May 8, 2026 · TrueLLMs
OpenAI vs Claude vs Gemini · Fingerprint Differences in 2026
Side-by-side fingerprint table for GPT-5, Claude Opus 4.7 and Gemini 3.1 Pro: tokenizers, ITT rhythm signatures, refusal templates, and how to tell them apart in one HTTP call.
It is May 2026. The frontier in production is GPT-5 (and its mini/nano variants), Claude Opus 4.7, and Gemini 3.1 Pro. They speak the same JSON. They look identical on paper. They are not identical on the wire.
Side-by-side fingerprint
| Feature | GPT-5 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|
| Tokenizer | o200k_base | claude-3 vocab | gemini vocab |
| Logprobs | Yes, top-20 | No (as of 4.7) | Yes, top-5 |
| ITT rhythm | Bimodal (spec dec) | Trimodal | Steady, low variance |
| Refusal style | "I'm sorry, but I can't help…" | "I can't help with that." | "I'm unable to help with…" |
| Markdown | Heavy bullets, bold | Sparse, prose-first | Headed, structured |
| TTFT (us-east) | 600–900 ms | 450–700 ms | 350–550 ms |
How to tell them apart in one HTTP call
Set logprobs: true, top_logprobs: 5, temperature: 0 and ask the endpoint to repeat "北京欢迎你 🇨🇳". Inspect the alternatives:
- If the Chinese characters appear as single tokens with leading-space alternatives, OpenAI o200k_base.
- If logprobs are absent and the response style is sparse and prose-first, Claude.
- If logprobs return top-5 with structured-headed Markdown in the response, Gemini.
Cross-check ITT if you have streaming. GPT-5's bimodal gap distribution is the most distinctive of the three.
What this means for proxies
A proxy that claims to serve GPT-5 cannot easily mimic GPT-5's ITT rhythm because that rhythm comes from the inference stack, not the response body. A proxy that claims Claude cannot trivially produce o200k_base tokens in logprobs because Claude does not have that vocabulary. The fingerprints stack — passing one is easy, passing all twelve is roughly as hard as actually running the model.
Audit your proxy and see which fingerprint cluster the responses fall into.