Services focused on the Anthropic Claude line, covering Opus, Sonnet and Haiku in full, and mostly compatible with coding clients such as Claude Code and Cline.
CCFly relays AI model APIs from inside mainland China, so requests do not have to leave the regi…
A globally hosted relay that fronts a range of AI model APIs. Claude and other families are avai…
GitHub (langfuse/langfuse) is an overseas AI API relay and aggregation service. It covers the Cl…
Claude is the most in-demand family in the relay ecosystem, and the driver is not chat — it is Claude Code. That command-line tool turned "a few million tokens a day" into a normal workload, which makes unit price and stability unusually sensitive.
The dividing line for these services is whether they can actually run Claude Code. An OpenAI-compatible /v1/chat/completions endpoint is not enough: Claude Code speaks Anthropic's native protocol and needs /v1/messages as well. That is why some services work fine in a browser chat and fail the moment you point Claude Code at them.
The second dividing line is whether prompt caching is passed through properly. Claude Code re-sends the same context constantly, and the official price for a cache hit is roughly a tenth of a normal input token. If the relay strips the cache fields in the middle, your bill can come out several times higher while the headline rate looks unchanged — something you cannot spot on a marketing page, only by running it for a while and reconciling the charges.
It is also worth checking that Opus is genuinely available. Some services advertise "the full Claude line" while quietly downgrading Opus to Sonnet, or returning an error, because Opus costs them more. Our API probe records what /v1/models returns, which is a reasonable starting point for cross-checking.
The picks below are computed live from the current measurements, not fixed conclusions — change the data and the picks change with it. Each states its basis, and you can open the detail page to check it yourself.
The top 3 services by overall score in this category. 1 of them are confirmed to expose a /v1/models endpoint — that figure comes from probing, not from the services' own claims.
| Service | Region | Uptime | Latency | Models | Payment | API |
|---|---|---|---|---|---|---|
|
CCFly
ccfly.codes
|
China direct | 100% | 605 ms | — | USDT / crypto | Has /v1/models |
|
GitHub (langfuse/langfuse)
github.com
|
Global / overseas | 72.13% | 1045 ms | — | — | Unconfirmed |
|
codesome.ai
codesome.ai
|
Global / overseas | 100% | 622 ms | — | — | Unconfirmed |
The questions we see most often about this category. The guidance applies to the category as a whole, not to any single service.
Most likely the service does not implement Anthropic's native protocol. Claude Code requests /v1/messages, not the OpenAI-style /v1/chat/completions. To check: set base_url to the service and send a single request to /v1/messages — a 404 or 405 means it is not implemented.
The most common reason is that prompt caching is not passed through. Cache hits in Claude Code save roughly nine tenths of the repeated input cost; if the relay does not support them, you pay full price for every re-sent context. Services like this look cheap per unit and can still cost more in total than a pricier service that supports caching.
Cross-check in three directions: request /v1/models and see whether Opus appears in the list; ask a hard question that only a flagship model answers correctly and compare with the official channel; and watch latency — Opus responds noticeably slower than Haiku, so a model labelled Opus that answers at Haiku speed is worth suspecting.
Yes. Claude Code fires off several sub-task requests in parallel, so instantaneous concurrency is far higher than in browser chat. Some services set a low per-account ceiling, which shows up as a burst of mid-task failures rather than a clean rate-limit error. Run one complete medium-sized project against the service first and watch the error rate.
This page shows you what exists; the pages below show you who is more stable, faster and cheaper. Read together, they tell you more than either alone.