The best model routing is task-specific

A new wave in model routing

Over the past 1.5 months, model routing techniques have been emerging one after another, including OpenRouter's Fusion, Cognition's Devin Fusion, Factory's router, and Vercel's routing rules. The background is simple: top-spec models are so expensive per token that they can't be used for every task, and the realization that not every task actually requires frontier-level performance.

The limits of general-purpose routers

Frontier models can burn $600 per hour, but at the upper end of the performance curve, cost drops sharply while quality barely changes. In this range, Factory achieved a ~99%-level pass rate at 20% lower cost than Claude Opus 4.7, and Cognition's Devin Fusion delivers Fable 5-level performance 35% cheaper.

However, general-purpose gateways have a fundamental limitation. When 3 pages of a scanned loan document are dense tables requiring an advanced vision model, while 1-2 pages are plain text where cheap extraction suffices, a general-purpose router can't tell the difference — because it doesn't understand any task in depth.

Success stories from task-specific teams

Harvey (legal) decomposes requests into sub-tasks and then selects the appropriate model for each. High-volume Vault tasks use fast models like Sonnet 4.6 and Gemini 3 Flash, and where needed, its own custom-trained open-source models, achieving frontier-level legal performance at low cost.

Decagon (customer support) follows a clearer strategy. New use cases require top-tier general-purpose models, but once a use case matures, the argument is that "general intelligence is overhead" and "the smallest, fastest model fine-tuned for the specific task" is optimal. Decagon now runs ~90% of its workload on fine-tuned open-source models.

The persistent gap in document AI

It was expected that frontier VLMs would completely replace document parsing, but reality differs. On the ParseBench benchmark, GPT-5-mini and Haiku score under 10% on visual localization, while specialized parsers reach 55-80%. For Gemini, going from "minimal" to "high thinking" increases cost by 4x while improving score by only about 5 points.

LlamaParse systematically exploits this gap. It combines agent-based automatic routing, a custom engine that cuts vision tokens by 50-90% on text pages, fine-tuned VLMs for tables and charts, and an output verification agent. Its agent mode dominates 4 out of 5 categories at 84.9%, and its cost-efficient mode ranks at the top at about 0.3 cents per page.

Division of roles between the two layers

General-purpose gateways like OpenRouter handle broad provider-level routing and ensembling (model availability, real-time pricing, assembling panels for hard problems). The excess returns come from the task-specific layer, accruing to teams that go deepest into a single workflow — Factory (software engineering), Cognition, Harvey (legal), Decagon (support), and LlamaParse (documents).