AI API Relay Channel Types Explained: Max, Reverse-Engineered, Kiro, Official

Highest price

The relay holds its own official API key, and your request goes straight through to Anthropic, OpenAI, or Google's official endpoint unmodified — same behavior, same fields, same rate limits as going direct, and the fullest feature set. Because it's billed at official rates, this is typically the priciest tier across every channel type.

Best for: Best for production workloads where response consistency and field compatibility really matter.

Mid-to-low

The relay bulk-buys official subscription plans (like Claude Max or Pro, typically around $200/month tiers), pools a batch of accounts, then splits and resells the usage by token. The underlying accounts are genuine paid subscriptions, so this isn't inherently shady — but it runs into the rate limits the provider sets for individual subscribers, and multiple users can be sharing the same quota pool. Some providers here only work through the official client (e.g. only official Claude Code), so third-party tools can't connect.

Best for: Good for budget-conscious users who can tolerate occasional throttling and happen to use a supported client.

Depends on the subscription

These channels only recognize requests from the official CLI client (identified by signals like the User-Agent header) — Cursor, Cline, Continue, and other third-party editors will most likely error out or get rejected. The channel itself isn't especially risky; the problem is that it locks you into a narrow set of supported clients.

Best for: Fits people who stick to the official command-line tool and don't need third-party editor plugins.

Mid-range

Calls Claude indirectly through Google Cloud's Vertex AI or AWS Bedrock. It's genuinely the Claude model underneath, but it goes through the cloud provider's own request structure — fields, response format, and some edge-case parameters or metadata don't line up perfectly with Anthropic's native API.

Best for: Fits users who can tolerate minor field differences in exchange for pricing somewhat below straight official pass-through.

Extremely cheap

Rides the free Claude quota Amazon has been giving away during the public beta of its Kiro IDE — underneath, it's still running through AWS Bedrock's Claude. Pricing can be nearly free, but this entire supply depends on Amazon's own promotional strategy: the beta ending, quotas tightening, or the rules changing without notice can all make it disappear.

Best for: For users purely chasing the lowest price who can accept the quota being cut or pulled with no warning.

Extremely cheap

Same idea as Kiro, just riding a different vendor's promotion — in this case, the free Gemini quota bundled with Google's AI coding IDE. The logic is nearly identical: it's cheap because it's living off a promotional window, and the risk is that window closing at any time.

Best for: For Gemini-only users who are extremely price-sensitive and can tolerate it going away without notice.

Cheap

Skips the paid API entirely and reverse-engineers the private communication protocol behind the web chat interface (Claude.ai, ChatGPT's web app) — an automated script logs into a Pro account and drives the web chat, then repackages the output as an API response. Because it completely bypasses official billing, pricing can go very low, but details like stop_reason, streaming format, and context-length limits don't match the official API, it can break entirely whenever the provider changes their web app, and long conversations are more prone to unexplained truncation or silent downgrades to a weaker model.

Best for: Fine for low-stakes experimentation where you can tolerate instability and sudden outages — not recommended for anything you're shipping.

Looks extremely cheap

Instead of billing by token, this is priced as "$X per request" or bundled request packs. The headline price looks tempting, but nearly every pay-per-request plan hides a context-length or output-length cap — go over it and you either get truncated or charged extra. Long conversations and long-document processing are exactly where this hidden cost sneaks up on you, and the real effective price can end up far higher than it looks.

Best for: Best for short, simple, single-turn Q&A with no long context involved.