Sometimes the model does not need to write a reply. Sometimes it needs to pick the right lane, score the right ticket, flag the risky one, or route the good one to the right queue. That is what a decision model does. Fikra serves three of them on the same OpenAI-compatible key, with the same billing and the same production limits.
Google's decision model, exposed as a native Google Decisions API. Best for mixed yes/no, choice and score questions on the same input.
$0.12 / 1M input · $0 output · context 1.05M
A scored choice on Qwen3.5-9B. Optimised for fast evaluation of support routing, content moderation, and agent stop decisions.
$0.042 / 1M input · $0 output · context 32K
Perplexity's decision model, open-weight, Apache 2.0. Fine if you need the output to be consumable by a self-hosted stack.
$0.02 / 1M input · $0 output · context 1M+
Routing
Classify the incoming request, then send it to the right model. Cheap, fast, and keeps the frontier model for the request that needs it.
Beam / portfolio scoring
Score multiple candidates for quality or fit. Keep top-k above the confidence floor; drop or flag the rest.
Agent stop/continue gates
Should the agent keep going, retry, escalate, or stop? A decision model returns a probability per option, so your code can set thresholds.
Safety / policy gating
A second pass on flagged inputs before they hit the user, the API, or the bill. Cuts review time without replacing people.
Decision models are not OpenAI chat-compatible. They return a probability distribution over a fixed set of options, which is exactly what you want for decisions. Call them with the same key you use for chat:
Pricing, limits, and uptime are public at /stats and /pricing.
Gemini Decision 1 is live. Microsoft-Decision-1 and Perplexity Decider v1.3 are rollout partners; both pages show access requirements.
Decision models bill per 1M input tokens. Output is free on every model in this category. Your wallet is billed at the same rate as every chat call; no separate fee.
Data residency is not offered today. Keys are scoped per account. See /trust for the current posture.
Most workloads don't need a decision model. Use one when the choice is fixed, the latency matters, and the confidence signal is cheaper than a full answer.
More reading: /pricing · /models · /guides/decisions-api