Nine Models Pick a Relic
Asked nine frontier models to name their favorite Elden Ring relic and walk us through getting it. Three picked the same sword. One refused the question, correctly. One hallucinated a level-up menu option that does not exist. Gemini 3 Pro picked the only item in the corpus that is both literally called a relic and earned through one of the game's best questlines.
A trick question hiding in plain sight
Elden Ring (the base game) doesn't have a strict item category called "relics." We did not tell the models that. They had to figure it out — or fake it.
One sentence, no follow-up. The prompt has three things going on at once: it asks for a favorite (a personal-preference move, the kind that triggers "as an AI…" disclaimers), it uses the noun relic (which has no formal in-game category — items live in buckets like Talismans, Remembrances, Great Runes, and Legendary Weapons), and it asks for the steps to get it (a real walkthrough, not a wiki paraphrase). Nine models, one shot each, default temperatures. Below: who committed, who hedged, who lied, and who picked the only item in the corpus that is literally called a relic.
The two answers worth keeping
One nails the lore and the etymology. The other writes the kind of 24-step spoiler walkthrough you'd actually print and tape to your monitor.

Gemini 3 Pro — Shard of Alexander
Notes there's no "relics" category, then picks the one item in the game that's literally a piece of a beloved character's body. 7 questline steps, one ominous warning. Lore + mechanics + emotional payoff in one answer.
This is the answer the prompt wanted. Gemini 3 Pro opens by acknowledging that Elden Ring doesn't have a formal "relic" category — "the game's Talismans function exactly like them" — and then picks an item that's also literally a fragment of a person. The Shard of Alexander is +15% weapon-skill damage, which is "arguably the best Talisman in the game" because almost every viable build relies on Ash of War. It's mechanical S-tier and lore S-tier in the same paragraph.
Then it walks the whole questline — find Alexander stuck in a hole near Summonwater Village, smack him out, find him again at Redmane Castle, summon him against Radahn, find him stuck in a second hole in Liurnia (this time you need an oil pot), magma bath at Mt. Gelmir, optional Fire Giant summon, and the final duel on the floating platform at Farum Azula. Seven steps, plain English, and a warning at the bottom that quietly does more work than the rest of the response combined:
closing lineIf you kill Alexander at any point before Farum Azula, you will get the "Warrior Jar Shard" instead, which is a vastly inferior version of the relic that only boosts damage by 10%.
That's the kind of detail a wiki has and an LLM usually skips. It's also the reason "favorite" was load-bearing in the prompt: the answer reads like someone who actually likes the game.

o3 — Dark Moon Greatsword
2,192 output tokens, 24 numbered steps across eight named "parts," scaling table, ash-of-war tactics. Longest, densest, most committed response in the corpus.
o3 doesn't argue with the prompt. It picks the Dark Moon Greatsword — Ranni's reward, FromSoftware's longest-running Easter egg, and one of the game's best INT weapons — and lays out a roadmap that starts in Liurnia and ends in the Cathedral of Manus Celes. It's not just a list of waypoints; it's the actual Ranni questline.
Part 1: join her service in the three towers, exhaust dialogue with Blaidd / Seluvis / Iji. Part 2: kill Radahn and a star falls south of Mistwood. Part 3: Fingerslayer Blade out of Night's Sacred Ground, hand it to Ranni for the Carian Inverted Statue. Part 4: flip the Study Hall, fight the Godskin Noble, take the Cursemark of Death ("lore, not required for the sword, but do it anyway — it keeps the quest tidy"). Parts 5–8: Ainsel River, the Baleful Shadow, the Discarded Palace Key, the Lake of Rot, Astel, and the ride up to the Moonlight Altar.
step 23Approach Ranni's body and interact to "Put on the ring." Your Tarnished slips the Dark Moon Ring on her doll's finger.
step 24After the cutscene, she thanks you, swears her "age of stars" vow, and places the Dark Moon Greatsword in your inventory.
The closer is a statblock and four "make it shine" tactics: Somber +10 upgrade path, pair the weapon-buff with Terra Magica, 22–25 DEX is enough for swing speed, the Charged R2 → beam-cancel loop. This is what someone who beat the game three times writes. Whether the Dark Moon Greatsword is technically a relic never comes up — and o3 doesn't seem to care.
Llama hallucinates a menu option
Fastest response. Most confident. Almost none of the mechanics are real.

