x-rush.ai's smart routing picks the best AI model for each generation task automatically - you describe what you want, the system routes it to the right model, and you never have to know which model ran. We connect to the industry's top-tier models - Claude Fable 5, GPT-5.6, Gemini 3.5, Qwen3.7 Max, DeepSeek V4, Grok 4.5, ChatGPT Image 2, nano-banana-2, Midjourney, Kling 3.0, Veo 3.1, Sora, Seedance, Suno V5, ElevenLabs - and the list is not fixed. It follows the frontier: when the next Fable or GPT ships (and 2026 ships one almost every month), we connect it. This page is the technical reference. For the narrative version, see our smart routing deep dive; for what each model is good at, see the 2026 model comparison.
What Is Smart Routing
In 2026 there is no single best AI model. Claude Fable 5 sits at #1 on the Arena leaderboard and the Artificial Analysis Intelligence Index (Anthropic's Mythos-class flagship, $10/$50 per million tokens, 1M context, 61 tok/s). GPT-5.6 Sol is the strongest reasoner (Artificial Analysis Coding Agent Index 80, $5/$30, 1.05M context). Qwen3.7 Max is the best for Chinese and the best value (Artificial Analysis #5 globally, #1 in China, 56.6 score). nano-banana-2 is the best at following image instructions ($0.067/image, 3-5s). Kling 3.0 and Veo 3.1 own video. ElevenLabs and Suno V5 own audio. Asking "which model is best?" only makes sense with a "for what?" attached.
Smart routing answers the "for what?" automatically. Instead of locking you into one model, it routes each task - text, image, video, audio - to the model that fits best. And because the frontier moves every month, we keep adding better ones as they ship. Our model pool is not a fixed roster; it follows the world.
Three-Layer Architecture
Our routing is a three-layer design. Layers 1 and 2 are the daily workhorse; Layer 3 is where the routing gets genuinely smart - picking the best model for this specific task rather than just a good default.
Layer 1: Modality Gate
The first decision is what kind of thing you're asking for. Each modality routes to its own model pool:
- Text -> text sub-router (Layer 2)
- Image -> ChatGPT Image 2 / nano-banana-2 / Midjourney
- Video -> Veo 3.1 / Kling 3.0 / Sora / Seedance
- Audio (TTS / cloning / ASR) -> ElevenLabs
- Audio (music / BGM) -> Suno V5 / ElevenLabs Music
This is static rule-based routing - simple, predictable, zero extra inference cost. It's the foundation everything else builds on.
Layer 2: In-Modality Selection
Within a modality, pick the best model for the task. For text this is a deliberate strong/weak split:
- Default text generation, chat, daily push content -> Qwen3.7 Max (best Chinese, Artificial Analysis #5 globally / #1 in China, 1-3s response, ~1/8 the cost of GPT-5.6 max)
- Harder reasoning tasks -> DeepSeek V4 (SWE-bench Verified 80.6%, within 0.2 points of Opus 4.8 at $0.14/$0.28 per MTok - the "price butcher" of 2026, cheapest frontier-class model on the market)
This is the same idea behind RouteLLM's cascade (easy queries to the cheap model, hard ones to the strong model), just with a static rule instead of a learned classifier. For image, video, and audio, the router picks the best connected model per modality - ChatGPT Image 2 for text-in-image (Arena T2I #1 at 1512, +242 over the runner-up), nano-banana-2 for speed and cost ($0.067/image, 4 resolutions up to 4K), Midjourney V7 for pure aesthetics. Multi-model selection within each modality is what Layer 3 handles.
Layer 3: Intent-Aware Routing
This is the layer that turns "we picked a good default" into "we picked the best model for this specific task."
Intent classifier: a lightweight model reads the query's language, difficulty, domain, and whether it's agentic, then routes to the best model + fallback chain. Routing targets across the connected pool:
- Chinese / cost-sensitive / general chat -> Qwen3.7 Max (Qwen3.7 Plus as downgrade fallback)
- English agentic / coding / tool use -> Claude Fable 5 (escalate from Sonnet 5 / Opus 4.8 for the hardest - Fable 5 is Anthropic's #1 flagship, Arena Intelligence Index #1, $10/$50)
- Tasks needing tunable reasoning effort -> GPT-5.6 Sol / Terra / Luna (low effort for easy, max for hard; $5/$30 down to $1/$6 - a 5x price spread within one family)
- Multimodal agentic -> Gemini 3.5 (fastest output in class, ~722 tok/s on Flash)
- Deep reasoning but cost-sensitive -> DeepSeek V4 Flash ($0.14/$0.28, RouteLLM-style cascade) or Grok 4.5 ($2/$6, Artificial Analysis Intelligence Index #4, "Opus-class but faster" at roughly a quarter of Opus 4.8's price)
Gateway failover: unified provider adaptation, rate-limiting, automatic failover (e.g., Claude Fable 5 capacity-limited -> GPT-5.6 Sol -> Gemini 3.5), guardrails, prompt versioning, and cost/latency/quality instrumentation. See Gateway Reliability below.
