contract checks, not a matched accuracy benchmark or a production latency SLA.
Provider input price
Jev
$0.042 / 1M
OpenAI Decisions API
$0.10 / 1M
OpenRouter lists Jev at $0.042 and GPT-6 Luna Decisions at $0.10 per million input tokens. This platform charges 600 and 1,500 credits respectively, with a one-credit minimum and rounding per successful request. Output tokens add no charge.
OpenRouter · GPT-6 Luna DecisionsLatency evidence
Jev
Test your cases
OpenAI Decisions API
Measure your workload
The three text contract probes took 304–401 ms end to end from the verification environment. They are too small a sample for p50/p95, an SLA or a fair speed comparison with Jev.
OpenRouter · GPT-6 Luna DecisionsProbability calibration
Jev
RLCD
OpenAI Decisions API
Unconfirmed
TypeSafe describes RLCD training. The verified API responses do not establish calibration or comparative accuracy for GPT-6 Luna Decisions. Measure both on labeled data before selecting thresholds.
TypeSafe explanation of RLCDA bounded decision
Define the available branches and the evidence that separates them. A decision model returns a signal your code can use. A chat model with structured outputs can also fit a schema, but a valid JSON shape alone does not establish decision quality or calibration.
BILLING_SUPPORTTECH_SUPPORTREFUNDGENERAL_SUPPORTChoose around your workload
01
Start with its published Noul, Choice, and Score contract. In the playground, test normal, ambiguous, and incomplete inputs against your own expected answers.
02
Image requests are verified through OpenRouter. This platform accepts up to 4 PNG, JPEG or WebP images under its upload limits. Use the model page for a preselected image-capable playground.
03
Compare error rates, probability calibration, cost, and p50/p95 latency with representative labeled data. Define a review threshold and a fallback in application code. A confident score does not guarantee a correct decision.
From evaluation to integration
Jev · TypeSafe System One model
TypeSafe documents POST https://api.typesafe.ai/v1/systemone. Requests carry state, model, and typed questions; responses return answers keyed by question ID.
OpenAI · OpenRouter integration
Use openai/gpt-6-luna-decisions with state and a questions object at this platform’s /v1/systemone. OpenRouter converts between its Jev-style contract and OpenAI’s native Decisions API.
Side-by-side comparison
Sources and scope · October 7, 2026
OpenAI’s official changelog and Decisions reference establish the beta launch and native schema. OpenRouter’s model and endpoint listings establish this integration’s model ID, context and prices. Live synthetic probes verify all three answer types, text/object/array state, 8 questions, option boundaries, a long-state marker and contrasting images. A simulated refusal response verifies refund handling; the live refusal probe returned a normal answer. No matched accuracy or calibration benchmark was run.