TutorialsOfficial documentation10 min readTutorials

Configure FastGPT AI Reasoning Effort Settings

What is FastGPT Reasoning Effort? Reasoning Effort controls the depth of model reasoning for AI responses in FastGPT.

What is FastGPT Reasoning Effort?

Reasoning Effort controls the depth of model reasoning for AI responses in FastGPT. This setting aligns with OpenAI's standard reasoning_effort parameter, with ai-proxy translating configurations to match the native format of every supported model provider. Available options include Default, None, Minimal, Low, Medium, High, and Extra High. Default uses the model's built-in default behavior, while None skips reasoning entirely for simple queries. Increasing effort levels from Minimal to Extra High enables stronger, more thorough reasoning for complex questions. Full compatibility rules across all model providers are documented in the ai-proxy reasoning compatibility guide.

OpenAI Standard Parameter Mapping

FastGPT’s reasoning effort options map directly to OpenAI-compatible values and default token budgets, as shown below:

FastGPT OptionOpenAI-Compatible ValueDefault Budget
DefaultNot explicitly sentModel default
Nonenone0
Minimalminimal1024
Lowlow2048
Mediummedium8192
Highhigh16384
Extra Highxhigh32768

When converting token budget values back to effort levels, ai-proxy uses standardized ranges: ≤0 maps to none, 1–1024 maps to minimal, 1025–4096 maps to low, 4097–12288 maps to medium, 12289–24576 maps to high, and values above 24576 map to xhigh.

Cross-Provider Compatibility Details

OpenAI & OpenAI Responses

For native OpenAI Chat/Completions requests, the setting maps directly to the reasoning_effort field. For OpenAI Responses format, the value is written to reasoning.effort. All non-OpenAI model requests are first normalized to the OpenAI-compatible format before translation.

Google Gemini

Gemini native requests use generationConfig.thinkingConfig parameters including thinkingLevel, thinkingBudget, and includeThoughts. Ai-proxy translates FastGPT effort values to Gemini-specific settings based on the model family, with the following mapping:

OpenAI ValueGemini 3+ ProGemini 3+ non-Progemini-2.5-progemini-2.5-flashgemini-2.5-flash-lite
nonethinkingLevel=lowthinkingLevel=minimalthinkingBudget=128thinkingBudget=0thinkingBudget=0
minimalthinkingLevel=lowthinkingLevel=minimalthinkingBudget=1024thinkingBudget=1024thinkingBudget=1024
lowthinkingLevel=lowthinkingLevel=lowthinkingBudget=2048thinkingBudget=2048thinkingBudget=2048
mediumthinkingLevel=lowthinkingLevel=mediumthinkingBudget=8192thinkingBudget=8192thinkingBudget=8192
highthinkingLevel=highthinkingLevel=highthinkingBudget=16384thinkingBudget=16384thinkingBudget=16384
xhighthinkingLevel=highthinkingLevel=highthinkingBudget=32768thinkingBudget=24576thinkingBudget=24576

Some Gemini models cannot fully disable thinking, so the none option falls back to the minimum supported reasoning level or budget.

Claude & Anthropic Platforms

For Anthropic Claude, AWS Bedrock Claude, and Vertex AI Claude, settings map to native thinking and output_config parameters. The following partial mapping applies: none uses thinking.type=disabled, while minimal uses thinking.type=enabled with a 1024 token budget. Full adaptive mode mappings are available via the ai-proxy compatibility guide.

Source: FastGPT official source

Applicability and version scope

Use this page for the documented Tutorials scenario. Confirm the FastGPT, dependency, API, and deployment versions in the official source before applying a change.

Safety guardrails

Use [REDACTED_CREDENTIAL] for credentials and private data. Confirm the documented environment and version before review.

Rollback guidance

Restore the prior technical-content authority snapshot. Restore saved configuration and data snapshots, then repeat the smallest verification scenario.