Managed model access · Enterprise traffic control · Migration assurance

One API contract. Policy-first control. Evidence-led migrations.

Manage proprietary and open-source model access through one API, enforce spend and privacy policy before dispatch, and move workloads with evidence-backed rollout controls.

OpenAI-compatible APIHard budget limitsEvidence-gated routing
DEMO WORKSPACE · SIMULATED DATA
Overview
Routes
Models
Policies
Keys
Usage
Region policy: HK
Spend guard · organization
31.84M / 100.00M credits
31.8% settled · 218K reservedHard budget active
Support Copilot18.20M creditsOn track
Risk Review8.14M creditsOn track
Code Agents5.50M creditsWatch
Policy-governed route
support-refunds-prod
ActiveZDR requiredHK / SG
APIapi.arc-ai.example/v1
GLM-5.2 · PrivateHealthyP1
Claude · SingaporeHealthyP2
Kimi K3 · EvaluationRestrictedP3
220-case migration evaluationClaude source manifest → three candidate systems
Conditional approval
GLM-5.24 / 5Best result · rollout approval required
Kimi K33 / 5Adapter re-run
Tencent Hy32 / 5Evidence required

Model families represented across access, BYOK, and evaluation workflows

OpenAIClaudeGLMKimiQwenDeepSeek
One platform · Three layers

Start with access, control, or migration assurance.

Managed access. Runtime control. Migration evidence. Enterprise Traffic Control is the operating layer beneath every Arc.AI product.

01Managed Model Access

One API for approved model deployments.

Use familiar API contracts across proprietary and open-source models, with local billing and customer-managed provider credentials.

Unified APIModel catalogLocal billingBYOK
Explore models
03Migration Assurance

Prove the move before production.

Replay representative workflows against locked variants, review outcomes blind, and produce a decision memo with rollout and rollback thresholds.

Expert-ledPaired replayDecision memoRollout plan
Discuss an assessment
01

Pre-dispatch spend control

Reserve eligible cost before dispatch, then settle actual usage and release the remainder.

Four-level limits + reservations*
02

Evidence-gated routing

Evaluate region, retention, and provider evidence before dispatch—and fail closed when a requirement is unmet.

Fail-closed privacy policy
03

Build auditable migration evidence

Replay representative cases against pinned variants and compare quality, tool use, latency, cost, and failures.

Decision memo + controlled rollout

* Hard limits apply to eligible Arc.AI-billed calls after configuration and validation. BYOK, tools, taxes, and contracted fees are metered separately.

Pre-request spend control

Enforce spend before dispatch.

Arc.AI reserves the maximum eligible cost against applicable budgets before dispatch, then settles actual usage and releases the remainder.

ENFORCEMENT BEFORE PROVIDER DISPATCH
  • Limits cascade from organization to key
  • Retries and fallbacks require a new reservation
  • Threshold alerts explain every hard stop
01RequestOpenAI-compatible
02ReserveWallet + budget
03PolicyZDR + region
04RouteModel endpoint
05SettleUsage + evidence
Policy-aware failover

Fail over without breaking policy.

Routes fail over only across administrator-approved endpoints and never relax region, retention, or model constraints.

  • Explicit endpoint priority
  • No replay after output or side effects begin
  • Full attempt chain in every Runtime Trace
01GLM-5.2 · Hong Kong429 · retry scheduled
02Customer BYOK · SingaporeBlocked · ZDR evidence expired
03Claude · SingaporeFallback succeeded
Trace trc_01K2P9E · policy zdr-apac-v3 · metadata only
Finance operations

One auditable billing center

Reconciled
DocumentEntityAmountStatus
Usage statement · 2026-06Hong Kong28.41M creditsFinal
Payment receipt · 0718-0042Hong KongHKD 156,000Paid
增值税电子发票 · 0709-0088Mainland ChinaCNY 72,000Delivered
Usage details · 2026-06Hong KongCSVReady
Managed Model Access · Local billing

One billing view. Separate wallets by legal seller and currency.

Fund organizations with Arc.AI Credits where available, or use existing provider contracts. Legal sellers, currencies, refunds, and invoice sequences remain separate while each charge is attributed to the workspace, principal, application, key, model, and provider.

Cards*FPS / bank transfer*Enterprise invoicing*Local wallet methods*

* Payment rails, currencies, and tax documents vary by market, contracting entity, and approval. Arc.AI Credits are non-cash and non-redeemable.

Migration Assurance · Claude Migration Lab

Prove what can move off Claude—and what should not.

Evaluate a customer source manifest and pinned harness across separate GLM-5.2, Kimi K3, and Tencent Hy3 deployments. Representative cases and historical incidents are reviewed before rollout approval.

Open Claude Migration Lab
Refund eligibility & response · 220 evaluation casesClaude source manifest → three candidate systems
Evaluation only
GLM-5.24 / 5Conditional approval · shadow required
Kimi K33 / 5Tool adapter re-run required
Tencent Hy32 / 5Serving evidence required
ShadowCanary 1%Canary 5%Canary 20%Canary 50%Production 100%
ZDR requestPolicy enforced
RequestPolicy registryEligible endpoint
Prompt retentionEvidence requiredTrainingEvidence requiredRegionHong Kong / SingaporeEvidence checked2026-07-30 09:28 UTC
Evidence-based privacy

Route only when the evidence passes.

The provider registry versions privacy evidence and expires it on schedule. If no endpoint meets policy, the request fails closed instead of relaxing the requirement.

Review privacy controls
Managed Model Access · Familiar APIs

Change one URL. Keep your SDK.

Arc.AI supports familiar OpenAI and Anthropic interfaces. Each provider deployment passes capability and compatibility checks before production admission.

TypeScript
const client = new OpenAI({
  apiKey: process.env.ARC_API_KEY,
  baseURL: "https://api.arc-ai.example/v1"
});

const response = await client.responses.create({
  model: "arc/auto",
  input: "Resolve this support case",
  metadata: {
    workflow_id: "support_resolution_v3",
    cost_center: "APAC-CS"
  }
});

Reserved demonstration endpoint. It cannot receive production traffic, credentials, or customer data.

AI Access Center

Give teams governed AI access without shared provider keys.

Employees receive approved models, allowances, and setup guides while administrators manage workspace, model, budget, and expiry policies.

Open Access Center
DEVICE AUTHORIZATIONConnect OpenClaw
ZDR required
Workspace Customer OperationsProfile Support AnalystHard limit 250K credits
ONE-TIME CODE · 09:42R7KQ-X2P9No long-lived key appears in the browser or device log.

Start where risk is highest

Access models. Control production. Validate the move.

Open console Discuss an assessment