| Parameter | Surface |
OpenAI
|
Anthropic
|
Gemini
|
Exposure | Gov. Impact | What it controls |
|---|---|---|---|---|---|---|---|
| model Model | Inference | Yes | Yes | Yes | High | High | Specifies the exact model to run. Determines capability, cost, and latency. |
| temperature Sampling | Inference | Yes | Yes | Yes | High | Medium | Controls randomness of token selection and response variability. |
| top_p Sampling | Inference | Yes | Yes | Yes | High | Medium | Nucleus sampling: restricts token selection to a probability mass. |
| max_tokens Sampling | Inference | Yes | Yes | Yes | High | High | Maximum number of tokens the model may generate in one response. |
| response_format Structure | Inference | Yes | Part | Part | Partial | Medium | Constrains output format, such as strict JSON or a schema mode. |
| tools Tools | Inference | Yes | Yes | Yes | Partial | High | Declares callable external functions the model may invoke during generation. |
| reasoning_effort Reasoning | Inference | Part | Yes | Part | Contractual | High | Controls depth of internal reasoning and cost/latency trade-offs. |
| safety_settings Safety | Enterprise | Part | No | Yes | Contractual | High | Configures harm category thresholds for content filtering. |
| stream Streaming | Inference | Yes | Yes | Yes | High | Low | Streams tokens progressively rather than waiting for completion. |
| logprobs Logging | Observability | Yes | No | No | Rare | Medium | Returns token log-probabilities for uncertainty analysis. |