Portkey
Listed at https://portkey.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Free 层: 10,000 请求/月;Hobby $49/月(10万请求);Growth $249/月(100万请求);Enterprise 合同制 | per million tokens |
| Output | 请求量计费,不收 token 路由费;BYOK 免费;日志/可观测性/Caching 按订阅档位开放 | per million tokens |
🤖 Supported Models (200)
✨ Pros
- ✓🎯 AI Gateway + full-stack observability in one console: Gateway, Logs, Feedback, Guardrails
- ✓200+ models with unified OpenAI-compatible API, BYOK (Bring Your Own Key) at zero markup
- ✓Multi-provider strategies: Fallback / Load Balancing / A/B Testing / Canary
- ✓Prompt Playground + version management + multi-user review workflow
- ✓Guardrails for safety: PII redaction, jailbreak detection, keyword filtering (50+ built-in rules)
- ✓Open-source self-hosted edition (GitHub 9k+ stars, Apache 2.0), enterprise-ready
- ✓Slack/Discord/Email/WhatsApp alerts + Webhook-triggered workflows
⚠️ Cons
- ×China users need proxy to access portkey.ai console (self-hosted edition solves this)
- ×Free tier limited to 10K requests/month; Growth $249/mo more expensive than OpenRouter free
- ×Observability / Guardrails / Prompt management tightly coupled — overkill for minimal needs
- ×Learning curve: many configuration surfaces (Configs, Virtual Keys, Workspaces)
- ×Default SaaS logs stored on Portkey cloud (self-hosted edition avoids this)
🎯 Best For
Small/medium teams (2-50 people) needing AI Gateway + observability; multi-provider failover / A/B testing / cost attribution; prompt version management and team collaboration; production scenarios with Guardrails for safety/compliance
💰 Pricing & Plans
| Plan | Price | Recorded Logs | Retention | Best For |
|---|---|---|---|---|
| Free Forever | $0 | 10K / month | Logs 3 days · metrics 30 days | Prototyping, POCs, evaluating enterprise trials |
| Production (Popular) | $49 / month | 100K / month (+$9 per extra 100K) | Logs 30 days · metrics 90 days | Teams shipping LLM apps to production |
| Enterprise | Custom | 10M+ / month | Custom retention | High-volume workloads needing SSO, RBAC, data residency |
🔧 API & Developer Experience
- •Universal API: One OpenAI-compatible endpoint (`https://api.portkey.ai/v1`) that proxies 200+ models from OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI, Bedrock, Azure and more — switch providers by changing the provider string, not your code.
- •BYOK with Zero Markup: Bring Your Own Key passes tokens straight through to the upstream provider; Portkey charges only for its gateway features, so there is no per-token routing markup on model usage.
- •Routing & Resilience: Per-request provider selection with automatic fallbacks, weighted load balancing, A/B testing and canary releases across multiple providers — keeps traffic alive when one vendor rate-limits or degrades.
- •Caching: Simple and semantic caching reduce cost and latency on repeated prompts; semantic cache matches intent, not just exact strings, which the community reports cuts repetitive test calls dramatically.
- •SDKs & Streams: Official TypeScript/JavaScript and Python SDKs, plus OpenAI-compatible clients; SSE streaming passes through with low added latency for chat-style apps.
- •Guardrails: 50+ built-in safety checks including PII redaction, jailbreak and prompt-injection detection, and keyword filters, enforced before the request reaches the upstream model.
- •Key Management & Costs: Virtual keys let you mint per-app or per-user keys that map to real upstream keys, with granular budgets and rate limits — ideal for cost attribution across teams and customer projects.
🛡️ AI Gateway, Guardrails & Full-Stack Observability
Portkey's differentiator is pairing an enterprise AI gateway with full-stack observability in a single dashboard, instead of bolting a router onto some other tool. Every call you make lands in one console that combines logs, traces, request/response payloads, custom metadata and feedback, with per-request fallback, load balancing, A/B and canary routing on the same screen. That co-location is what makes Portkey practical in production: when a prompt regresses or a provider degrades, you see it in the traces and reroute in the same UI — no jumping between a gateway, a metrics tool and a prompt store. Guardrails (PII redaction, jailbreak and prompt-injection detection) are enforced inline before upstream calls, and prompt management with versions, playground and multi-user approval covers the team workflow around model changes.
🌐 China Access & Latency
Portkey is not a China-native service. The portkey.ai console and hosted gateway are reachable from mainland China but connectivity is inconsistent, so mainland teams typically report needing a proxy for stable use of the SaaS — and for latency-sensitive production they prefer the self-hosted option. Portkey ships an open-source edition (Apache 2.0, 9k+ GitHub stars) that you deploy inside your own VPC, which solves both the China access problem and the cloud-logging concern since data never leaves your infrastructure. Latency for the hosted API is naturally highest from mainland data centers and depends on your proxy path to Portkey's edge; self-hosting in a CN region removes that hop entirely. For mainland teams the practical recommendation is clear: proxy for the SaaS console, or self-host for production-grade latency and data residency.