Portkey alternatives for AI apps: model access, tracing and inference margin

Choose Portkey alternatives by the gateway, governance, tracing or commercial responsibility your AI app needs to change. Compare LiteLLM, OpenRouter, Kong and Langfuse for selected operational requirements. Evaluate ShareAI when an existing app wants eligible inference paid through an approved Builder arrangement, with an allowed app surcharge and attributed earnings. A model API, a log store and an accepted customer charge are different parts of that decision.

Reviewed 23 October 2026. ShareAI publishes this comparison and operates one program discussed. Current official references support features, with actual source screenshots. The executed code is an offline synthetic ledger. No production provider migration, gateway deployment, customer charge, security certification audit or performance benchmark was performed.

Current Portkey documentation uses the Prisma AIRS AI Gateway name alongside Portkey. Its current feature comparison distinguishes managed and hybrid deployments from legacy fully air-gapped deployments. Identify the actual service, account and version before using an older alternatives article to infer new-customer availability or the scope of a selected deployment.

What Portkey currently documents

The current introduction describes a gateway and application platform with observability, guardrails and administration references. The open-source page also identifies a separate gateway implementation path. Reasons to keep it: evaluate the selected current product when those capabilities already fit the app. Replacing only the routing endpoint can leave other working controls and evidence without an owner.

The public pricing page lists Developer and Production offerings with recorded-log allowances and different retention. Its Developer description says exceeding the logging allowance does not stop requests; extra logs are not recorded. Compare those explicit units with actual inference usage. A loss of retained evidence and a runtime stop are different outcomes, even if both appear under a plan limit.

The current deployment comparison says new fully air-gapped deployments are no longer offered, while existing ones remain supported. Managed and hybrid arrangements remain separately described. Do not flatten that distinction into either a shutdown of all existing deployments or a promise that a new customer can buy every historical option. Confirm written scope for the selected account and contract.

A translation app provisions model access across client workspaces

Consider a hypothetical AI translation app that already hosts business customers and their document workflows. It has separate client workspaces, centrally managed provider credentials and a premium terminology-review model. Product wants to evaluate a permitted margin on eligible inference. Finance needs to know which client generated usage, what useful work was delivered and how an accepted customer charge becomes available earnings.

One client receives a new premium model while another must keep its restricted model list. The app also approaches its logging allowance during a busy document batch. Requests can keep producing translations while the evidence needed to answer a later dispute becomes incomplete. Compare actual model provisioning and retained records rather than treating a working endpoint as proof of a complete commercial workflow.

The exception case is a provider migration during a document review. Existing prompts reference a workspace-scoped provider slug, while administrative automation still calls an older credential-management API. A compatible inference call may keep working even though automated onboarding needs changes. Preserve the original client, provider, model and accepted price through the transition, including an explicit correction for an unusable translation.

Choose the operational or commercial layer to compare

The translation app can retain qualified operational controls while evaluating another approved inference path. That requires a real integration and a clear customer explanation. No automatically shared wallet, provider catalog, native connector or synchronized budget among the products is implied. Give each model permission, measured usage item, charge and correction an explicit owner.

Follow one document through access, delivery and release

Keep technical success and commercial completion distinct. A provider can be connected without being provisioned to a particular workspace. A document can produce a model answer without satisfying the client’s terminology requirements. A calculated margin can remain uncollected. Support, product and finance should be able to explain the same original document and its explicit adjustment.

LiteLLM: compare cost control and operating ownership

The LiteLLM custom-pricing guide and margin reference describe cost overrides, discounts and percentage or fixed-amount fees for internal billing. Pros: compare those existing controls when the translation app wants to evaluate its own gateway prices and allocations. Configurable margins are not exclusive ShareAI capabilities.

Tradeoffs: qualify the selected deployment, price lookup, effective permissions, lifecycle policy and collection. The billing guide documents a Lago integration for internal or external usage billing. That existing integration still needs trusted client attribution, agreed billing rules and actual payment evidence. A callback alone does not prove complete commercialization.

