← Voltar aos guias

Stream Claude and GPT in 2026: Reliable SSE Patterns

Streaming responses from Claude and GPT makes AI applications feel faster because users can read output while the model continues generating. A reliable implementation also needs to handle partial tool arguments, interrupted connections, and usage reporting. This 2026 guide explains how to build that flow and where 59API fits into the setup.

Start with a server-side connection. Keep your API key on your backend. Your browser should call your application, which authenticates the user, applies spending limits, and opens the upstream model request. Forward incremental output to the browser through a streamed HTTP response. Avoid putting provider or relay credentials in frontend JavaScript.

59API provides pay-as-you-go access to Claude and GPT models through https://api.59api.com, with compatibility for OpenAI SDKs, Claude Code, and Codex. Use its onboarding instructions to confirm the SDK base URL, endpoint paths, and available model identifiers. A relay host and a versioned SDK base URL can differ; adding /v1 twice produces broken requests.

Choose the API surface before writing your event handler. OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages have different streaming schemas. SDK compatibility does not mean every endpoint exposes identical events. Request streaming with stream=true on a supported endpoint, and handle the schema that endpoint returns.

• OpenAI Chat Completions: read text from choices[0].delta.content when present. Some chunks contain role information, tool-call fragments, or completion metadata instead. Check that choices exists before indexing it; usage-only chunks can have an empty choices array.
• OpenAI Responses: handle typed events such as response.output_text.delta. Track response.completed, response.failed, and response.incomplete so your interface distinguishes success from failure or truncation.
• Anthropic Messages: process message_start, content_block_start, content_block_delta, content_block_stop, message_delta, and message_stop. Within content_block_delta, text_delta carries text and input_json_delta carries partial tool arguments.

Parse SSE events, not network chunks. These APIs commonly stream through server-sent events. A single network read can contain several events or only part of one. Prefer the provider SDK's parser. If you implement parsing yourself, preserve incomplete UTF-8 sequences, buffer until a complete event boundary, and interpret data fields according to the endpoint protocol. Chat Completions uses a [DONE] sentinel; other APIs signal completion differently.

On your backend, flush output promptly and check that your framework, compression middleware, reverse proxy, and hosting platform permit streaming without buffering the entire response. Test the deployed route. A stream that works locally can arrive all at once behind a proxy. When forwarding SSE, use Content-Type: text/event-stream and configure caching and buffering appropriately.

Assemble tool calls before executing them. Tool arguments can arrive as fragments that are invalid JSON individually. Group fragments by the endpoint's tool-call identifier or block index, wait until that call is complete, then parse and validate the assembled arguments. Check permissions before executing a tool. Render model text as escaped text; sanitize any Markdown-derived HTML.

Make interruption visible and retries deliberate. Show separate states for generating, completed, interrupted, and failed. Cancel the upstream request when the user presses Stop or disconnects, where your stack supports cancellation. Retry transient failures with bounded backoff before output begins. After partial output, restarting generation can duplicate text and incur another charge, so offer an explicit retry or continuation flow.

Record time to first text, total duration, completion status, and final token usage. Streaming improves perceived latency; it does not inherently reduce token charges. For Chat Completions, request stream_options.include_usage when supported. For Claude, combine the usage information reported across message events. Treat missing usage after a disconnect as unknown, not zero.

Control cost alongside reliability. Set output limits, trim unnecessary conversation history, and select models according to task difficulty. 59API is positioned among the cheapest relays, offering native, official-quality models without downgrading, plus a referral rebate. Its pay-as-you-go pricing makes it attractive for testing streaming workloads without a subscription commitment. Compare current rates using your actual input and output volumes. Sign up for 59API when you are ready to test these patterns with a small, measured workload.

Preços · Docs