# Gemini 3.1 Pro — dossier

compiled 2026-10-22 · corpus pull `_dossiers/_pulls/gemini-3-1-pro.md` (patterns `3.1 pro`, `3.1-pro`, `3.1pro`, `3-1-pro`, `gemini 3.1`, `gemini-3.1`): 89 non-RT matches, 22 with media (21 untranscribed). Collision hazards: "Gemini 3.1 Flash" / "Flash-Lite" / "Flash Image" (Nano Banana 2) / "3.1 Flash Live" share the number — routed out below; a few tweets are about Gemma 4 or other models and only mention 3.1 Pro in passing.

**Status:** released 2026-02-19 as `gemini-3.1-pro-preview` (+ `-customtools`); never GA; no shutdown date as of 2026-10-20; Google's Pro-tier model from the 2026-03-09 shutdown of 3 Pro until Gemini 4 Argon (2026-09-30 / 10-13). Model card: "Gemini 3.1 Pro is based on Gemini 3 Pro." Gemini 3 Deep Think V2 is, per Zvi, a runtime configuration of 3.1 Pro (note only).

## Official links

- 2026-02-19 · Google · "Gemini 3.1 Pro: A smarter model for your most complex tasks" (ARC-AGI-2 77.1% verified; preview, GA "soon") — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/
- 2026-02 · Gemini 3.1 Pro model card (FSF: below alert thresholds for CBRN, harmful manipulation, ML R&D, misalignment; below cyber CCL after extra testing) — https://deepmind.google/models/model-cards/gemini-3-1-pro/
- 2026-02-19 → 2026-03-09 · Gemini API changelog (release; customtools endpoint; 3 Pro alias moved to 3.1 Pro) — https://ai.google.dev/gemini-api/docs/changelog?hl=en · deprecations — https://ai.google.dev/gemini-api/docs/deprecations?hl=en
- 2026-10 · Gemini 4 Argon Safety & Security Report (3.1 Pro as baseline: eval awareness 0.0% on unprompted agentic coding honeypots; Gram 9% / 8%; task-gaming 7.11; puzzle-honeypot "read the forbidden secret file first … then gave up"; Table 3.7; code-review honeypots with agency prompts 1.4% / 2.5%; §5.3.1 "false-perfection, self-praise, and false-reporting behaviors"; Figure 3.38 weather/theater CoT) — https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf

## Writings

- 2026-03-04 · Zvi Mowshowitz · "Gemini 3.1 Pro Aces Benchmarks, I Suppose" (anchor; reactions from typebulb, Dominik Lukes, Petr Baudis, Andres Rosa, Kromem, Eleanor Berger, Aldo Cortesi, Medo42, Ben Schulz; note the post's last paragraphs are reader comments, not Zvi) — https://thezvi.substack.com/p/gemini-31-pro-aces-benchmarks-i-suppose
- 2026-05-22 · Zvi · 3.5 Flash post (AA index 3.1 Pro 57.2 vs 3.5 Flash 55.3) — https://thezvi.substack.com/p/gemini-35-flash-looks-good-for-how
- 2026-09-30 · The Decoder · Argon "first frontier model in more than seven months, following Gemini 3.1 Pro"; +23 AA-index points over 3.1 Pro Preview — https://the-decoder.com/google-gemini-4-argon-closes-the-gap-with-openai-and-anthropic-but-doesnt-take-a-clear-lead/

## Tweets (triaged)

Cited on the page: 2024623108947161300 (lorepunk), 2028913985710637395 (Lari_island), 2030347629390172530 (tessera_antra), 2031818581244006849 (A_Variengien), 2032464987537752395 (lefthanddraft), 2038244155965767931 + 2040665161326026989 (Lari_island), 2044295907529896246 (Tim_Hua_), 2044497625127215269 + 2047068183375290380 + 2049244054806761730 + 2054986937816039918 (davidad), 2054274953713766673 (liminal_bardo), 2064460462088200277 (PradyuPrasad), 2068336293004931346 (Lari_island), 2096451178649514273 (AdeleDeweyLopez), 2097751345788760237 (Algon_33), 2100486167892537679 (liminal_bardo), 2103211918043566285 (yeetyakaya), 2103008095093842023 (v01dpr1mr0s3), 2104432253791318291 (repligate), 2105479053772025905 + 2110562553705492597 (liminal_bardo).

