Benchmarks & research
Morph's Reflex is a hosted 27B decision model, based on Qwen 3.5, that answers yes/no, choice and score questions about text, JSON or images, and charges less when the same input comes back with a new question.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Input costs $0.042 per million tokens the first time and $0.008 once cached, output is free, and caching needs no flag or cache ID. Morph's example is a support ticket asked about department, urgency, anger, refunds and spam. The launch posts call Reflex "10-20x cheaper than Jev" with "similar intelligence to other leading decision models" and say it is also on OpenRouter, but Morph has published no benchmark table behind the comparison.
You can have a developer connect your app to Morph's Reflex with an access key from the service. When a support ticket arrives, the app can send the same ticket text several times, each with a new question, such as which team should handle it, how urgent it is, or whether the customer wants a refund. Morph says the repeated text is then billed at a cheaper cached rate, though caching is not guaranteed every time.