Factory's Droid Router Cuts AI Coding Costs 25% by Picking Models Automatically
Factory's new Router automatically picks the best LLM per task and fails over across providers, cutting token spend 20-25% while keeping sessions alive.
Takeaways
- Factory shipped Factory Router, an automatic model picker layered on top of its Droid coding agent.
- Cuts token spend 20-25% while hitting 96-99% of Claude Opus 4.7's pass rate on internal benchmarks.
- Auto-failover across model providers keeps Droid sessions alive when an LLM endpoint degrades.
- Enterprise tier includes guaranteed TPM reservations and access to US-hosted open-source models.
- Droid works across Slack, desktop, CLI, CI pipelines, and SSH from mobile.
- Routes across Claude, GPT, Gemini, GLM-5, Kimi K2.6, and MiniMax M2.7 from one harness.
The AI coding agent space has spent the last year arguing about which model is best. Factory just shipped a feature that makes the question moot: let the router decide. The company behind Droid, the autonomous coding agent that runs across Slack, desktop, CLI, CI, and even SSH from a phone, has rolled out Factory Router, a model-routing layer that picks an LLM per session, fails over when providers degrade, and reserves dedicated throughput for enterprise customers.
The pitch is simple. Instead of forcing engineers to manually choose between Claude, GPT, Gemini, or open-source alternatives for every task, Factory Router makes that decision automatically based on task complexity, then re-routes if anything breaks downstream.
Why a router, why now
Factory has been explicit about its bet on model agnosticism. Droid is model agnostic and can change models mid-session, with models from all major providers including Anthropic, OpenAI, Google, and NVIDIA. The new router operationalizes that flexibility at the org level rather than leaving it to individual engineers.
The cost story is the headline. Factory Router cuts token spend by 20-25% while maintaining frontier performance by automatically selecting the right model for each task and routing across providers if an endpoint degrades. On benchmarks, the routing approach holds up: 99% of Claude Opus 4.7's pass rate on Terminal-Bench 2 at 20% lower cost per session, and 96% of Opus 4.7's pass rate on Legacy-Bench at 25% lower cost per session.
The framing matters. Enterprise AI costs are rising, and usage alone does not prove value. A higher token bill does not mean more engineering work is getting done. Often, it means simple work is running through the most expensive models unnecessarily.
How the routing actually works
There are essentially four moving parts under the hood:
- Automatic model selection. Factory Router chooses the optimal model for each task, drawing from a diverse pool of frontier and efficient models. If the selected model struggles to complete the task, Factory Router moves the session to a more capable model to reliably ensure high-quality outcomes.
- Provider failover. If one provider path is unavailable, Factory Router can keep sessions running through another healthy path.
- Dedicated TPM allocation. Enterprise customers get reserved throughput for critical Droid work instead of depending only on shared public capacity.
- US-hosted open-source fallback. Teams that need cost-efficient or controlled model options can route eligible work to US-hosted open-source models, so work keeps moving through provider issues, capacity limits, and model availability changes.
The aggregate claim: Droid sessions keep working when providers degrade, rate limits hit, or capacity gets constrained, with 99.9%+ request reliability by routing across models, providers, and capacity sources.
The model pool is wider than you'd expect
Beyond the obvious frontier closed models, Droid ships with a notable open-source roster. Droid Core variants based on GLM-5, GLM-5.1, Kimi K2.5, Kimi K2.6, and MiniMax M2.7 are available in the CLI, useful for air-gapped environments, cost-sensitive projects, privacy requirements, and experimentation with open-source capabilities. MiniMax M2.7 is the cheapest model available, with a 0.12x multiplier and Low/Medium/High reasoning toggles.
Where Droid is strong, and where it isn't
Factory's own Terminal-Bench results show that the harness matters more than the model. Droid with Sonnet outperforms all other agents using Opus, and Droid with Opus (58.8%) and Sonnet (50.5%) beats Claude Code with Opus (43.2%), while Droid with GPT-5 (52.5%) tops Codex CLI (42.8%). A lot of that comes from boring-but-effective engineering: implementing the grep tool using ripgrep helps Droid work with large repos more efficiently. And counterintuitively, using short default timeouts led to better average performance by cutting long, unproductive waits and failing fast.
The weak spots are honest ones. Factory Router only routes a session to a cheaper model when that model can handle the work; the savings come from sessions that did not need a frontier model, and every session that does keeps it. Routing is a best-effort optimizer, not magic, and teams with mostly hard tasks will see less of the 20-25% cost reduction.
Getting access
Factory Router is gated for enterprise rollout. With Factory Router enabled, your Software Factory routes each Droid session to the model that fits the work, with provider routing for reliability and Enterprise Controls for governance, and teams interested in enabling it for their organization need to reach out to the Factory team.
For individual developers, the existing Droid CLI still ships with manual model selection and BYOK support for Ollama, Groq, OpenRouter, Fireworks, and others, so you can keep driving the routing decision yourself if you want fine-grained control. The Router is the version of that workflow built for organizations that would rather not think about it.