Problem

Dari addresses the cost and reliability problem of running AI agents across a growing assortment of models. Agent workloads generate many requests, but not every step requires the most capable—and most expensive—model. Teams therefore face a tradeoff between cost, latency, quality, and policy, while switching models mid-session can also disrupt prompt-cache accounting. Dari’s published benchmark illustrates the economics: its router reports 79.8% accuracy on Terminal-Bench 2.1 for $76 in total cost, compared with other model runs ranging from 66% to 84% accuracy at $200 to $2,060.

The clearest current use case is high-volume coding and other production agent workloads, where a team wants to run more agent steps without proportionally increasing inference spend. Dari originated from the company’s work on browser-automation agents, but its current positioning is broader: it provides the control layer for teams deploying coding, research, operations, and other agents.

Product / Service

Dari is a hosted, OpenAI-compatible model router. An application or coding client sends requests to one Dari endpoint, while the customer configures the candidate models and provider keys; Dari then selects a model for each request. Routing can use a small selector model, deterministic heuristics based on evaluations and prices, or customer-defined instructions. The same endpoint is designed to work with OpenAI SDKs and agent clients such as Claude Code, Codex, and Pi, and responses expose which model was selected.

The benefit is a drop-in way to optimize model choice without rewriting an agent for every provider. Dari supports customer-managed or Dari-managed provider keys, tracks routing decisions and costs, and uses cache-aware switching so changing models does not silently invalidate the economics of a long agent loop. The commercial model is early and usage-oriented: the public pricing page lists a free tier with one-time credits and a $200-per-month Scale plan with monthly credits and concurrency limits, while stating that Dari initially charges no markup on token costs; enterprise deployments are handled through a call.

Market

Dari competes in developer tools and B2B AI-agent infrastructure, specifically the emerging model-routing and AI-gateway segment. Its adjacent alternatives include LiteLLM, OpenRouter, Portkey, Vercel AI Gateway, Bifrost, Martian, and other routing platforms. Dari’s differentiation is its explicit focus on agent workloads, including coding-agent traces, quality-versus-cost routing, cache-aware model switching, and policy-controlled access to a configured model set. Its earlier browser-agent positioning also places it near browser automation and robotic-process-automation platforms, although the current site emphasizes routing rather than a standalone browser-agent builder.

The company appears to be an early-stage startup rather than a business with publicly demonstrated scale. Y Combinator lists it as an active Fall 2025 Developer Tools B2B company founded by Benjamin Hong and Avyay Varadarajan; another company profile lists two employees in San Francisco, and a third-party traction profile reports $500,000 raised. The public evidence shows a live product, benchmark claims, self-serve pricing, and an enterprise sales motion, but it does not disclose customer counts, revenue, or usage, so the available record cannot establish whether Dari is already revenue-generating or still pre-revenue.