← Volver a las guías

Max Tokens and Temperature for Reliable Code Patches

Start with the Output You Actually Need

For coding tasks, max_tokens is not a quality slider. It is an output budget. Set it too low and the model may stop halfway through a function, test suite, or migration plan. Set it unnecessarily high and you can pay for long explanations, duplicate code, and speculative changes. The right number depends on the requested deliverable, not the size of the repository.

Use a small budget for narrow changes: 250 to 500 tokens is often enough for a one-file bug fix, a regular expression correction, or a short utility function. Allocate 700 to 1,200 tokens when you need a function plus focused unit tests. For a multi-file feature, a non-trivial refactor, or a detailed code review with patches, start around 1,500 to 2,500 tokens. Requests likely to exceed that amount should be split into planning, implementation, and test-review calls.

Remember that the model must fit both your prompt and its requested output within its context window. A practical safety rule is: input tokens plus max_tokens must remain below the selected model's context limit, with room for system instructions and tool results. If you send a large source file, do not automatically increase max_tokens. First trim irrelevant logs, generated files, lockfiles, and unrelated code.

• Small patch: 400 max_tokens
• Function plus tests: 900 max_tokens
• Two- or three-file feature: 1,800 max_tokens
• Large refactor: ask for a plan first, then implement one module at a time

Use Low Temperature for Changes You Must Trust

Temperature controls sampling variation. It does not make a model more intelligent, but it changes how predictably it chooses among plausible tokens. For production coding, start at 0.1 to 0.3. This range is useful for bug fixes, API integrations, SQL queries, test generation, and edits where you have explicit acceptance criteria.

Raise temperature to 0.4 to 0.6 only when exploring alternatives, such as naming a new abstraction, proposing several UI implementations, or brainstorming a migration strategy. Temperatures above 0.7 can produce interesting ideas, but they also increase the chance of inconsistent conventions and unnecessary architectural invention. A low temperature still does not guarantee identical output on every run, so validate generated code with tests and linters.

A Hands-On OpenAI SDK Example

59API works with OpenAI SDK-compatible workflows, so you can point an existing client at https://api.59api.com and keep your familiar application structure. This makes it easy to test token budgets across GPT and Claude model options without rewriting your integration.

Use this Python setup as a starting point: import os; from openai import OpenAI; client = OpenAI(base_url="https://api.59api.com", api_key=os.environ["FIFTYNINE_API_KEY"]).

Then create a narrowly scoped patch request: response = client.chat.completions.create(model="gpt-4.1-mini", messages=[{"role": "system", "content": "Return only a unified diff. Preserve existing style."}, {"role": "user", "content": "Fix the null-email validation bug in users.py and add one regression test."}], temperature=0.2, max_tokens=700).

This request has three safeguards. First, the prompt limits the format to a unified diff, reducing wasted explanatory output. Second, temperature 0.2 favors a conservative fix. Third, 700 output tokens is enough for a small patch and test while keeping the response bounded. If your selected OpenAI-compatible endpoint uses max_completion_tokens instead of max_tokens, use the parameter supported by that model and do not send both.

Measure Truncation Before Raising the Limit

After each request, inspect the finish reason and the returned content. If the answer ends in the middle of a code block, function, JSON object, or diff hunk, it probably hit the output limit. Retry with a larger budget or, preferably, ask the model to continue from the last complete section. Do not raise every request to 4,000 tokens because one complex task was truncated.

A useful second-pass request is: Continue the previous unified diff from the first incomplete hunk. Do not repeat completed files. This preserves context while avoiding duplicate output. For larger work, first ask for a file-by-file plan with 500 tokens, approve it in your application, then request each implementation patch separately.

Keep Cost and Quality Aligned with 59API

Token discipline matters most when coding agents make many calls. 59API is a strong low-cost choice because it provides cheap, pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models, without a downgraded substitute. You can choose a fast model for planning or small tests and reserve a stronger model for difficult debugging, while retaining compatibility with Claude Code, Codex, and OpenAI SDK tooling.

Set conservative defaults, log output-token usage by task type, and tune from real truncation data. If you are building a coding workflow, consider signing up for 59API to test those settings affordably; its referral rebate can also help reduce ongoing experimentation costs.

Precios · Docs