Actualyze vs Portkey
Actualyze. Personal data masked from the first request, inside the most complete AI governance platform.
Portkey is an AI gateway, now sold as Prisma AIRS AI Gateway, and Actualyze is the AI gateway and AI governance platform for the enterprise. With Actualyze, your rules reach every model, from public APIs to servers inside your network, with no inbound port.
Competitor facts checked against Portkey's documentation on .
In short
- Actualyze. A default guardrail nobody can delete masks from the start. By default, a failed Portkey check logs and lets the request pass.
- Actualyze. Servers inside your network join through an outbound-only relay. On Portkey's SaaS gateway, custom model hosts must be public URLs.
- Actualyze. Routing reads each request, with no metadata to tag. A Portkey config routes on metadata, request parameters, or URL path.
Actualyze
Your organization
policy from Team Owners and Budget Managers
Inference
- public APIs
- brokers
- GPU clouds
- your own servers
Portkey
Portkey platform
policy from integration settings
Inference
- supported providers
- custom hosts
Side by side
Portkey is an AI gateway that Palo Alto Networks acquired in May 2026 and now sells as Prisma AIRS AI Gateway.
| Actualyze | Portkey | |
|---|---|---|
| Pricing and routing | ||
| Where cost figures come from | Priced per model and provider. Unknown cost is never $0 | A central price list that updates automatically A model with no tracked price shows a cost of $0.00. |
| One model, many providers | One model through any provider the caller can use | A model alias can load-balance or fail over across providers Request logs show the provider model that served, not the alias. |
| Routing across models | Sizes each request, picks the cheapest model strong enough A Virtual Model, one name over several models, picks only among models the caller can already use. | Configs route on metadata, request parameters, or URL path |
| Provider prompt caches | Stays on the cached model unless moving saves money | Sticky load balancing keeps a user or session on one target Opt-in per config, for an hour by default. Routing tied to provider prompt caches isn't documented. |
| Access and spend | ||
| Who approves access and spend | Each model names its own approver. Budget Managers own spend | Each integration lists the workspaces that may use it For each workspace, admins set budget limits, rate limits, and model access. |
| Budgets and limits | Budgets nest. A call past the limit is rejected | Cost or token limits per workspace, set on each integration |
| Chargeback and the month | Finance closes the month on one statement | Cost analytics grouped by AI service, model, or metadata |
| Protection | ||
| On from the first request | Personal data masked from the first request The default guardrail masks personal data in prompts and responses, and nobody can delete it. | A guardrail runs once attached, and only logs by default A failed check still lets the request through. |
| Weakening protection | Protection can't be quietly switched off | A workspace can be exempted from org guardrails by API A written reason or an alert for weakening a guardrail isn't documented. |
| Records and reach | ||
| Audit record | Each routed call records which model answered and why The record can't be edited, and it names the model the app asked for. | Audit logs of admin actions, kept indefinitely |
| What the controls reach | Any LLM, anywhere, behind an OpenAI- or Anthropic-style API Your own servers too, through a relay you run that only connects outward. | On SaaS, custom model hosts must be publicly reachable The SaaS gateway doesn't route to private or internal network addresses. |
| Bringing a model online | Each model goes live only after its own approval | Admins allowlist the models each integration exposes |
| Deployment | ||
| Deployment | Hosted today. On-prem deployment coming in 2027 | SaaS (AI Gateway Cloud) or a hybrid deployment |
Actualyze. Personal data is masked from the first request.
Actualyze
your app sends
Refund jane.doe@example.com for order 4471.
the provider receives
Refund [REDACTED:EMAIL_ADDRESSES] for order 4471.
guardrailthe default one, on since the workspace was created
The provider gets a prompt Actualyze has already masked.
Portkey
your app sends
Refund jane.doe@example.com for order 4471.
the provider receives
Refund jane.doe@example.com for order 4471.
guardrailattached to a config, the check failed and was logged
Portkey passes the prompt to the provider and logs the check.
Actualyze protects every new workspace before the first prompt leaves.
Masking starts the day an Actualyze workspace does, through a default guardrail that catches personal data such as email addresses and card numbers. It masks the prompt before the provider sees it and the response before your app does, with nothing to attach or configure. Nobody can delete it, and switching it off takes a written reason that every Workspace Admin sees.
A Portkey guardrail runs once attached, and only logs by default.
A Portkey guardrail checks either the input or the output of requests whose config includes it. By default, a failed Portkey check still lets the request through. Portkey organization owners can enforce guardrails on every request, and a workspace can be exempted through the Admin API.
Actualyze. Any model, anywhere, under one set of rules.
Actualyze
Actualyze governs it like a public API, with no inbound port.
Portkey
Portkey's SaaS gateway needs a public URL for it.
Actualyze governs the servers inside your network as easily as any public API.
The Actualyze Inference Relay is a small component you run in your private network, VPC, or data center. The relay connects outbound only, so you open no inbound firewall port and deploy no gateway in each network. The vLLM and Ollama servers you already run then get the same approvals, budgets, masking, and audit trail as OpenAI or Anthropic, all from the hosted platform.
On Portkey's SaaS gateway, custom model hosts must be public.
Portkey connects to the providers it supports, and a Portkey custom model can point at its own host URL. On the SaaS gateway, Portkey doesn't route to private or internal network addresses.
Actualyze. The best model routing in the field.
Actualyze
- Model A$0.15 / Mnot strong enough
- Model B$0.60 / Mstrong enough
- Model C$3.00 / Mstrong enough
- OpenAI
- Azure
- Anthropicstandby
- Bedrockserving
- Vertexstandby
Actualyze answers with the cheapest model strong enough.
Portkey
- Model A
- Model Bweighted load balance
- Model C
- OpenAI
- Anthropic
- Vertex AI
- Bedrock
Portkey returns the answer from that target.
