dari.dev

Most routing libraries ask you to write the policy. Dari asks you to trust a router model it trained and publishes evals for. That's the real trade: less control, far less glue code. Pick it when your spend is dominated by multi-turn agent loops with stable prompt prefixes. That's where cache-aware selection matters — a cheaper model that blows the cache can cost more than the expensive one you were avoiding. The GPT-5.6-sol ($5.00 in, $0.50 cache read) versus GLM 5.2 ($1.40 in, $0.14 cache read) comparison on the homepage is exactly the math the router is doing for you. Pick it if you're already on Claude Code, Codex or Pi. The install is one curl and `dari claude`; there's no migration project. Any OpenAI-compatible base URL works too. Pass if you need contractual, guaranteed pricing today. Early access means terms and roadmap are in motion. Pass if your workload is a handful of long prompts where you've already tuned the model by hand — the router has little to work with. And pass if you want a routing ecosystem with years of third-party adapters; this is young. The closest comparison is LiteLLM. LiteLLM gives you a library and you define the rules; Dari gives you a router that decides, plus a hosted runtime. If your team enjoys owning the policy, LiteLLM is fine. If nobody wants to maintain a cost-routing heuristic, Dari earns its fee. Helicone sits in a different slot — observability, not selection — and the two are complementary. Where it bites: the accuracy figures come from Dari's own benchmarks. Vendor evals are a signal, not proof. Run the 47-task Next.js Self-Bench or the 103-task Snowflake bench against your own prompts, then watch cost per completed task for two weeks.