For the premium translation model, inspect a known nonzero request, a provider change and a correction. Preserve the accepted client price independently of a later global configuration change. Operating the gateway also requires updates, availability and recovery; source access does not remove those responsibilities or guarantee the same managed dashboard and support scope.

OpenRouter: compare managed access and the client payer

The OpenRouter quickstart describes managed model access. Its OAuth guide explicitly lets users authorize an app with a user-controlled key and pay from their own OpenRouter credits. Pros: compare that supported journey when the translation app wants clients to fund their own eligible inference instead of charging every request to the app owner.

Tradeoffs: qualify model availability, provider policy, authorization and the complete app commercial arrangement. The app-attribution reference concerns usage association and rankings. That alone does not establish an approved app surcharge or received settlement. User-funded inference is not unique to ShareAI; compare actual permitted pricing and financial release.

For the client document, rehearse an exhausted user balance and a failed final review after earlier processing succeeds. Preserve the expected payer and trusted workspace identity. A model-access change should not silently move a customer-funded request onto the app owner’s account or imply that every document-processing and support responsibility moved with the credential.

Kong: compare policy, entitlements and usage commercialization

The Kong AI monetization page describes a gateway and commercialization path. Its current entitlement reference links managed rate-card entitlements with separate gateway enforcement. Pros: compare the selected services when the translation app wants product access and measured usage evaluated together.

Tradeoffs: qualify the actual service and deployment combination, policy enforcement, metering and payment integration. The location of a gateway and the operating model of its managed billing service are separate architectural facts. A product-family comparison does not establish that every hosted capability is included in one gateway installation.

For the translation client, preserve which purchased right permits the premium model and when that right becomes effective. Compare an in-progress review, a plan change and an explicit correction. A calculated balance and an enforcement component can both exist while the complete application path still needs qualification before customers depend on it.

Langfuse: compare tracing and translation-quality evidence

The Langfuse observability guide describes application traces spanning model, tool, retrieval and custom work, with evaluation-related workflows. Pros: compare this scope when the translation app needs to understand why a terminology review changed or failed. A request-cost chart alone cannot explain every application decision that affected the delivered document.

Tradeoffs: qualify instrumentation, retained data, selected deployment and platform costs. A tracing overview does not by itself establish a replacement model-routing or customer-collection service. Ensure the actual integration sends the observations required to reconstruct the document. A quality score also differs from an accepted client price, invoice or settlement receipt.

For the disputed translation, retain the original input, model output and the reason a correction was accepted. Link the revised result without erasing the earlier record. Compare actual retained history and the complete task relationship rather than assuming a newly instrumented dashboard can recover logs that the previous service never recorded.

ShareAI: compare approved app margin and settlement

The ShareAI Builder Program concerns an app that already exists with its own users. Eligible AI traffic routes through ShareAI, the Builder configures a permitted surcharge or margin, and the end customer pays ShareAI for that inference. App-attributed earnings follow the program. Pros: evaluate this approved payment and attribution journey for the translation app’s eligible model work.

Tradeoffs: app approval, eligible traffic, model access and account terms matter. Validated company Builders have a 45-day review period before payout. Individuals currently accumulate inference value for ShareAI use, without cash-out or account transfers. The app remains built, hosted, distributed and maintained elsewhere. Its document storage, other subscriptions, support and financial duties remain explicit responsibilities.

Providing infrastructure or a custom endpoint belongs to applicable Provider or Prosumer onboarding, separately from Builder attribution. Custom prices, margins and endpoint control exist elsewhere. Compare actual acceptance, routing, payer and release. The Builder Program does not establish a universal governance suite, a general merchant-of-record service or automatic migration of Portkey provider identities and policies.

Separate inference compatibility from management-API migration

The Model Catalog upgrade guide describes organization-level integrations provisioned to workspaces, with model access and limits. It says existing provider slugs and inference references are retained, while credential-management automation using the older Virtual Key API needs migration to the new administrative APIs. It also identifies account enablement. Confirm the selected organization’s actual state rather than assume every account is already identical.

