Temperature Calibrator
Understand how the Temperature parameter changes LLM outputs.
Use presets below for common LLM tasks instead of tuning blindly.
0.0 = deterministic, 2.0 = extreme randomness
0.70Creative
typed entry (0–2)
0.01.02.0
Active preset: GPT-4o Default
0.0 = most compact, 1.0 = full vocabulary pool
1.00Open
typed entry (0–1)
0.00.51.0
Advanced sampling parameterstop-k · min-p · penalties
Top-k 00 = off
Min-p 0.000 = off
Frequency penalty 0.00−2..2
Presence penalty 0.00−2..2
Model defaults
Change temperature OR top-p, not both. o-series ignore temperature.
Sampling visualizersynthetic distribution after your settings
Current settings are within common production guardrails. Adjust one setting at a time for clearer behavior changes.
Effect on Output
Creative and varied. Better for brainstorming, brand copy, and storytelling. At top-p=1.0, the full distribution remains available.
Temperature strengthControls how sharp or broad the token probabilities are before sampling.
Top-p strengthCuts the head of high-confidence candidates and keeps only the smallest set that sums to p.
Copy-ready OpenAI-compatible parameters
Provider snippets
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
temperature=0.70,
top_p=1.00,
frequency_penalty=0.00,
presence_penalty=0.00,
)URL hash and preset JSON are 100% client-side — no accounts, no backend, nothing leaves your browser.
Rule-of-thumb heuristics
0.0 — Greedy / deterministic
Code, math, extraction, structured output.
≤ 0.3 — Precise
Factual Q&A, classification, function calling.
0.7 — Balanced
General chat, summarization, standard writing.
0.9–1.0 — Creative
Brainstorming, brand copy, storytelling.
≥ 1.2 — Exploratory
High novelty; expect repetition and drift.
top_p 1.0 — No pruning
Full distribution; pair with temperature.
top_p < 0.5 — Narrow
Tighter, safer outputs; often better than high temp.
One knob — At a time
Change temperature OR top-p, not both (OpenAI guidance).
Provider reference
| Provider | temp default | range | top_p | Other params | Deterministic |
|---|---|---|---|---|---|
| OpenAI (GPT-4o) | 1.0 | 0–2 | 1.0 | top-k (legacy), freq/presence −2..2, seed | temp 0; o-series ignore temperature |
| Anthropic (Claude) | 1.0 | 0–1 | 1.0 | top_k (0 = off) | temp 0 for code/structured; no penalties |
| Google (Gemini) | 1.0 | 0–2 | 0.95 | top_k 40, seed (some models) | temp 0 + seed where supported |
| Cohere (Command R+) | 0.3 | 0–5 | 0.75 | top_k, freq/presence (k/m/p) | lower temp + top_k pruning |
| DeepSeek | 1.0 | 0–2 | 1.0 | freq/presence penalties | reasoner ignores temperature |
| Mistral | 0.7 | 0–1 | 1.0 | random_seed | temp 0 + seed |
| Ollama | 0.8 | 0–2 | 0.9 | top_k 40, freq/presence, mirostat, seed | temp 0 + seed; mirostat off |
| Llama 3.x (Meta) | 0.6 | 0–2 | 0.9 | freq/presence (varies) | 0.6 creative, 0 deterministic |
Model notes
Exact defaults vary by provider and model; treat these as common behavioral tendencies, not absolute contracts.
OpenAI
Many chat models cap temperature at 2.0.
Anthropic
Temperature range is 0–1; no frequency/presence penalties.
Google
Gemini usually performs best under 1.0.
Ollama / Local
Common defaults are 0.7–0.9; higher often repeats.
Task playbook
Open preset recommendations for common LLM tasks.