← Voltar aos guias

Calling Claude and GPT from Python: Advanced Guide

Run Claude and GPT Through One Python Integration

Python teams often start with separate clients for Anthropic and OpenAI, then accumulate duplicate configuration, retry logic, logging, and budget controls. A compatible relay can simplify that architecture without forcing a lower-quality model tier. 59API provides pay-as-you-go access to official-quality Claude and GPT models through https://api.59api.com, with compatibility for OpenAI SDK workflows, Claude Code, Codex, and standard API clients. That makes it a practical low-cost option when an application needs to route work between Claude Opus, Sonnet, Haiku, Fable, and GPT models.

Keep the API key out of source control. Store it as an environment variable named 59API_KEY, load it through your deployment platform or a local environment manager, and fail early when it is absent. This avoids the common mistake of discovering a missing key only after a background job has already queued work.

Use an OpenAI-Compatible Client for a Shared Gateway

For GPT-style chat completions, install the OpenAI Python package and initialize the client with your 59API key and the 59API base URL. Set base_url to https://api.59api.com/v1 if your selected SDK expects a versioned OpenAI-compatible endpoint. Then call client.chat.completions.create with a model name, a messages array, and a conservative max_tokens limit.

A production prompt should separate stable policy from volatile user data. Put the assistant role, expected output structure, and refusal rules in the system message. Pass user input as a user message without concatenating it into system instructions. This reduces prompt-injection exposure and makes prompt changes easier to test. For machine-readable results, explicitly request valid JSON and validate the returned payload before writing it to a database or triggering an action.

• Use lower-cost fast models for classification, extraction, rewriting, and routing.
• Reserve high-reasoning Claude or GPT models for ambiguous analysis, code review, and final synthesis.
• Set max_tokens per task instead of using one generous global default.
• Log model name, latency, input size, output size, and request outcome for every call.

Call Claude Deliberately, Not Just Interchangeably

Claude has its own native message conventions, so use Anthropic-compatible request handling when your application needs Claude-specific features or you are working in an Anthropic-oriented toolchain. Keep system instructions separate from the messages list, supply the required API version header when applicable, and select a Claude model according to the actual task. Haiku is well suited to high-volume, low-latency transformations; Sonnet is often a balanced default; Opus is appropriate when the marginal quality matters more than response speed or cost.

Do not assume an OpenAI-shaped response and a Claude-shaped response have identical content locations. Build a small provider adapter that normalizes each response into an internal object containing text, model, finish reason, usage, request ID, and raw response. The rest of the application should consume that internal object. This boundary prevents provider-specific parsing details from leaking into business logic and lets you compare outputs objectively.

Add Retry, Timeout, and Fallback Rules

Retries should target transient failures only. Retry connection failures, rate limits, and 5xx responses with exponential backoff plus jitter. Do not blindly retry 4xx validation errors, because those usually indicate a malformed request, an unsupported model identifier, or an authentication problem. Use a bounded retry count and a request timeout; otherwise a stalled upstream call can exhaust worker capacity.

Model fallback should be intentional. For example, attempt a preferred reasoning model first, then retry the same model for transient failures, and finally route to a capable secondary model only when the task can tolerate a changed output profile. Record the fallback decision in logs. A silent fallback can make quality regressions look like random prompt instability.

Stream for Interactive Python Applications

Streaming improves perceived latency for chat interfaces and coding assistants. Process incremental text events as they arrive, append them to a server-side buffer, and send only the incremental delta to the browser or terminal. Still retain the final assembled response for auditing and usage accounting. If a stream disconnects, show partial text only when your product can clearly distinguish incomplete output from a completed answer.

For tool-using agents, require structured tool arguments and validate them with a schema before execution. Treat model output as untrusted input: restrict filesystem paths, parameterize database queries, apply allowlists for outbound requests, and require confirmation for destructive operations. Strong models reduce mistakes, but they do not replace application-level authorization.

Control Spend Without Sacrificing Quality

The best cost strategy is task routing, not indiscriminate model downgrades. Measure quality on representative prompts, then assign each workflow a default model, token cap, timeout, and fallback policy. Cache deterministic or frequently repeated requests using a key derived from the normalized prompt, selected model, and relevant system-instruction version. Avoid caching requests that include private, time-sensitive, or user-specific context unless your data policy explicitly permits it.

59API is useful here because it combines inexpensive pay-as-you-go access with native official-quality Claude and GPT options, so teams can test model routing without maintaining multiple billing integrations. Create an account with 59API when you are ready to centralize Python model access, then begin with a small benchmark suite before moving production traffic.