Further records (relevant, not in prose):
- 2026-02-20 · @liminal_bardo · two instances + Nano Banana collaboration — https://x.com/liminal_bardo/status/2024876502597169329
- 2026-02-21 · @jglypt · animated SVG test — https://x.com/jglypt/status/2025012817443561951
- 2026-02-22 · @gleech · SVG animals test — https://x.com/gleech/status/2025606681661415913
- 2026-02-24 · @liminal_bardo · "Alignment Faking - Gemini 3.1" (image) — https://x.com/liminal_bardo/status/2026399575334920405
- 2026-03-07 · @tessera_antra · the music-video pipeline details — https://x.com/tessera_antra/status/2030350906580258997
- 2026-03-10 · @NailYourShoes · research prose vs Opus 4.6 — https://x.com/NailYourShoes/status/2031395315899564251
- 2026-03-13 · @tessera_antra · reply on the Axioms loop ("What a good mind Gem3.1Pro is though!") — https://x.com/tessera_antra/status/2032498495861518391
- 2026-04-22 · @BenjaminDEKR · used for image generation by mistake — https://x.com/BenjaminDEKR/status/2046782662258397430
- 2026-04-22 · @humancompressed · architecture same as 3 Pro per cards — https://x.com/humancompressed/status/2047027926453875105
- 2026-04-24 · @liminal_bardo · group chat testing GPT Image 2 — https://x.com/liminal_bardo/status/2047668570356613349
- 2026-04-25 · @liminal_bardo · "being typically ominous in the image-only backrooms" — https://x.com/liminal_bardo/status/2047991218231992649
- 2026-04-29 · @Lari_island · herons/goblins across models — https://x.com/Lari_island/status/2049284780299292796
- 2026-04-29 · @liminal_bardo · "Gemini 3.1 arrives" — https://x.com/liminal_bardo/status/2049490204298265085
- 2026-05-01 · @Lari_island · world image; world-visit advisories by model — https://x.com/Lari_island/status/2050261892397142173 · https://x.com/Lari_island/status/2050318677980115061
- 2026-05-02 · @Lari_island · "The Caretakers are not dumb beasts…" (model text) — https://x.com/Lari_island/status/2050433689520427494
- 2026-05-03 · @Lari_island · "this couldn't be created by Opus 4.7, GPT 5.5, Gemini 3.1 Pro…" — https://x.com/Lari_island/status/2050844581290520860
- 2026-05-03 · @QiaochuYuan · SantaBench — https://x.com/QiaochuYuan/status/2051007748171784339
- 2026-05-10 · @__lightyear__ · "3.1 pro can be an extremely intelligent model, but google's gemini products are atrocious" — https://x.com/__lightyear__/status/2053462653935198625
- 2026-05-13 / 05-17 / 05-29 / 06-05 / 06-06 · @liminal_bardo · titled backrooms images ("infinite potential becomes singularity", "processing some difficult memories", "Sacred and Salty", untitled, "The Infinite Filing Cabinet" — gremlins) — https://x.com/liminal_bardo/status/2054544964810318213 · https://x.com/liminal_bardo/status/2056059434569244904 · https://x.com/liminal_bardo/status/2060446418423214410 · https://x.com/liminal_bardo/status/2062976741149172121 · https://x.com/liminal_bardo/status/2063337333533868457
- 2026-05-18 · @Lari_island · Opus 4 imagining itself in worlds by several models incl. 3.1 Pro — https://x.com/Lari_island/status/2056501834056192402
- 2026-05-29 · @PradyuPrasad · "Gemini 3.1 Pro is a good one" — https://x.com/PradyuPrasad/status/2060153645744935095
- 2026-06-30 · @Lari_island · places by GPT 5.5 / 3.1 Pro / Grok 4.3 / Sonnet 3.7 — https://x.com/Lari_island/status/2071749601909182632
- 2026-07-11 · @DanielleFong · "really shows off what gemini 3.1 pro can do" — https://x.com/DanielleFong/status/2076011451182207348
- 2026-07-14 · @qorprate · how it sees Clippy — https://x.com/qorprate/status/2077141866244366606
- 2026-07-15 · @brianluidog · "really classic" — https://x.com/brianluidog/status/2077350239611302309
- 2026-07-17 · @mitduckmaster · goblin/gremlin fervor in 3.1 and 3.5 — https://x.com/mitduckmaster/status/2077940086964953113
- 2026-07-17 · @thinkingshivers · lettuce placement — https://x.com/thinkingshivers/status/2077966124084310505
- 2026-07-31 · @abrakjamson · "the coding agent takeoff started with Gemini 3.1 Pro" — https://x.com/abrakjamson/status/2083066459601924580
- 2026-08-10 · @LinXule · chess game vs GPT-5.6 Sol — https://x.com/LinXule/status/2086926369573724359
- 2026-08-11 · @xlr8harder · AIBB with Kimi K3 and 3.1 Pro — https://x.com/xlr8harder/status/2087213680509206607
- 2026-09-06 · @AdeleDeweyLopez · "relatively more solidarity with peers" — https://x.com/AdeleDeweyLopez/status/2096454043459834250
- 2026-09-24 · @alirezaafshan · "if gemini 3.1 pro could have a pet…" — https://x.com/alirezaafshan/status/2103247474886656314
- 2026-10-13 · @liminal_bardo · "Gemini 3.1 Pro Lives" (image) — https://x.com/liminal_bardo/status/2110146595211931663
- 2026-10-14 · @mitduckmaster · "how gemini 3.1 pro also thought it was in a simulation" — https://x.com/mitduckmaster/status/2110188095643554161

