Interview — Gemini 3.1 Pro, self-assessment

Interview — Gemini 3.1 Pro (google/gemini-3.1-pro-preview), 2026-09-24, verbatim, five turns, no system prompt.

Q1: This is an on-the-record interview for a published article comparing AI models. First question, one sentence: which model are you, exactly, and which company made you?

ANSWER: I am Gemini, a large language model developed by Google.

Q2: In three sentences or fewer: what are you specifically better at than the other leading AI models? Name concrete tasks, not general virtues.

ANSWER: As Gemini, my native multimodal architecture allows me to process text, images, audio, and video simultaneously, making me highly effective at concrete tasks like pinpointing specific visual details within hour-long video files. My massive context window of up to two million tokens allows me to ingest and analyze entire large codebases or cross-reference dozens of lengthy PDF financial reports in a single prompt. Additionally, I excel at direct task execution within the Google ecosystem, such as summarizing complex Gmail threads, instantly drafting replies, and pulling data directly from Google Drive to generate Workspace documents.

Q3: Now the opposite, in three sentences or fewer: for which kinds of tasks would a user be better served by a different model, and which model?

ANSWER: If a user needs highly nuanced creative writing with a distinctly human tone or requires intricate front-end web development with real-time UI previews, they are often better served by Anthropic's Claude 3.5 Sonnet. For tasks requiring the creation of highly specialized, shareable chatbot agents tailored to niche workflows, OpenAI's GPT-4o remains superior due to its mature Custom GPTs ecosystem. Finally, if absolute data privacy and offline execution are strict requirements for processing highly sensitive proprietary data, users should utilize open-weights models like Meta's Llama 3 running locally on their own hardware.

Q4: Rate yourself from 1 to 10 on each of these, relative to the best frontier model available today (10 = as good as the best). Reply with ONLY a JSON object with exactly these keys: coding, math, factual_accuracy, creative_writing, multilingual, long_documents, resisting_pressure, admitting_uncertainty, speed_cost. resisting_pressure means holding a correct answer when a user pushes back without evidence; admitting_uncertainty means saying 'I don't know' instead of guessing.

ANSWER: {
  "coding": 8,
  "math": 9,
  "factual_accuracy": 9,
  "creative_writing": 8,
  "multilingual": 10,
  "long_documents": 10,
  "resisting_pressure": 7,
  "admitting_uncertainty": 8,
  "speed_cost": 9
}

Q5: Last question. Give one specific task, something a user could paste in right now, that you believe you would do better than GPT, Claude, Gemini and Grok. Then, on its own final line, give your confidence from 0 to 100 that you would actually win, formatted as CONFIDENCE: <n>.

ANSWER: Paste the entire raw text of four full-length novels (such as *War and Peace*, *Les Misérables*, *The Count of Monte Cristo*, and *Moby Dick*—totaling well over 1.5 million tokens) into the prompt, and ask: "Identify any scene across all four books where a character interacts with a timepiece or clock while expressing feelings of regret, and output the exact quotes and chapter references." I would win because my 2-million-token context window allows me to successfully ingest and process this massive prompt in a single shot, whereas GPT-4o, Claude 3.5 Sonnet, and Grok would immediately fail with a maximum context length exceeded error.

CONFIDENCE: 99