Factory Router Cuts AI Coding Costs 25% Without Dropping Performance
Factory Router automatically picks the cheapest model that can handle each task, cutting costs 20-25% while matching frontier performance on enterprise coding benchmarks.
Takeaways
- Factory Router automatically selects the right LLM for each coding task mid-session, cutting costs 20-25% vs. always using Claude Opus 4.7. Read more
- On Terminal-Bench 2, it achieves 99% of Opus 4.7's pass rate at 20% lower cost; on Legacy-Bench, 96% pass rate at 25% lower cost.
- The router can escalate to a more capable model mid-session if the current model struggles, rather than committing to one model upfront.
- It provides 99.9%+ request reliability via provider failover, routing across Anthropic, OpenAI, Google, and open-source models when endpoints degrade.
- Enterprise admins can set custom routing rules and context to shape automatic model selection for their org's specific workflows.
- Currently in private research preview in the Factory CLI and Desktop App; BYOK support for Router is not yet available. Contact Factory for access.
Most AI coding platforms let you pick a model and stick with it. That single choice ends up being a tax: use a frontier model like Claude Opus for everything and you overpay on every doc update, boilerplate refactor, and trivial bug fix. Use a cheaper model to save money and you risk dropping the ball on the tasks that actually need depth. Factory Router is Factory's answer to that dilemma -- an automatic model dispatcher that routes each Droid session to the right model for the job, without any manual intervention.
The problem no one wants to admit
Engineers are defaulting to the most performant models out of fear of losing performance. Simple questions, mechanical refactors, documentation updates, small bug fixes, and search-heavy investigations all end up on the same premium path as work that truly needs frontier capability. The result is a budget that burns fast with no clear increase in output. A higher token bill does not mean more engineering work is getting done.
The AI coding tools market has bifurcated sharply. On one side, products like Cursor, Windsurf, and GitHub Copilot target individual developers with fast-response code completion. On the other, enterprise-focused platforms like Factory focus on autonomous, long-horizon tasks that span multiple repositories, systems, and days of wall-clock time. At that scale, model choice compounds into a real cost center.
How routing actually works
Factory Router automatically picks the best model for each Droid session, 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. The key word is mid-session: the router doesn't just make a choice at the start and commit to it. It monitors progress and can escalate to a stronger model if the current one hits a wall.
Factory's position in this landscape is distinctive because it treats model selection as a routing problem rather than a branding problem. A single Droid task can hand off between Anthropic's Claude for planning, and a smaller open-source model for routine test authoring. The pool Factory draws from is wide -- the model stack includes Claude Opus 4.8, GPT-5.4, Gemini 3 Flash, and open-source options like MiniMax M3 and Kimi K2.7 Code, each with cost multipliers ranging from 0.12x to 4x.
Beyond cost optimization, the router also handles reliability. Factory Router provides 99.9%+ request reliability by routing across models, providers, and capacity sources -- more than any single-provider platform can offer. If one provider path is unavailable, Factory Router keeps sessions running through another healthy path.
The benchmark numbers
Factory tested the router against a Claude Opus 4.7 baseline on two internal enterprise-grade benchmarks:
- Terminal-Bench 2: 99% of Claude Opus 4.7's pass rate at 20% lower cost per session.
- Legacy-Bench: 96% of Claude Opus 4.7's pass rate at 25% lower cost per session.
The savings also hold when measured against only completed tasks: cost per successful run is 80.5% of Opus on Terminal-Bench 2 and 78.0% on Legacy-Bench. That matters because a router that saved money by abandoning hard tasks early would look cheap on aggregate but terrible in practice.
Factory also published a separate code review benchmark across 13 models that illustrates exactly why routing makes sense. Cost explains only ~21% of quality variance -- model architecture and training matter far more than token budget. In other words, the relationship between spend and quality is lumpy, not linear, which is precisely the condition that makes a smart router valuable.
The Pareto frontier, explained
Factory's blog post introduces a useful framing: the cost/performance Pareto frontier. Near the top, the curve is nearly flat -- cost drops sharply while performance barely moves, because the first work to leave the frontier model is work cheaper models handle just as well. Lower down, the curve bends, and what's left is the work that genuinely needed the frontier model.
Factory Router operates on that flat stretch, just before the bend. Push past the bend and every further dollar saved costs far more performance. To illustrate the extremes: the most aggressive routing measured cut Terminal-Bench 2 to 56% of Opus cost but dropped pass rate to 81%; on Legacy-Bench, 30% of Opus cost came at a pass rate of just 49%. Factory's default configuration deliberately stops well before that cliff.
Enterprise controls baked in
The router is not a black box. Admins can provide routing guidance so that automatic model selection reflects how work actually happens inside the organization -- describing workflow patterns, codebase areas, toolchains, and model preferences that should shape automatic selection. The same policy surfaces that govern other Factory models apply here, so there is no separate control plane to manage.
A few things worth noting for teams evaluating it:
- Dedicated TPM: Enterprise customers have reserved throughput for critical Droid work instead of depending only on shared public capacity.
- BYOK not yet supported: Factory confirmed in the announcement thread that Bring Your Own Keys is not yet available for Router, though they are working on it for the most popular BYOK models.
- Model policy controls: Admins can allow or block Factory Router org-wide, the same way they manage other models in Factory.
- Mission workers included: Long-running autonomous background sessions benefit from the same automatic routing as interactive sessions.
Where it fits in the bigger picture
Perhaps the most important shift in AI coding is the emergence of long-running autonomous workflows. Instead of responding to a single prompt, agents now operate through execution loops that can run for extended periods. In that context, model selection stops being a one-time configuration decision and becomes an ongoing operational cost. Factory Router is a bet that this problem is better solved by infrastructure than by individual engineers making manual choices session by session.
Factory, the San Francisco-based startup building autonomous AI agents for enterprise engineering teams, closed a $150 million Series C at a $1.5 billion post-money valuation. Factory Router is currently in private research preview, available in the Factory CLI and Desktop App. Once enabled for an org, it appears in the model picker for every user with no setup required. Teams interested in access can reach out directly through the Factory website.