A provider's upstream channel determines both price and stability. Understand the channel groups before judging the price tag.
Official Direct
Risk LowPrice Highest
How it works:The provider holds official API keys and forwards requests untouched to official endpoints — behavior, fields, and rate limits match the official API exactly, with full feature support.
Best for:Production workloads that demand stability and full field compatibility
Max / Pro Subscription Pools
Risk MediumPrice Moderate
How it works:Bulk official subscriptions (e.g. Claude Pro/Max at $20–$200/mo) split and resold by token. Watch out: personal-subscription rate limits, shared quota pools, and some providers only support the official client.
Best for:Budget users who tolerate occasional rate limits
Claude Code / Codex Only
Risk LowPrice Moderate
How it works:Only recognizes official CLI client requests (via User-Agent fingerprints); third-party tools like Cursor, Cline, Continue will most likely error out.
Best for:Users who only use official CLI tools
Vertex AI / AWS Bedrock
Risk MediumPrice Moderate
How it works:Claude accessed indirectly via Google Vertex AI or AWS Bedrock — the model is real, but fields and response formats differ subtly from the official Anthropic API.
Best for:Users who accept minor field differences in exchange for better pricing
Kiro
Risk MediumPrice Dirt cheap
How it works:Free Claude quota from Amazon's Kiro IDE beta (backed by AWS Bedrock). Depends on Amazon's promo strategy — can die anytime the beta ends or quotas tighten.
Best for:Bargain hunters who accept sudden discontinuation
Antigravity
Risk MediumPrice Dirt cheap
How it works:Free Gemini quota bundled with Google's AI coding IDE, same logic as Kiro — can tighten anytime with Google's promo policy.
Best for:Gemini-only users on a tight budget who accept sudden loss of access
Reverse-engineered
Risk HighPrice Cheap
How it works:Reverse-engineers the private protocol of web apps (Claude.ai, ChatGPT web), scripting Pro-account chats wrapped as an API. Streaming format and context length differ from official APIs; breaks on any redesign; long conversations get truncated or silently swapped to weaker models.
Best for:Cheap experiments — not recommended for serious projects
Per-request billing
Risk HighPrice Dirt cheap
How it works:"¥X per request" or "N-request packs" usually hide context/output length caps; overages get truncated or billed extra — looks cheap but long-context use costs a lot.
Best for:Light users doing only short single-turn Q&A