Allow choosing max_tokens or max_completion_tokens for OpenAI-compatible requests #1133

Problem

ChatGPTBox's OpenAI-compatible Chat Completions client always sends a token limit, but there is no setting to select the request field name (max_tokens versus max_completion_tokens).

This breaks custom API endpoints and models that only accept max_completion_tokens, even when the model is configured under OpenAI (API) with a custom API URL. The Max Response Token Length slider only changes the value, not the field name.

Example error:

{
  "error": {
    "message": "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.",
    "type": "invalid_request_error",
    "param": "max_tokens",
    "code": "unsupported_parameter"
  }
}

Why the current automatic handling misses valid configurations

• src/services/apis/openai-token-params.mjs uses a provider + model-name regex to select max_completion_tokens. Unrecognized model IDs (for example, o3) fall back to max_tokens, even on the built-in OpenAI provider.
• For custom providers, provider !== 'openai' so a model ID such as gpt-5.2 also falls back to max_tokens, regardless of what the upstream API accepts.
• Using Advanced → API Url → Custom OpenAI API Url does not solve requests that use a model name outside the built-in regex.
• Updating model-name recognition alone cannot cover proxies, custom aliases, and future models.

The issue is visible in the v2.7.1 release and the current source as of 2026-11-01. Reported with Firefox and a configured OpenAI-compatible endpoint (the user's exact provider/model ID is not available for this report).

Desired behavior

Add a setting under Advanced → API Params, for OpenAI-compatible Chat Completions requests:

• Auto (default): preserve the existing provider/model detection.
• max_tokens: explicitly send the legacy field.
• max_completion_tokens: explicitly send the newer field.

Keep the current token-count slider and existing /v1/completions handling unchanged. The chosen field should not be duplicated by extra request-body parameters.

This allows people to use OpenAI-compatible proxies and nonstandard model IDs without modifying or rebuilding the extension.

Related

• Please add an option that lets us use "max_completion_tokens" instead of "max_tokens" #908 (closed): initial GPT-5/5.1 parameter support; automatic handling was improved, but there is still no manual override.
• Support OpenAI reasoning models via OpenAI API #792 (open): broader reasoning-model support.

Would a small backward-compatible, opt-in override be welcome?