参考

模型目录

你可以作为 model 字段传入的每个公开别名。价格以每 100 万 token 的美元计价，与 OpenAI、Anthropic 和 Google 在其价格页发布的单位一致。中转计费按 token 粒度计量，界面金额以美元显示，钱包按美元分扣款。

165 个模型在线。 你也可以在开发者控制台中浏览此列表。

Interchangeable model groups

Aliases in the same group are automatic substitutes when every route of the requested model is down (unless you send allow_substitution: false).

ds-v4-flash: deepseek-v4-flash ⇄ volc-deepseek-v4-flash

ds-v4-pro: deepseek-v4-pro ⇄ volc-deepseek-v4-pro

gpt5-flagship: gpt-5.4 ⇄ gpt-5.5

路由与故障转移

你始终传平台别名（例如 gpt-5.2），而不是上游特定的 id（例如 openai/gpt-5.2）。在幕后：

1. 每个别名映射到一个或多个按优先级排序的上游路由。
2. 每个路由从按提供方划分的凭证池中取用凭证。如果某个凭证被反复拒绝，它会被冷却 60 秒，下一个请求改用同组的其他凭证。
3. 单个可重试的上游错误（5xx、401/403 token 问题、429 速率限制）会触发中转自动重试下一个候选——你的客户端只会看到最终结果。
4. 如果每个候选都耗尽，你会得到一个明确的 503 和清晰的消息，而不是一个卡住的请求。

关于每个状态码的含义及如何应对，参见常见错误。

Automatic model routing (auto)

Send model: "auto" and the platform picks a concrete model for you, balancing cost and latency. Three intent variants steer the choice: auto:cheap (lowest price), auto:fast (lowest latency) and auto:quality (highest quality tier).

curl https://chinzy.com/v1/chat/completions \
  -H "Authorization: Bearer tsk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto:cheap",
    "messages": [{"role": "user", "content": "Say hi in one sentence"}]
  }'

• Server switch required. Auto routing is a server-side feature flag (AUTO_ROUTING_ENABLED). When it is off, auto requests fail with 400 auto_routing_disabled — fall back to a concrete alias.
• Text chat only. Requests with images or tool definitions must name a concrete model; auto returns 400 for them.
• The billed model is always reported. Check the model field of the response, and the autobadge next to each call in the console usage page's recent-calls list.
• Candidate set = the Auto column above. A model marked ✓ eligible can be picked by the auto router (text + enabled + priced + at least one active route).

Cross-model failover & allow_substitution

When every route of the model you asked for is down, the platform automatically substitutes an equivalent model from the same interchangeable group (see the groups listed above). The response and your usage page always report the alias actually billed, with the original request preserved as the source annotation.

allow_substitution is a platform-reserved request field: it defaults to true, and only an explicit false opts out of substitution. The field is stripped before the request is forwarded, so it never leaks to the upstream provider. It can be sent at the top level of the request body on both the OpenAI and the Anthropic protocol.

Protocol matrix