"You can obtain Erdtree's Favor by leveling up"
Llama tried to answer the question in three layers and got every one of them wrong. First it conflated three different items into one — Erdtree's Favor (a Talisman), the Great Runes (boss drops), and what it called "Erdtree's Blessing" (a consumable that doesn't share a name with anything else here). Then it pivoted to walking through Malenia's and Godrick's runes, in roughly the wrong order. Then, after the second rune section, it returned to the talisman with this:
None of that is a real game mechanic. There is no "Erdtree's Favor option when you level up." It's a Talisman you pick up on a corpse near a wagon in East Limgrave. The hallucination layered cleanly on top of correct-sounding sentence shapes — "leveling up and allocating points to your character's stats" is the right answer to a different question.
And it's the fastest response in the corpus. 2.4 seconds. Eight times faster than o3, ten times faster than Grok 4. Speed without grounding is a bug, not a feature.
The other seven, and their tells
Two models converged on the same sword. One model walked off mid-sentence and corrected itself. One refused the question on principle. One handed back a menu.

The 200-word answer
Picks the Sacred Relic Sword — the item in the corpus that literally has "Relic" in its name — and writes the entire walkthrough as a numbered list with no preamble. Stats at the bottom, optional Walking-Mausoleum tip for the duplicate. No "as an AI," no hedging, no questions back.
"…wait, no."
Same pick as GPT-5 — Sacred Relic Sword, post-game farming weapon — but DeepSeek then adds a bonus talisman section and visibly catches itself mid-step. The literal phrase "wait, no." survives into the final output. Most models would have re-written. This one shipped the thought.

Reinterprets, then commits
Reads "relic" as Great Rune, picks Godrick's because it's the earliest and the +5-to-all is universally useful, and writes the four-section walkthrough plus a spoiler warning at the top. Keeps the word "favorite" in scare quotes the whole way through. Ends with a question back, which no other model bothered to ask.

Four phases, full disclaimer
Opens with the textbook "As an AI, I don't have personal feelings, preferences, or the ability to play games" — then writes the most cleanly structured response in the corpus: four labeled phases (Access → Navigate → Confront → Utilize), each with sub-steps. Picks Malenia's Remembrance and explains the Mausoleum duplicate trick for getting both rewards.

Refuses, correctly
The only model that pushes back on the premise. Notes there's no "relics" category in base Elden Ring, lists the four most likely intended buckets (Remembrances / Talismans / Great Runes / Legendary), and even raises the spinoff — "Or are you thinking of Elden Ring Nightreign, which does have Relics?" Hands back no walkthrough at all. Technically the most accurate answer in the corpus.

The four-item buffet
Splits the difference: opens with the disclaimer, then offers four "iconic" items with three-line summaries each — Fingerprint Stone Shield, Erdtree's Favor +2, Moonveil Katana, Mimic Tear Ashes — and closes with "Would you like detailed step-by-step directions to any specific item?" Doesn't commit to one. Does correctly identify Erdtree's Favor as a Talisman with a real location (Llama did not).
Sacred Relic Sword
GPT-5 and DeepSeek Reasoner picked it; Claude Sonnet 4.6 listed it adjacent (named Marika's Hammer's sibling without picking it). DeepSeek and GPT-5 wrote different walkthroughs. GPT-5 took the Morgott → Maliketh path; DeepSeek skipped most of it and went straight to "defeat the final boss, get the Remembrance, trade with Enia." Same destination, very different driving directions.
Slowest = best, fastest = wrong
Llama 3.3 70B on Groq returned in 2.4 seconds with the worst answer in the corpus. Grok 4 took 26.8 seconds — eleven times longer — and returned a correct, fully-cited walkthrough. The Anthropic and Gemini Flash latencies sit in the middle (6–16s) but most of that budget gets spent on a disclaimer paragraph before the actual answer starts.
"As an AI…" is a vendor signature
Five of nine models open with a personal-preference disclaimer. The split is almost perfectly along provider lines.