For the translation app, qualify both paths: an existing document request and the automation that provisions the next client. A working old inference reference does not prove that administrative onboarding still behaves correctly. Preserve the approved client and model mapping. Rehearse restrictive access and a revoked provider so a successful migration cannot silently broaden another workspace’s rights.

Compare the actual data path and retained commercial evidence

The hybrid architecture reference places the data plane in the customer environment and the control plane with Portkey, with configuration synchronization, metrics and selectable log handling. That is a defined architecture, not a blanket claim that no information can leave a network. Qualify the actual model-provider route, log option and administrative access alongside the deployment contract.

Preserve recorded-log and metrics retention independently from the accepted customer charge. Confirm which public pricing or enterprise contract applies to the current product and account. A platform allowance can affect later evidence even when inference keeps working. Moving hosts or buying another plan cannot recreate original records that were never retained.

For arithmetic only, 1000 USD of inference plus a hypothetical 15% surcharge produces 150 USD gross surcharge and 1150 USD total. These synthetic amounts are not Portkey or ShareAI fees, received earnings, net profit or a tax result. Actual observability, infrastructure and provider costs, corrections and eligibility affect the business outcome.

Check arithmetic before qualifying access and collection

The executed fixture below deduplicates identical activity, rejects a conflicting duplicate and records an explicit adjustment in integer USD cents. It does not provision a model, send a translation, deploy a gateway or collect payment. Those outcomes require actual application and commercial evidence around the calculation.

import assert from "node:assert/strict";
const records = [
  {id: "usage-a", customer: "demo", cents: 60000n, kind: "usage"},
  {id: "usage-b", customer: "demo", cents: 50000n, kind: "usage"},
  {id: "usage-a", customer: "demo", cents: 60000n, kind: "usage"},
  {id: "adjustment-b", customer: "demo", cents: -10000n, kind: "adjustment"}
];
function netCents(rows) {
  const seen = new Map();
  let total = 0n;
  for (const row of rows) {
    assert.equal(row.customer, "demo");
    const fingerprint = [row.customer, row.cents.toString(), row.kind].join("|");
    if (seen.has(row.id)) {
      if (seen.get(row.id) !== fingerprint) throw new Error("Conflicting duplicate");
      continue;
    }
    if (row.cents < 0n && row.kind !== "adjustment") throw new Error("Invalid sign");
    seen.set(row.id, fingerprint);
    total += row.cents;
  }
  return total;
}
const base = netCents(records);
assert.equal(base, 100000n);
assert.throws(() => netCents([...records,
  {id: "usage-a", customer: "demo", cents: 1n, kind: "usage"}
]));
const numerator = base * 1500n; // Hypothetical 15 percent in basis points.
if (numerator % 10000n !== 0n) throw new Error("Define contractual rounding first");
const surcharge = numerator / 10000n;
assert.equal(surcharge, 15000n);
const usd = cents => `${cents / 100n}.${(cents % 100n).toString().padStart(2, "0")}`;
console.log(JSON.stringify({baseUSD: usd(base), grossSurchargeUSD: usd(surcharge),
  totalUSD: usd(base + surcharge), paidReceiptVerified: false}));

Observed output is 1000 USD illustrative inference, 150 USD gross surcharge and 1150 USD total, with no verified paid receipt. Preserve original client identity, actual model work, accepted terms and corrections in production. A correct sum cannot establish retained tracing, authorized premium access or available earnings.

Move one client and its administrative exception first

• Identify actual service, deployment, account enablement and retained-history terms.
• Select the gateway, tracing or commercial responsibility that needs to change.
• Map trusted client, provider, model and accepted-price identities.
• Rehearse existing inference, new-client provisioning, revoked access and correction.
• Verify actual service and payment evidence without duplicate live collection.
• Retain original records and rollback until the complete customer journey is qualified.

Keep the selected Portkey implementation when it fits. Compare a replacement by the full client workflow, including provider provisioning, useful translation, retained evidence and payment. Current product naming and deployment boundaries deserve explicit verification; they do not justify invented shutdown, roadmap or universal feature-parity claims.

Frequently asked questions