cascaio.com

What It Is

Casca is an LLM API cost-optimization routing engine designed for teams spending around $10K–$200K per month, or more, on large-model APIs. It is not a model itself, nor is it simply a model aggregator. Instead, before requests reach the model, it performs complexity classification, cache matching, and model selection, aiming to reduce bills without changing prompts or rewriting business logic.

Core Capabilities

At its core is LOW/MED/HIGH/CACHE tiered routing: simple queries can be sent to Gemini Flash, mid-level generation to GPT-4o-mini or Claude Haiku, while high-risk or complex tasks remain on GPT-4o/Claude Sonnet. The materials state that classification latency is under 1ms, and that the production engine includes 160 rules, with support for MiniLM fallback, Auto-Learn, and semantic caching. For scenarios with many repeated requests, such as customer support, e-commerce, HR, and insurance, the official modeling suggests savings of up to 55%–75%. For code generation, however, the figure is only 19%–31%, indicating that Casca is better suited to businesses with a high share of simple or repetitive traffic.

Pricing and Trial

Casca offers a Free plan with 10M tokens and BYO API keys. Starter is $299/month, Growth is $999/month, and Scale starts at $2,499/month. It can also be billed at 12% of verified savings. The materials mention both a 60-day trial and a 30-day free trial; the actual terms should be confirmed at signup. Under the BYO-key model, LLM usage fees are still charged directly by OpenAI, Anthropic, Google, and other providers, while Casca charges for routing.

Pros and Cons

The main advantage is very lightweight integration: it is compatible with the OpenAI SDK, requiring only a base_url change, and supports a CASCA_BYPASS=true bypass for quick fallback during outages. It explicitly supports 14 languages, including Simplified Chinese and Traditional Chinese, and provides a Dashboard, auditing, quality SLA, Provider Pool, and Zapier API. On privacy, it also states zero-log handling, no prompt training, no data persistence, API key isolation, and DPA support.

The downside is that results depend heavily on workload structure, so “60% savings” cannot be assumed universally. Its savings benchmark is mainly modeled against a GPT-4o flat-rate baseline, while real bills will also be affected by retries, cache hit rates, and provider pricing. The text indicates that SOC 2 Type II is still in progress, so customers with strict compliance requirements should conduct further due diligence.

Who It’s For and Access from China

Casca is suitable for AI SaaS, customer support, finance, e-commerce, HR, and insurance teams that already have stable LLM traffic and want to reduce costs without building their own routing layer. For individual developers or low-usage projects, the free tier can be tested, but its commercial value may be limited. Access from mainland China is not clarified in the materials, and because Casca relies on overseas services such as OpenAI, Anthropic, and Google, network connectivity and payment may be uncertain. If domestic compliance and direct connectivity are required, it may be worth evaluating model gateways from Chinese cloud providers or local large-model platforms as well.