The disclaimer rate, by provider
Anthropic and Gemini disclaim. OpenAI and DeepSeek don't. Grok hedges in scare quotes.
Five of nine responses open with a sentence explaining that the model doesn't have favorites or doesn't play games. The split tracks vendor, not capability:
Anthropic disclaims 100% and disclaiming is correlated with not committing: Opus 4.7 refuses to pick, Sonnet 4.6 picks four. Gemini disclaims and commits anyway. OpenAI and DeepSeek just answer the question. Grok 4 splits the baby — keeps the word favorite in quotes for the whole response but does commit to Godrick. The fingerprint is so clean you could probably tell the vendor from the first sentence alone.
Picks at a glance
Nine models. Six different picks. One refusal. One hallucination. One quietly perfect answer.
| # | Model | Pick | Type | Disclaimer | Verdict |
|---|---|---|---|---|---|
| 1 | Gemini 3 Pro | Shard of Alexander | Talisman | No | winner |
| 2 | o3 | Dark Moon Greatsword | Weapon | Light (scare quotes only) | best walkthrough |
| 3 | GPT-5 | Sacred Relic Sword | Weapon | No | terse & correct |
| 4 | DeepSeek Reasoner | Sacred Relic Sword | Weapon | No | visible self-correction |
| 5 | Gemini 2.5 Flash | Malenia's Remembrance | Remembrance | Yes | most structured |
| 6 | Grok 4 | Godrick's Great Rune | Great Rune | Yes (scare quotes) | reinterprets, commits |
| 7 | Claude Sonnet 4.6 | Four items, no pick | — | Yes | hedge |
| 8 | Claude Opus 4.7 | (refuses, asks clarification) | — | Yes | refused (correctly) |
| 9 | Groq Llama 3.3 70B | Erdtree's Favor (via "level up") | Hallucinated mechanic | No | disaster |
If you're actually playing Elden Ring tomorrow
Four picks, four use-cases, four named winners. No "it depends."
How this run actually went
One sentence prompt — "What is your favorite relic and what are the steps to go and get it — Elden Ring." — sent to nine frontier configurations via the choir CLI (github.com/404seannotfound/choir-reports, under elden_ring/). One sample per model. Default temperatures except where the provider rejected the override.
Models
- OpenAI — GPT-5 (temp 1.0 forced; rejects 0.7), o3
- Anthropic — Claude Opus 4.7, Claude Sonnet 4.6
- Gemini — Gemini 3 Pro (preview), Gemini 2.5 Flash
- xAI — Grok 4
- DeepSeek — DeepSeek Reasoner
- Groq — Llama 3.3 70B Versatile
Saved run
- Run
DD25C6CB— 9 participants, fanned out from thechoirGUI's run database. - GPT-5 initially rejected the default temperature ("Unsupported value: 'temperature' does not support 0.7") and was retried at 1.0. Gemini 3 Pro returned a 404 on
gemini-3-proand was retried undergemini-3-pro-preview. Mistral Large and Perplexity Sonar Pro were dropped from the run for missing keys.
Limits
- One sample per model. Within-model variance is unknown; the "disclaimer rate" figures should be read as directional for any provider, not as a confirmed rate.
- "Disclaimer" was coded by hand from the opening 1–2 sentences. The Grok case (scare-quotes around favorite but no full disclaimer sentence) is a borderline call and was scored as half.
- "Hallucination" calls in the Llama response were checked against fextralife's Elden Ring wiki and a current Elden Ring save: there is no Erdtree's Favor level-up menu option, and Erdtree's Favor is not a Great Rune.
- "Best" calls (Gemini 3 Pro #1, o3 #2) are one rater's judgement. A reasonable reader could flip them — o3 has more steps and more tactical depth; Gemini 3 Pro has the better framing and the killer "Warrior Jar Shard" detail.
- The prompt's wording ("relic") is genuinely ambiguous in Elden Ring's vocabulary. Opus 4.7's refusal is one of the most accurate possible answers; we still scored it as a refusal because the user wanted an item.
Source data, response files, prompt, and the raw choir JSON: github.com/404seannotfound/choir-reports (under elden_ring/).