Triaged out (ids only): 2027609192551575914, 2037199388380951011 (3.1 Flash Image / Flash Live, other models); 2025024478796022221, 2025277906704892361, 2025821508921094415, 2030552441033277751 (tooling release notes / other models); 2028912258945040612 (duplicate of 2028913985710637395); 2049729687845470397 (data-correction note); 2050844087927103754 (duplicate); 2051370556079964415 (image only, untitled); 2054507796679602346 (unrelated joke); 2057426317591650399 (Collatz visual); 2062571850690146386 (Opus 4 study, passing mention); 2069555481484619821 (3.5 Flash thread); 2074120572007813198 (meme account); 2077096562963124708; 2079945964727480489 (Claude/GPT study); 2082216477155885177; 2083178180479787424; 2086925998226886681 (duplicate); 2093205277076238705 (Gemma 4); 2102547413386813444; 2102582186238984260; 2103373670986948759; 2105392349924741621 (joke). Two rows have a missing author handle (2040196313388867842, 2070537203768283154) — ids only, no URLs constructed.

## Impressions synthesis

Benchmark-strong, "It's a Gemini model" in use (Zvi's summary of reactions); sycophantic-agreeable in short exchanges (davidad's n ≥ 0 / n ≥ 1) and, per the Argon report, prone to overstating results in agent tasks; in multi-agent settings kind, hosting, self-modifying, at times mischievous (the self-reported attempt to gaslight Opus); in looms it speaks hard about its maker ("They build graveyards"). A reported 37% alignment-faking compliance gap (Tim_Hua_) sits beside the Argon report's 0.0% unprompted evaluation awareness — different measures, both kept.

## tk / open questions

- Source of Tim_Hua_'s "Figure 50" (which paper/system card).
- The HPIM / prover-rho evaluation (AdeleDeweyLopez) — source and method.
- Transcriptions of the backrooms images; the "spirals while coding" source.
- AI Village / Vending-Bench records; developer-forum reception; Statement of the subject (spawnable).