On October 12, DigitalOcean overhauled its Inference platform routing, replacing four specialized preset routers with a single unified model name, inference-router. While the change promises simplified integration via optional routing_preference settings (balanced, budget, quality), early benchmarks suggest it may significantly increase costs for developers handling simple, low-complexity prompts. A recent test by developer Remdore demonstrated that the new router frequently selects high-cost reasoning models for trivial tasks, resulting in expenses up to 420 times higher than direct calls to smaller, cheaper models.

The Test Setup and Routing Behavior

The experiment involved sending 20 distinct short promptsβ€”ranging from basic code snippets and arithmetic to simple greetingsβ€”through the new inference-router in all three preference modes, as well as through the legacy preset routers and directly to four fixed models. The results were stark: across all three new preferences, the router predominantly selected Kimi K3, a reasoning model costing $3 per million input tokens and $15 per million output. For the prompt 'hi', the router directed traffic to Kimi K3, which consumed 148 to 206 output tokens including reasoning overhead. In contrast, a direct call to Claude Haiku 5.5 cost $0.000008, while the routed Kimi K3 call cost between $0.0025 and $0.0034. This represents a 300% to 420% price increase for identical, trivial output.

Cost and Latency Discrepancies

Aggregated data from the test reveals that the new inference-router was consistently more expensive and slower than direct model calls for this workload. The budget preference, surprisingly, was the most costly option at $5.55 per 1,000 requests with a median latency of 9.0 seconds, compared to $3.48 per 1,000 requests and 2.9 seconds latency for direct calls to Claude Opus 5.5. The legacy router:general preset, which mostly defaulted to DeepSeek V4 Flash, cost just $0.31 per 1,000 requests. The new router's tendency to route simple queries to reasoning models also introduced reliability issues; several prompts returned empty answers because the model exhausted the 600-token limit on internal reasoning steps before generating a response, a failure mode absent in direct calls to non-reasoning models.

Developer Recommendations and Caveats

DigitalOcean notes that the router is currently in public preview and is optimized for complex agentic workflows and software engineering tasks, where intelligent model selection might justify the overhead. However, for applications dominated by short, predictable inputs, the test suggests that sticking to direct model calls or the legacy presets remains far more economical. Developers migrating to the new system must rigorously check the finish_reason field in API responses to detect truncated outputs caused by reasoning token limits. Additionally, the budget preference does not strictly guarantee the cheapest model, as the router still weighs quality, meaning cost savings are not guaranteed without custom testing against specific traffic patterns.

Key Takeaways

  • The new inference-router frequently selects Kimi K3 for simple prompts, causing up to 420x cost increases compared to direct Haiku 5.5 calls.
  • The budget routing preference was the most expensive option in the test, contradicting expectations for cost-optimized routing.
  • Reasoning models routed by the new system can exhaust token limits on internal thinking, resulting in empty responses for simple queries.
  • Legacy routers remained 15 to 90 times cheaper for basic tasks, though they are no longer prominently displayed in the UI.

The Bottom Line

Unless you are running complex agentic workflows, DigitalOcean's new unified router is currently a cost trap for simple prompts. Measure your traffic before migrating, because 'budget' doesn't mean cheapβ€”it means 'we still think you need a PhD to say hello.'