Portkey review: the gateway is now Prisma AIRS AI Gateway

Start with the acquisition

Every other Portkey review you will read was written before this, so it goes first.

Palo Alto Networks positions the acquired technology as the AI gateway inside Prisma AIRS, its AI security platform: a control plane that inspects AI traffic at runtime, authenticates agent identity through its Idira technology, and enforces just-in-time tool access. The GA announcement quotes sub-millisecond routing latency and 99.999% availability.

None of that makes the product bad. It makes it a different bet than it was in 2025. If you are a two-person startup picking a gateway for the next three years, weigh that; if you are an enterprise already buying Palo Alto Networks, this probably just got easier to procure.

What Portkey actually does

Four products sold as one, which is the reason it is expensive to replace and the reason people like it.

Latency is honest in the docs: Portkey estimates 20 to 40ms of total added latency versus a direct API call, offset by caching and routing. The open-source gateway claims sub-1ms of its own processing overhead; the 20 to 40ms is the network hop to Portkey's edge workers. Two different numbers measuring two different things, and vendors are rarely this clear about which is which.

from openai import OpenAI

client = OpenAI(
    api_key="PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-provider": "anthropic",
        "x-portkey-virtual-key": "your-virtual-key",
    },
)

Open source versus hosted

The open-source gateway is the strongest part of the story and the reason Portkey survives an acquisition as a technology choice. It is MIT-licensed, written in TypeScript, 12.8k stars, sub-1ms overhead, and a 122kb footprint that runs on edge runtimes.

Pricing

Two things about that table. First, the meter is logs recorded, not tokens or dollars of inference, so a cheap high-volume workload costs more here than an expensive low-volume one. Exceeding the log limit does not block requests; the logs beyond the limit simply are not recorded, which is a sane degradation but means you can silently lose observability at exactly the moment traffic spikes.

Second, none of this is your model spend. Portkey charges for the control plane; OpenAI, Anthropic, and Google still bill you separately for every token. A $49 line item on top of a $4,000 inference bill is cheap. A $49 line item on top of a $60 inference bill is not.

The candid take

What it is genuinely good at

• Guardrails. 40-plus built in, applied on input and output, is the most complete out-of-the-box set in the category. Most competitors expect you to bring your own.
• Breadth in one product. Gateway plus prompts plus observability plus governance means one vendor, one config, one dashboard.
• The open-source core. MIT, small, fast, and edge-deployable. Not a crippled community edition.
• Honest latency documentation. Publishing "20 to 40ms added" rather than only the sub-1ms internal figure is unusual.

Where it is weak

• Strategic uncertainty. A developer-tools company inside a security company. The self-serve tier is not the acquirer's business.
• Log-based pricing punishes high-volume, low-cost traffic, which is exactly the shape a cheap-model routing strategy produces.
• 30-day retention on Production is short for anything audit-adjacent.
• The feature surface is large. Prompt management and virtual keys are real lock-in; plan the exit before you adopt them, not after.
• Model counts vary across Portkey's own pages (250+, 1,600+). Treat catalogue claims as marketing and verify the specific models you need.

Questions people ask

Keep reading

Continue through LLM gateways and routing with Portkey alternatives: what to move to, sorted by the job.