New AI Router Reaches Near-Claude Opus Performance While Cutting Costs 25%

Factory has launched Factory Router, a new AI model-routing system designed to reduce enterprise software engineering costs by automatically selecting the most appropriate model for each coding task. 

The product entered private research preview on June 1 and is available through the company’s CLI and desktop applications.

The company said Factory Router reduces token spending by 20–25% while maintaining near-frontier performance by dynamically assigning workloads across a mix of frontier and lower-cost models. 

If a selected model struggles with a task, the system can escalate the session to a more capable model and reroute workloads across providers when endpoints degrade or capacity becomes constrained.

Factory benchmarked the system against Anthropic's Claude Opus 4.7. According to the company, Factory Router achieved 99% of Opus 4.7's pass rate at 20% lower cost per session on Terminal-Bench 2, and 96% of Opus performance at 25% lower cost on Legacy-Bench.

The launch targets a growing concern among enterprises deploying AI coding tools at scale: rising inference costs. 

Factory argued that organisations frequently route routine work, such as documentation updates, code refactoring, bug fixes, and search-heavy investigations, through premium reasoning models, thereby increasing token consumption without necessarily improving engineering output.

Rather than introducing a new model, Factory Router serves as an orchestration layer that evaluates each Droid session and selects from a pool of frontier- and efficiency-focused models. Less demanding tasks can be handled by lower-cost models, while more complex engineering work can be routed to frontier systems.

Factory CEO Matan Grinberg wrote in a post on X, “If your daughter needs tutoring in algebra, you can probably find someone cheaper than Albert Einstein.” 

“Giving every task to GPT-5.5 or Opus 4.8 is overkill.” He said many software engineering tasks can be completed “10x cheaper and 10x faster” using smaller models, forming the basis for Factory Router's automated model-selection system.

Factory said Router delivers 99.9%+ request reliability through provider failover, dedicated throughput allocations for enterprise customers and routing across multiple providers and capacity sources. Enterprise administrators can further customise routing behaviour through policy controls, workflow guidance and organisation-wide model governance settings.

Factory published additional benchmark data describing what it calls the cost-performance “Pareto frontier” for AI coding workloads. 

According to the company, Router operates near the point where cost reductions remain substantial while benchmark performance declines only marginally. More aggressive routing configurations reduced Terminal-Bench 2 costs to 56% of the Opus baseline but lowered pass rates to 81%, while Legacy-Bench costs fell to 30% of baseline costs with pass rates dropping to 49%.

The company raised a $150 million Series C in April, led by Khosla Ventures, with participation from Sequoia Capital, Blackstone and Insight Partners, valuing the company at $1.5 billion.

The announcement also drew attention from executives across the AI tooling ecosystem. 

Reacting on X, Hugging Face CEO Clément Delangue said automatic routing could “redistribute value capture and usage towards many more models than just frontier ones,” particularly open-source and lower-cost alternatives. 

Delangue argued that requiring users to manually select models creates enough friction that many default to frontier models for all tasks.

The launch comes as enterprises increasingly scrutinise AI spending amid a broader debate around maximising AI usage and token consumption as a measure of productivity. 

While companies have aggressively rolled out coding assistants and AI agents, investors and executives are increasingly asking whether rising inference bills are translating into proportional gains in software output.