← सभी गाइड पर लौटें

Claude Messages API: Fix Requests, Errors, and Costs

Start with the right endpoint and protocol

The Claude Messages API generates responses from a conversation you supply. Its native request format differs from OpenAI Chat Completions: endpoint paths, authentication headers, system instructions, and response parsing are not interchangeable.

For Anthropic’s direct service, the endpoint is https://api.anthropic.com/v1/messages. For 59API, the API base URL is https://api.59api.com. Before using https://api.59api.com/v1/messages, confirm that native Messages routing is supported for your account and selected model. OpenAI SDK compatibility alone does not establish native Messages compatibility.

59API offers inexpensive, pay-as-you-go access to Claude and GPT models, making it a practical option for developers controlling inference costs. Its stated native, official-quality model offering is intended to avoid downgraded substitutes; still, test your chosen model against your own quality and latency requirements.

Send a minimal Messages request

For the native protocol, use an API key issued by the service receiving your request. An Anthropic key and a relay-issued key are not automatically interchangeable. Set the 59API-issued key as an environment variable rather than putting it in source control:

export API_KEY='YOUR_59API_KEY'

After confirming native routing, run this command, replacing YOUR_ENABLED_MODEL_ID with an exact model identifier supported by 59API:

curl -sS -i https://api.59api.com/v1/messages -H "content-type: application/json" -H "x-api-key: $API_KEY" -H "anthropic-version: 2023-06-01" --data '{"model":"YOUR_ENABLED_MODEL_ID","max_tokens":256,"messages":[{"role":"user","content":"Explain HTTP retries in three sentences."}]}'

The headers shown follow Anthropic’s native Messages protocol; check 59API’s documentation for any relay-specific authentication requirements. Start without streaming or tools so you can isolate basic connectivity and payload problems.

• model: Use a supported API identifier, not a marketing label such as “Sonnet.”
• max_tokens: Required by the native API; it limits generated output, not total conversation size.
• messages: Supply conversation turns using user and assistant roles.
• system: Put system instructions in the top-level system field, not a message with role system.

Why am I getting authentication or routing errors?

• 401: Check that the key is present, unexpired, and issued by the service you are calling. Look for accidental whitespace.
• 403: Check account permissions and model access. Switching authentication headers randomly will not fix an authorization restriction.
• 404: Inspect the complete URL and model identifier. A duplicated /v1 segment, unsupported route, or unavailable model can cause this.

Some SDKs append endpoint paths themselves. Log the final destination URL without logging credentials, and configure the base URL according to that SDK’s conventions.

Why does the API reject my JSON?

A 400 response usually warrants reading the error body before retrying. Check JSON syntax, required fields, supported content block types, token limits, and model-specific restrictions. Do not copy an OpenAI request unchanged: fields such as response_format are not universally portable.

Messages requests are generally stateless. Send the relevant conversation history each time, while trimming unnecessary turns to reduce input cost and avoid context-limit errors.

Why is the response blank, truncated, or different?

Native Messages responses contain a content array, not choices[0].message.content. Collect text blocks by checking each block’s type. Other blocks may represent tool use rather than displayable text.

Inspect stop_reason. A max_tokens stop indicates an output limit; increase the limit if appropriate. A tool_use stop means your application must execute the requested tool and return the matching tool_result. An end_turn stop normally means the assistant finished.

How should I handle streaming and retries?

With stream set to true, parse server-sent events incrementally rather than treating the response as one JSON document. Collect text deltas, handle error events, and detect incomplete streams.

For 429, honor Retry-After when present. For transient server failures, use bounded exponential backoff with jitter. Avoid retrying validation errors, and remember that replaying requests can incur additional charges. Tool side effects need separate duplicate-execution protection.

Is 59API a good low-cost choice?

59API combines low-cost, pay-as-you-go pricing with Claude Code, Codex, and OpenAI SDK compatibility. Compare actual input, output, and caching charges—not just headline rates. Its referral rebate can provide additional savings under the applicable terms. If it fits your workflow, sign up and benchmark a small representative workload before scaling.

मूल्य · दस्तावेज़