For teams spending $10K–$200K / month on LLM APIs

Cut your LLM bill 30–75%.
One line of code.
Numbers backed by published research.

A research-backed routing engine that classifies each request in under 1 ms and sends it to the model that actually fits, not the most expensive one. Same outputs, same SLA. Up to 97% per token; 30–75% on real bills depending on workload.

✓ Live in under 30 minutes ✓ Prompts never stored or trained on ✓ One-click bypass if anything goes wrong

# 現有程式碼 / your current code client = OpenAI(api_key="sk-...", base_url="https://api.openai.com/v1") # 改成這樣,其餘不動 / change to this, nothing else client = OpenAI(api_key="sk-casca-...", base_url="https://api.cascaio.com/v1") # ✓ # 退路 5 秒 / escape hatch in 5 seconds export CASCA_BYPASS=true
30–75%real bill reduction (by workload)
<1 msclassification time per request
14languages classified natively
160rules in the v2.6.2 production engine
Core technology

Zero-compromise cost optimisation,
four systems working together.

Classify complexity, protect quality, cache answers, learn automatically.

Complexity-aware routing

Every prompt is classified HIGH, MED or LOW in under 1 ms. Simple queries route to Gemini Flash (up to 97% cheaper per token); critical analysis stays on GPT-4o or Claude Sonnet. No manual rules to maintain: the production engine handles 14 languages natively, with an L2 MiniLM fallback for ambiguous cases.

160 RULES · 14 LANGUAGES · L1+L2 PIPELINE

SLA quality protection

Legal, compliance and medical prompts are force-routed to GPT-4o / Claude Sonnet, always. If quality drops below your threshold for three consecutive days, you get a full refund. Written in the contract, not a promise.

FORCE HIGH · ONE-CLICK ROLLBACK · SLA GUARANTEE

Semantic caching

"What is an API?" gets asked 200 times a day. Same question, same answer, zero cost. The global knowledge cache matches semantically: typos, rephrasing and multilingual variants all hit cache at $0.

FUZZY MATCH · LEVENSHTEIN < 5 · GLOBAL POOL

Auto-learn flywheel

Ambiguous prompts ("handle this for me", "fix it") enter the AMBIG queue for review; every resolution trains the engine. Engine versioning is public: v1.2 (26 rules, peer-reviewed paper) → v2.6.2 (160 rules, current production). Your savings compound monthly.

AMBIG QUEUE · v2.6.2 PRODUCTION · COMPOUNDING
Honest math

Most "save 60%" claims are fiction.
The truth is industry-dependent.

We modelled 8 representative workloads through the v2.6.2 routing engine. These are the numbers we actually saw, including the industries where Casca saves less. We publish the underperformers because honesty scales better than marketing.

IndustryBill reductionWhy this savings level
Fintech support65–75%Balance, transaction and KYC lookups dominate; almost all LOW-tier traffic
E-commerce CS60–72%Order status, returns, sizing: the same questions repeat at scale
HR helpdesk60–75%Every employee asks the same PTO / benefits / 401k questions
Insurance support55–68%Policy lookups and claim status checks are repetitive and simple
Marketing / content tools41–55%Mid-tier content generation dominates; fewer LOW, fewer HIGH
EdTech / online learning38–52%Mixed Q&A and tutoring; broad complexity distribution
B2B SaaS in-app AI30–45%Mixed reasoning workloads: analyse / draft / summarise blend
Developer tools / code gen19–31%Most queries genuinely need reasoning; less to optimise

Numbers represent modelled workloads on engine v2.6.2 against a GPT-4o flat-rate baseline ($5.00 / M tokens). Real customer bills depend on traffic mix, retry policies and overhead. The published v1.2 paper documents the foundational methodology; the v2.6.2 industry benchmark is coming.

Get my workload analysed → Read methodology & paper →

Compare

Why not just use OpenRouter, Helicone or Portkey?

Real comparison, no marketing spin. Casca is a routing engine, not an aggregator or observability layer, and we publish our research.

CapabilityOpenRouterHeliconePortkeyCasca
Auto-route by request complexity——Manual rules✓ Native
Quality SLA with monthly refund———✓ Contractual
Public technical paper———✓ Zenodo
Per-industry savings benchmark———✓ 8 verticals
Auto-learn from real traffic———✓ Flywheel
14-language native classificationEN-only routingN/ALimited✓ Built-in
Cross-customer semantic cache——✓✓ + isolation mode
OpenAI SDK compatible (drop-in)✓✓✓✓ One line
Bypass switch (≤ 5 s rollback)——Config-based✓ Env var

All comparisons based on public documentation as of May 2026. Disagree? Open an issue on GitHub and we will update the table.

Pricing

Pay for what you save.
Not what you spend.

Self-serve plans for predictable pricing, or the outcome-based tier that only charges when you save more than the subscription would have cost. All BYO-key plans: your API keys never leave your infrastructure; LLM costs are billed directly by OpenAI / Anthropic / Google and Casca charges the routing fee only.

