FlashLabs Launches 'Routing DSL' for OrcaRouter — Achieving Fable 5-Level Reasoning Intelligence at Minimal Cost, Enabling Full Control of AI Workloads via YAML + CEL-Defined Inference Graphs

FlashLabs Inc. (Headquarters: Chiyoda City, Tokyo; CEO: Yoichi Hosoi; hereinafter 'FlashLabs') announces the launch of a new feature, 'Routing DSL,' for OrcaRouter — an AI inference gateway developed by U.S.-based Continuum AI and exclusively distributed in Japan by FlashLabs. Routing DSL is a domain-specific language that enables declarative definition of inference graphs using YAML and CEL (Common Expression Language). By combining multiple LLMs based on prompt complexity and task type, it aims to deliver frontier-model-level intelligence — exemplified by Anthropic's Claude Fable 5 — at significantly lower costs.

Background and Objective

By 2026, enterprise AI adoption is shifting from 'which model to use' to 'how to combine models.' Anthropic's Claude Fable 5 achieved first place in the Artificial Analysis Intelligence Index (score: 64.9), setting a new benchmark for reasoning intelligence. However, only limited workloads can justify the cost of using such high-performance models continuously.

Most production workloads consist of simple tasks such as extraction, classification, and formatting. Continuously using frontier models like Fable 5 is not cost-effective. On the other hand, replacing all models with low-cost open-source alternatives compromises quality when advanced reasoning is required.

'Routing DSL' addresses this by making routing logic a declarative, auditable, and version-controlled code, providing the control needed for agent-era production systems. Developers can now design inference graphs using a single YAML file — for example, routing certain prompt types to Fable 5-class models and others to fast, low-cost models.

Overview of Routing DSL

Launch Date: June 15, 2026 (Monday)

Access:

Documentation: https://docs.orcarouter.ai/ja/routing/routing-dsl

Dashboard: OrcaRouter Dashboard → Routing → Strategy → DSL

Syntax: YAML + CEL (Common Expression Language)

Five Routing Strategies

Routing DSL enables the construction of inference graphs by combining the following five strategies:

1. Route by Difficulty

Automatically assesses prompt complexity, routing high-level reasoning tasks to frontier models like Claude Opus or GPT-5.5, and routine tasks to open models like DeepSeek V4 Pro or Qwen3.6.

2. Route by Task

Selects the optimal model based on task type — such as 'code generation,' 'summarization,' 'translation,' or 'data extraction.' For example, Claude Sonnet for coding, DeepSeek for extraction.

3. Fan-out to Multiple Models

Sends the same prompt in parallel to models like GPT-5.5, Claude Opus, and Gemini 3.1 Pro, then aggregates the results to achieve response quality beyond any single model.

4. Fallbacks & Judges

Automatically evaluates model responses and triggers fallback to another model if quality falls below a threshold. Supports mid-stream switching during response generation.

5. Optimize for Cost, Latency, or Quality

Applies policies such as 'cost-first,' 'latency-first,' or 'quality-first' across the entire graph, enabling declarative optimization aligned with business requirements.

Article with Detailed Explanation

Value for Enterprises

1. Achieve Fable 5-Level Reasoning at Lower Cost

By combining multiple models in a graph structure instead of relying on a single frontier model, Routing DSL delivers intelligence comparable to Fable 5 at a fraction of the cost. There is no need to use the most expensive model for every prompt.

2. Drastically Reduce Operational Burden with Declarative Configuration

Declarative configuration via YAML eliminates the need for hard-coded if/else logic or code changes with every model update. When a new model is released, updating a single line in the YAML file propagates the change across the entire inference graph.

3. Maintain Transparency and Auditability

The decision logic of inference graphs built with Routing DSL is fully visualized on a per-request basis. The dashboard and response headers allow complete traceability — showing which prompt was routed to which model and under what conditions fallbacks occurred.

Technical Features

Routing conditions in Routing DSL are written in CEL, enabling expressions such as:

prompt.difficulty >= 0.8 → Frontier model

prompt.task_type == "code_generation" → Code-specialized model

response.quality_score < 0.7 → Automatic fallback

cost_budget.monthly_remaining > 100 → Cost-constrained routing

Routing DSL integrates with OrcaRouter’s existing features, including LinUCB contextual bandit-based adaptive routing, mid-stream failover, and access to over 200 available models.

Example Available Models

Anthropic Claude Opus 4.8 API

OpenAI GPT 5.5 API

Gemini 3.5 Flash

MiniMax M3

DeepSeek V4 Pro API

Qwen3.7 Max

Future Developments

OrcaRouter plans to release a template library for inference graphs built with Routing DSL, promoting community-driven best practices. It also intends to enhance performance analytics and add automatic optimization of routing strategies via A/B testing.

Executive Comment

Yoichi Hosoi, CEO of FlashLabs Inc.

'Fable 5 is one of the most intelligent AI models available today. However, processing every enterprise production request with Fable 5 is neither cost-effective nor practical. With Routing DSL, we’ve created a world where high-difficulty reasoning uses the best models, and routine tasks use cost-optimized models — and the blueprint for this combination can be declared in YAML. We will continue evolving OrcaRouter as the foundation that empowers Japanese enterprises to leverage top-tier AI without fear of cost.'