Actualyze sizes the request itself, so your app sends nothing extra to be routed well.
An Actualyze Virtual Model is one name for several models. A routing model we train reads what each request needs, and Actualyze picks the cheapest model strong enough, with no metadata to tag and no condition to maintain. Every model behind the name runs over the provider routes the caller can use, and when one route fails, a standby route takes the call.
A Portkey config routes on conditions you set.
A Portkey config can route on metadata, request parameters, or URL path. Portkey can also split traffic across targets by weight or fall back down an ordered list. Each request is answered by one target.
Actualyze. Priced at what you really pay, down to the call.
Actualyze
modelnew-model-preview
througha GPU cloud
tokens1.4M
cost
unknown, not $0
statusserved, and flagged to FinOps
Actualyze serves it from day one and never books it as $0.
Portkey
modelnew-model-preview
througha GPU cloud
tokens1.4M
cost
$0.00
budgetsnot counted toward budget limits
Portkey shows a cost of $0.00 for it.
In Actualyze, an unknown cost is never $0, and finance sees every dollar at risk.
Actualyze prices calls at your negotiated rate or the list price, with a separate price for each provider, so one model through Anthropic, Google, and Bedrock carries three. When a team starts using a model with no price anywhere, Actualyze serves the call, marks it unpriced, and shows finance the exposure until someone confirms a rate. Nothing slips into the month's totals as free.
Portkey prices calls from a central list that updates automatically.
Portkey admins can override a model's input and output rates. A Portkey model alias can spread traffic across Anthropic, Vertex AI, and Bedrock, or fail over between them. A model Portkey has no price for shows a cost of $0.00.
Six questions every AI gateway should answer live
If your AI gateway can't tell you who approved a model or what last month cost each team, it's just moving traffic. Put these questions to every vendor on your list, and ask to see each answer working in the product. Actualyze answers every one today, and each comes with a check you can run in your own trial.
Who approves a team's access to each model?
The leading delegation and approval system for AI spend starts here. In Actualyze, each model names its own approver.
Request a model that needs approval as a team member, and check who receives the request and where the decision is recorded.
What decides which model answers each request?
In Actualyze, a routing model we train sizes each request and picks the cheapest model strong enough, the best model routing in the field.
Send an easy question and a hard one to one Virtual Model, then see which model each was routed to, and why.
What happens to a call when its budget runs out?
Actualyze rejects it before any model is contacted, so a budget is a real limit and never a surprise on the bill.
Set a small test budget and confirm the next request is rejected once it reaches its limit.
Does an unknown cost ever show up as $0?
An unknown cost is never $0 in Actualyze, which marks it unpriced and flags it for FinOps, so no spend hides behind a zero.
Connect a third-party endpoint with no catalog price, call it, and check the call reads unpriced, not $0.
What is protected before anyone writes a policy?
Personal data is masked from the first request in Actualyze, in prompts and responses, by a default guardrail nobody can delete.
Open a new workspace's guardrails before writing any policy, and confirm the default guardrail is already enforcing.
Can finance close the month on AI spend?
Finance closes the month on AI spend with one Actualyze statement, a row for each budget, reviewed and then settled.
Generate a statement for a month with traffic, then move it from draft to reviewed to settled.
Common questions
What is the difference between Actualyze and Portkey?
Does Actualyze routing keep our prompt-cache savings?
Can Actualyze stream its audit trail to our SIEM?
Sources
24 pages, checked
- Prisma AIRS AI Gateway model pricing and cost management
- Portkey combining routing strategies
- Prisma AIRS AI Gateway conditional routing
- Prisma AIRS AI Gateway load balancing
- Prisma AIRS AI Gateway workspace provisioning
- Prisma AIRS AI Gateway budget limits
- Prisma AIRS AI Gateway analytics
- Prisma AIRS AI Gateway guardrails
- Prisma AIRS AI Gateway enforcing org-level guardrails
- Prisma AIRS AI Gateway audit logs
- Prisma AIRS AI Gateway custom hosts
- Prisma AIRS AI Gateway AI model provisioning
- Portkey home page
- Palo Alto Networks completes acquisition of Portkey
- Prisma AIRS AI Gateway simple setup
- Prisma AIRS AI Gateway fallbacks
- Prisma AIRS AI Gateway overriding model details
- Prisma AIRS AI Gateway workspace budget and rate limits
- Prisma AIRS AI Gateway integrations
- Actualyze access, budgets, approvals, and cost accounting
- Actualyze guardrails and audit
- Actualyze Virtual Models and routing
- Actualyze endpoints and integrations
- Actualyze pricing
Portkey, Palo Alto Networks, Prisma AIRS, and other product names mentioned here are trademarks of their respective owners. Actualyze is not affiliated with or endorsed by them.
More comparisons
Actualyze vs TrueFoundry
Actualyze. Every model and every budget gets its own approver, and finance closes the month on one statement.
Actualyze vs LiteLLM
Actualyze. The whole package for governing AI, with masking from the first request and no unknown cost as $0.
Actualyze vs Kong AI Gateway
Actualyze. Approvals that scale to thousands of people, and personal data masked from the first request.
Actualyze and OpenRouter
Actualyze. Govern AI wherever you buy inference. Keep OpenRouter as one source among many.
Actualyze and Vercel AI Gateway
Actualyze. Keep Vercel AI Gateway. Governance shouldn't depend on whose inference you buy.
Actualyze and Cloudflare AI Gateway
Actualyze. Keep Cloudflare AI Gateway, and give every AI dollar an owner from day one.
Give every model an approver and every call a budget.
Connect the providers you already use and test every claim on this page against your own traffic. Or book a demo. We'll walk you through it, and send the Virtual Model engine whitepaper if routing is on your list.