Portkey Review: Pricing & Comparison

What problem Portkey solves

Your AI application calls the Claude API. One day Anthropic hiccups, and response time jumps from 200ms to 3 seconds. Your product monitoring has no metrics specific to LLM latency, so you don’t even know whether it’s your own code or something upstream.

Two weeks later you land an enterprise client who requires that outputs never mention competitor names. You need to write content-filtering logic in your code, then test, deploy, and maintain it.

A month after that, your boss asks where last month’s AI API spend went — broken down by model, by feature, by user segment — and you can’t answer.

Portkey (portkey.ai) exists to systematically solve exactly these kinds of problems. It isn’t a relay — it doesn’t just forward your Claude calls with a markup tacked on. Portkey is an AI Gateway: once all your model calls are routed through Portkey, you automatically get a set of engineering-grade capabilities.

Three core capabilities

1. Universal API

Portkey connects to 1600+ models, including:

• Anthropic: Claude Fable 5, Opus 4.8, Sonnet 4.6, Haiku 4.5
• OpenAI: GPT-5.5, o3, GPT-4o
• Google: Gemini 3.5 Pro/Flash
• Meta: the Llama 4 series
• Open source: DeepSeek, Qwen, Mistral, Llama…
• Plus enterprise cloud endpoints like Azure, AWS Bedrock, and Vertex AI

Switching models only requires changing one string — your SDK code stays the same:

from portkey_ai import Portkey

portkey = Portkey(
    api_key="your Portkey key",
    virtual_key="anthropic-virtual-key"  # points to Claude
)

# Switching to OpenAI only requires changing virtual_key
portkey = Portkey(
    api_key="your Portkey key",
    virtual_key="openai-virtual-key"
)

2. Guardrails

50+ built-in guardrail rules automatically check requests as they come in or responses as they go out:

Guardrail type	Example	Purpose
PII detection	Redact names/phone numbers/emails	Data privacy compliance
Content filtering	Block harmful output	Brand safety
JSON format	Enforce valid JSON output	Structured extraction
Regex validation	Custom output format	Formatting scenarios
Keyword blocklist	Block competitor names	Enterprise compliance

Guardrails are defined in a YAML config file, no code logic changes needed.

3. Observability

Every LLM call is automatically logged with:

• The full input/output content
• Latency (TTFT, total duration)
• Token usage and cost
• Metadata tags (by feature, user ID, etc.)

The dashboard provides session tracing, cost analysis, model comparison, and prompt version management. Building this kind of system yourself would take at least a couple of backend engineers several weeks.

When the free tier is enough

Portkey’s Developer free tier provides 10,000 recorded logs per month (beyond that only stops recording logs, requests are unaffected), including all core features: tracing, logging, guardrails, A-B testing.

For an early-stage AI application (DAU < 5,000), the free tier is usually enough. Wiring up Portkey the first time you write the code means that by the time you need to upgrade to the Production plan, you’ll already have accumulated enough observability data to make well-informed optimization decisions.

Portkey vs. direct connection vs. relays

Dimension	Direct to Anthropic/OpenAI	Domestic relay (e.g. ShenmaAPI)	Portkey
Price	Full price	Below full price (¥1≈$1 or less)	Full price + gateway fee
Usable in mainland China	Needs a proxy	Direct connect	Needs a proxy
Observability	None	None/basic	Complete
Guardrails	None	None	50+ rules
Model switching	Multiple SDKs	Single endpoint	Single endpoint

Portkey isn’t a fit for individual users chasing low prices, but for enterprise teams that need production reliability, compliance controls, and cost visibility, the ROI of adding a gateway layer is usually positive.

Information verified 2026-08-11. Defer to portkey.ai/pricing for free-tier quotas and paid-plan details.

Related reviews

• Mistral AI API: a European open-source AI leader, GDPR-compliant, low-priced Codestral code generation
• OAIPlus: an OpenAI-enhanced relay, clean interface, beginner-friendly, mainland direct connect
• Novita AI: 200+ open-source models, focused on image generation, low-cost global inference
• Anyscale: from the team behind the Ray framework, enterprise-grade high-concurrency open-source model deployment and fine-tuning