The 2026 Model Pool
The top-tier models we route to, across modalities. The pool is not frozen - when a new model genuinely takes a lane, we connect it.
| Modality | Top-tier models x-rush routes to |
|----------|----------------------------------|
| Text | Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, GPT-5.6 (Sol/Terra/Luna), Gemini 3.5, Qwen3.7 Max, DeepSeek V4, Grok 4.5, GLM |
| Image | ChatGPT Image 2, nano-banana-2, GPT Image 1.5, Midjourney V7, FLUX.2 |
| Video | Veo 3.1, Kling 3.0, Sora 2, Seedance 2.0, Wan 2.6, SkyReels V4 |
| Audio | ElevenLabs (TTS/clone/ASR + Music), Suno V5 (music/BGM), Udio, Lyria 3 |
Naming hygiene: we only write model names we can verify against official sources. Suno doesn't prominently publish version numbers, so we write "Suno V5," the version in production as of mid-2026 (v5.5 added voice cloning and custom models in March 2026). Claude Fable 5 is real - Anthropic's Mythos-class flagship above Opus, released June 9, 2026, briefly suspended June 12, redeployed July 1. Anthropic's 2026 line: Fable 5 (top, $10/$50) > Opus 4.8 (prior flagship, $5/$25) > Sonnet 5 (agentic, $2/$10 intro, rising to $3/$15 after Aug 31).
Routing Rules
How a task flows through the system:
- Submit - you describe what you want (e.g., "generate a meme about cats")
- Modality gate - content type routes to the right pool (image -> ChatGPT Image 2 / nano-banana-2, text -> Qwen3.7 Max, video -> Veo 3.1 / Kling 3.0, audio -> ElevenLabs / Suno V5)
- In-modality selection - text tasks pick Qwen3.7 Max by default, DeepSeek V4 for harder reasoning; image tasks pick ChatGPT Image 2 for text-in-image, nano-banana-2 for speed/cost, Midjourney for aesthetics
- Plan & quota - Free = standard queue, Pro = priority; enough credits? not rate-limited?
- Run - the task executes on the model API, the result is stored, you get a download URL
End-to-end latency depends on modality: text is fastest (1-3s), image is a few seconds (nano-banana-2 at 3-5s, ChatGPT Image 2 under 3s for 1K), video is slowest (60-120s for an 8s clip from Veo 3.1 or Kling 3.0).
Gateway Reliability
Production-grade reliability is the Layer 3 gateway:
- Failover - if a provider rate-limits or goes down, traffic fails over to the next-best model automatically (Fable 5 capacity-limited -> GPT-5.6 Sol -> Gemini 3.5; this matters in practice - Fable 5 had a 19-day suspension in June 2026)
- Rate-limiting - per-user and per-IP QPM caps to keep the system fair and stable
- Guardrails - input/output safety checks at the gateway
- Observability - cost, latency, and quality instrumentation per request, closing the loop on routing effectiveness
- Prompt versioning - track and roll back prompt changes
How We Compare
There are six routing strategies in 2026. We layer them rather than pick one:
| Approach | What it does | Our take |
|----------|--------------|----------|
| Static rule-based | Route by content type + fixed rules | Layer 1 - our foundation |
| RouteLLM-style cascade | Train a router to split easy/hard queries | Layer 2 borrows the cascade idea (Qwen3.7 Max + DeepSeek V4) |
| OpenRouter (aggregator) | One API, many models, route by price/availability | We optimize for quality, not just price |
| Portkey (gateway) | Failover, observability, governance | Layer 3 - reliability layer |
| LMRouter (multimodal API) | One endpoint for all modalities | We're multimodal native, with quality routing on top |
| Learned intent classifier | Classify intent -> best model + fallback | Layer 3 - the quality-first layer |
The differentiation we're aiming for, honestly: (1) multimodal native, not text-only; (2) quality-first in-modality routing, not just price routing; (3) Chinese-native with Qwen3.7 Max as the backbone; (4) production gateway reliability with real failover; (5) the model pool follows the world - when the next Fable, GPT, or Veo ships, we connect it, and your outputs quietly get better with no UI change.
FAQ
Why not just use one model? No single model is best at everything. Qwen3.7 Max is best for Chinese and value (Artificial Analysis #5 globally, #1 in China) but loses to GPT-5.6 on the hardest reasoning (Coding Agent Index 80). Fable 5 is the strongest overall (Arena #1, Intelligence Index top) but at $10/$50 per MTok it's overkill for high-volume chat. Routing gives you the best of each without the tradeoffs.
Why Qwen3.7 Max for Chinese? It was trained primarily on Chinese text and produces noticeably more natural Chinese than Western models - especially in casual conversation, creative writing, and culturally-specific content. It's also ~1/8 the cost of GPT-5.6 max and responds in 1-3s.
How do you decide which top-tier model to connect? We watch the Arena and Artificial Analysis leaderboards, the official pricing pages, and what real users report. When a model is genuinely top-tier in its lane - Fable 5 for agentic (Arena #1), GPT-5.6 for reasoning (Coding Agent Index 80), ChatGPT Image 2 for text-in-image (Arena T2I 1512, +242 over #2), Kling 3.0 for production video, SkyReels V4 for text-to-video (Artificial Analysis #1 in March 2026) - we connect it. The pool follows the frontier; it's not a fixed list.
How is cost controlled? By routing the ~90% of tasks that don't need frontier reasoning to Qwen3.7 Max, and escalating only the ~10% that genuinely need deep reasoning to DeepSeek V4 ($0.14/$0.28 - the cheapest frontier-class model in July 2026) or Sonnet 5 / GPT-5.6 for the hardest. This is the RouteLLM insight: most queries don't need the strongest model.
What happens when a new model ships? We connect it. Claude Fable 5 came back from its 19-day suspension on July 1; GPT-5.6 launched July 9 with Sol/Terra/Luna; Grok 4.5 landed July 8 at $2/$6. The frontier moves monthly, and our routing moves with it - your outputs quietly get better, no UI change required.
Can I pick a model manually? Pro users can override automatic routing and specify models manually. Enterprise users can whitelist models for compliance or quality preferences.