Free
$0 / mo

10M tokens included · BYO API keys · no overage, quota-capped

For founders trying Casca on a side project or proof of concept.

  • 3-tier intelligent routing
  • Cross-customer semantic cache
  • Real-time savings dashboard
  • Community support (GitHub)
  • Bring your own OpenAI / Anthropic / Google keys
Start free →
Starter
$299 / mo

100M tokens included · BYO keys · overage $0.10 / 1M tokens routed

For Series A AI-native startups spending $5K–$30K / month on LLMs.

  • Everything in Free
  • Email support · 24h response
  • Routing analytics + alerts
  • Quality SLA (automated)
  • Pause subscription anytime
Start 60-day trial →
Growth most popular
$999 / mo

500M tokens included · BYO keys · overage $0.05 / 1M tokens routed

For Series A/B teams spending $30K–$80K / month on LLMs. The routing fee pays for itself on day one.

  • Everything in Starter
  • Slack support · same-day response
  • Custom routing rules + force-HIGH lists
  • Multi-provider failover
  • Audit logs · 90-day retention
Start 60-day trial →
Scale
From $2,499

2B tokens · or 12% of savings · two pricing options

For mid-market teams spending $80K–$300K / month on LLMs. Choose flat or outcome-based.

  • Everything in Growth
  • Outcome-based pricing available
  • Provider Pool add-on available
  • Dedicated success manager
  • Custom SLA + audit log retention
See both options ↓

Scale tier: two ways to pay

Switch anytime

Option A · Predictable

Flat subscription

$2,499 / mo

2 billion tokens included; overage $0.05 / 1M tokens. A predictable monthly bill: easy for procurement to approve, easy for finance to forecast.

  • Predictable revenue line for finance
  • No data sharing required
Option B · Aligned

Outcome-based

12% of verified savings

Connect read access to your provider billing. We verify savings against a GPT-4o flat-rate baseline and bill 12% of what you actually save. Floor: $0 if no savings. Cap: 1.5× Option A, never more than the flat plan.

  • Floor $0: zero risk if the engine underperforms
  • Cap 1.5× Option A: no surprise bills
  • Includes a monthly peer-benchmark report
Optional add-on · Scale tier and above

Casca Provider Pool: future-proof your AI stack

One integration, every major LLM provider and every credible second-tier alternative. New models are available the day they launch; when OpenAI changes pricing you switch in one click; when a new provider beats GPT-4o on price/quality, you benefit immediately. Built for teams that do not want their AI strategy tied to one vendor's roadmap.

OpenAIAnthropicGoogleGroqMistralTogether AIFireworks AICohere+ more

Pre-negotiated rates · unified DPA · auto failover · one invoice across all providers

Add-on +$499 / month (included in Enterprise) Add to Scale plan →

Enterprise

Custom annual contract · outcome-based primary

For teams spending $300K+ / month on LLMs. Outcome pricing 12–15% of verified savings (lower at higher volumes). Provider Pool included by default. Custom SLA, named-account support, private deployment, custom rate-card negotiation with providers on your behalf.

  • Provider Pool by default
  • Private deployment available
  • Quarterly business reviews
  • Setup fee $15K–$30K

Contact sales →

The math

E-commerce CS workload, $50K / month on GPT-4o

Casca routes about 60% to cheaper models, so the LLM bill drops to about $15K. Add the Growth plan: 5B tokens × $0.05 + $999 ≈ $1,250. Total Casca cost $1,250 / month; net savings $33,750 / month; ROI 27:1.

Other industries see different ratios; see the honest-math table above.

Service commitment

Designed so you can leave anytime.

The best SaaS commitment is not lock-in; it is making cancel-anytime so frictionless that customers stay because they want to. Here is exactly what we promise, in plain English.

1

60-day savings guarantee

If your verified bill reduction is under 30% after 60 days of active use, we refund the subscription fee for the trial period. No questions, no negotiation; once per customer account.

2

Cancel anytime

Cancel from your dashboard at any time. No annual lock-in on self-serve plans, no salesperson to call. Service runs to the end of the billing cycle; no refund of the unused portion. CASCA_BYPASS=true reverts traffic to your provider in under 5 seconds.

3

Pause subscription

Pause for 1–3 months at any time: zero subscription fees, routing engine inactive (traffic falls back via bypass), all settings, keys and dashboard data preserved. Resume anytime within 90 days, no penalty.

4

Automated quality SLA

Objective metrics, automatic credits, no claim required and no subjective dispute: routing accuracy under 90% on a 7-day rolling window; bypass latency over 5 seconds; p99 latency over 1.5× the provider baseline for three consecutive days. Enterprise customers receive a custom MSA with named-account commitments.

FAQ

Frequently asked

Stop burning money.
Start today.

High-touch ($50K+/month teams): send 100 anonymised real requests, we run them through the v2.6.2 engine and return a per-tier savings projection within 24 hours; no commitment, no sales call. Self-serve: create a free account, get a key, swap your base_url, start routing in under 30 minutes.

✓ Free to start ✓ No credit card ✓ Cancel anytime · reply within 24h, directly to the founder