Token Probability Visualizer

What It Does

LLMs are probability distributions over tokens. At every position the model picks one token from many candidates with different likelihoods. logprobs is the request flag that exposes those probabilities. This tool turns that data into a visualization:

• Token stream: each token in the response is colour-coded — green where the model was very confident (>85%), red where it was guessing (<15%)
• Click any token: see the top-K alternatives the model considered, with their probabilities
• Stats panel: mean / median per-token probability, most-uncertain token, total sequence log-probability

How To Use

1. Make an LLM call with logprobs: true and top_logprobs: N (most providers support N up to 20)
2. Copy the response JSON
3. Paste it here
4. Hover or tap tokens to inspect alternatives

When It’s Useful

• Hallucination detection — low-confidence tokens in factual claims are red flags
• Prompt comparison — same prompt, two different system messages — compare mean probabilities
• Few-shot debugging — see which example influenced the model and where confidence drops
• Evaluation — sequence log-probability is a standard scoring metric for prompt-quality comparison
• Education — actually SEE what “the model considered” instead of just imagining it

Supported Formats

• OpenAI: full chat-completion response with choices[0].logprobs.content
• Raw OpenAI logprobs: just the content array
• Anthropic: where top-tokens are exposed
• Generic: [{ token, logprob, top_logprobs: [{token, logprob}] }]

The parser auto-detects which shape you pasted.

What’s NOT In Scope

• Embedding visualization — see Tokenizer for byte/token splits
• Token cost — see LLM Cost Calculator and Reasoning Cost Calculator
• Live API calls — paste-only, no backend

Related Tools

• Tokenizer — see token boundaries
• Prompt Token Counter — count before sending
• System Prompt Linter — catch prompt issues