Reasoning Model Outputs

Some 4ALL API models provide reasoning-tier variants, allowing you to choose the reasoning strength through the model name suffix without any extra parameters:

Tier Suffixes

Suffix	Meaning	Example
-low	Low reasoning strength, faster and more economical	gpt-5.5-low, gemini-3.1-pro-low
(no suffix)	Default tier	gpt-5.5
-high	High reasoning strength, more reliable for complex problems	gpt-5.5-high, gemini-3.5-flash-high
-xhigh	Highest tier (some models)	gpt-5.5-xhigh
-thinking	Claude thinking mode	claude-opus-4-6-thinking

How to Call

Exactly the same as a normal chat request, just change the model name:

curl https://api.4allapi.com/v1/chat/completions \
  -H "Authorization: Bearer $AIPROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.5-high",
    "messages": [{"role": "user", "content": "证明根号2是无理数"}]
  }'

Output Parsing

• The reasoning process of thinking models may appear in reasoning_content (or the same field in streaming deltas), while the main text remains in content; for UI display, it is recommended to collapse the reasoning process;
• The higher the reasoning tier, the more output tokens, the greater the latency, and the higher the cost — reasoning tokens are billed as output;
• max_tokens needs to leave room for the thinking process; setting it too small can easily result in only the reasoning being output before it is truncated (finish_reason: length).

Selection Recommendations

• Use the default tier for everyday Q&A; switch to -high / -thinking for math, code debugging, and multi-step planning;
• For batch pipelines, first run everything with the low tier, then re-check low-confidence results with the high tier for the best cost efficiency.