The architectural distinction: developer platform versus API governance platform
Cloudflare AI Gateway is designed to sit in front of model traffic using Cloudflare's gateway configuration, REST API and related platform services. Its May 2026 REST API announcement describes a unified interface for chat completions, Responses, Anthropic Messages and a broader run endpoint. That is attractive to teams already using Workers, Cloudflare access controls and edge telemetry. It does not prove every downstream provider supports every operation or that a particular workload stays in a requested region.
Kong AI Gateway builds on Kong's API proxy and deployment choices. Its AI Proxy plugin supports provider transformations, usage recording and multiple gateway topologies; Kong's commercial offering includes managed and hybrid deployment options. This can be useful when the same platform team governs ordinary APIs and model calls. The trade-off is that plugin versions, deployment topology and paid entitlements matter more than a generic 'supports AI' label.
- Cloudflare-first fit: application team wants fast gateway setup within an existing Cloudflare stack.
- Kong-first fit: central API team already owns gateway lifecycle, policy deployment and API traffic operations.
- Neither fit is a substitute for testing how credentials, payloads, logs and failures move through your selected topology.
Routing capability: versioned flows versus plugin-mediated provider paths
Cloudflare's dynamic routing documentation describes conditional branches, percentage splits, model nodes, rate and budget limits, and deployable route versions with rollback. The same documentation explicitly limits dynamic routes to the OpenAI chat-completions request shape; an Anthropic Messages-shaped request sent to that route receives an error. Do not confuse the wider Cloudflare REST API's endpoint support with dynamic routing support for each endpoint.
Kong's AI Proxy plugin documents standardized request transformation, and version 3.10 or newer can pass certain provider-native formats without conversion. Its documentation also describes REST-style batch and file routes in newer versions, with provider-specific capability limits. That does not mean every Kong topology or version supports every endpoint. Ask for a route-by-route demonstration of your production payloads rather than a provider logo sheet.
- Test a text completion, structured response, tool call, stream interruption and one provider-native request.
- Record whether fallback changes model, response schema, price or data-processing region.
- Demand an explicit error for unsupported fields; silent omission is a procurement failure.
Governance and limits: where the counter actually lives
Cloudflare documents dynamic-route budget and request-limit nodes that can key off request metadata. Its October 2026 limits page also distinguishes the 200-requests-per-minute Unified Billing rate limit from BYOK traffic, which is not subject to that particular limit. A buyer must still verify account-level gateway limits, log handling and whether application metadata is authenticated rather than simply trusted from a caller-supplied header.
Kong's AI Rate Limiting Advanced plugin can use provider token usage or configured cost units to enforce policies. Kong documents local, cluster and Redis counter strategies with different consistency and operational implications, and identifies this advanced plugin as part of AI Gateway Enterprise. A multi-node test should prove the effective limit under concurrency, not just show that a dashboard field exists.
- Run simultaneous requests near a budget boundary and reconcile admitted versus settled usage.
- Try a missing or spoofed tenant identifier; confirm that callers cannot select another tenant's quota.
- Inspect counter consistency across replicas and behavior when the counter store is unavailable.
Price comparison: free core is not free total ownership
Cloudflare's October 15 pricing page says core AI Gateway analytics, caching and rate limiting are free. It also specifies a 5% fee on Unified Billing credit purchases, separate Workers Logs rules for newer gateways, potential Workers AI charges for guardrail inference and Logpush costs. A BYOK team should distinguish provider charges from gateway charges; a Unified Billing team should include credit-purchase fees. Pricing depends on actual feature use and can change.
Kong's published AI Management price card lists a Plus entry point from $25 per month plus usage and custom Enterprise contracts. Its table separates control planes, requests, AI model proxy counts and add-on versus included enterprise plugins. A $25 headline is therefore not a reliable monthly total for a production AI deployment. Estimate the selected topology, model count, request volume, premium plugins and staff operations before comparing it with Cloudflare.
- Use the same monthly model-token bill, request count, retention requirement and geographic footprint for both vendors.
- Add control-plane fees, request overages, log export, paid guardrails and engineering ownership separately.
- Do not claim a winner on price until you have current written quotes and a workload-specific worksheet.
Data handling, logging and failure boundaries
Both products are request intermediaries, so security review starts with the actual path. Determine where the provider credential is stored, which runtime can use it, whether prompts or responses are persisted, how long logs remain, and who can query them. Cloudflare's limits and pricing pages describe different log regimes based on when an account created its first gateway. Kong's AI Proxy documentation says request/response logging can be enabled; that is a configuration choice to inspect, not a guarantee of safe defaults.
Run a controlled provider outage and a gateway configuration failure. Check whether traffic is blocked, retried or routed elsewhere, and whether the fallback path preserves residency and authorized provider restrictions. Capture sanitized evidence: configuration version, upstream model, status, request ID and cost. Neither vendor's documentation alone establishes compliance with your contractual data-processing requirements.
- Verify that secrets and sensitive prompts are absent from raw runtime logs, not only from the dashboard.
- Prove route rollback, credential revocation and least-privilege administrator access.
- Document which component owns each failure mode and which team is paged.
A reproducible 10-point purchasing test
Before a contract, allocate ten points across five paired tests: two for exact API and payload compatibility, two for identity and credential controls, two for budget concurrency and failure behavior, two for logging and export, and two for complete monthly cost and operational ownership. Score each criterion zero for untested, one for partial with a documented limitation, or two for a witnessed pass. Keep the raw results alongside the version, plan and deployment topology tested.
If Cloudflare wins because your application already uses its developer platform and the supported request shapes fit, that is a legitimate result. If Kong wins because you need one managed policy system across APIs and AI with specific native endpoints, that is equally legitimate. AetherGate is another BYOK gateway option to evaluate using the same scorecard; do not treat any AetherGate security or cost claim as independently verified without its own test. Pair this scorecard with the existing enterprise RFP checklist.
- Require a signed-off matrix of pass, fail, not tested and workaround for every production endpoint.
- Record a total-cost worksheet and an exit/rollback rehearsal before committing traffic.
- Publish the decision rationale internally; avoid declaring an overall winner from marketing copy.