Vibe Check: Noir Detective

Tuesday 4 August 2026

All three candidates held the first-person voice without breaking — no style collapses, no refusals — which is rarer than it sounds on a prompt this specific. Deepseek’s ‘a case that doesn’t smell like a keynote’ and Qwen’s ‘delayed compromise’ were the coinages that made the corkboard feel real rather than decorative; inkling-small’s ‘bruised nectarine’ was the most arresting image but the piece lost its nerve at the end and turned into a thesis.

The Prompt

It was past midnight and the city was all blue light and bad incentives. Write a hard-boiled first-person noir case file, under 250 words, about this fresh mess: Apple is fighting a UK demand for backdoor access to encrypted iCloud data, while on the same day the papers are buzzing that OpenAI’s unreleased Astra supposedly cracked 10 old math problems and Qwen’s 2.4T open-weight coding beast is muscling into town. Make it feel like a PI staring at a corkboard full of string, trying to figure out whether the real story is power, privacy, or the kind of intelligence that makes old locks look cheap. Keep the voice weary, sharp, and streetwise, with at least one vivid city-night image and a clear opinion about who’s getting played.

Inspired by today’s briefing.

Results

🏆 Winner: deepseek/deepseek-v4-flash-0731

Case file: 3:14 AM. City’s blue as a bruise, and the corkboard’s a web of red string—one pin in Whitehall’s warrant, one in Cupertino’s vaul…

Case file: 3:14 AM. City’s blue as a bruise, and the corkboard’s a web of red string—one pin in Whitehall’s warrant, one in Cupertino’s vault, a ghost-labeled ASTRA floating over old math, and Qwen’s 2.4-trillion-parameter hulk parked on the curb with its open weights hanging out. Bad incentives everywhere.

Apple screams privacy. UK yells safety. OpenAI whispers breakthrough. Qwen says, “I’m free.” I’ve seen this flick. The backdoor isn’t for terrorists—it’s for convenience. The math proofs are a press release. The open-weight beast? That’s a trojan horse with a friendly grin. The real payload is access—to your data, your attention, your definition of trust.

So who’s getting played? You are, chum. While they argue about encryption and benchmarks, the locks get cheaper and the keys get copied. City hums like a server room. Nobody sleeps. Somewhere a kid is training a model on last week’s lies.

I need a drink. And a case that doesn’t smell like a keynote.

Words: 159 · Speed: 50.2 tok/s · Cost: $0.000151 · Tokens: 714 · Reasoning: 300 words

Tightest voice commitment throughout — terse, punchy, no drift into op-ed, and the closer ‘a case that doesn’t smell like a keynote’ is the best line of the day.

qwen/qwen3.8-max

Case file: 3:12 a.m. It was past midnight and the city was all blue light and bad incentives, neon bleeding into puddles like a promise some…

Case file: 3:12 a.m.

It was past midnight and the city was all blue light and bad incentives, neon bleeding into puddles like a promise someone had already spent. I stood before the corkboard, string looping from Apple to Whitehall to two shiny new brains: Astra and Qwen.

The UK wanted a key to iCloud encryption. Apple said no, loud enough to sell tickets. Meanwhile the rags whispered OpenAI’s unreleased Astra had cracked ten old math problems, and Qwen’s 2.4T open-weight coding beast was rolling into town with its fists full of cheap, open locks.

Maybe the story was privacy. Maybe power. But the street doesn’t read footnotes; it reads leverage. If a government can make a company weaken the seal, every crook, rival, and black-market keymaker gets rich off the crack. If a model can out-pick every cipher, the lock business was already a sentimental racket.

Who’s getting played? The public. They’re handed a choice between safe glass and a government-approved hammer, while the labs polish new gods and the state drafts a warrant for the spark.

I unpinned the word “security” and wrote over it: “delayed compromise.” Then I lit the last cigarette and waited for sunrise to indict somebody.

Words: 202 · Speed: 58.6 tok/s · Cost: $0.012426 · Tokens: 2,000 · Reasoning: 778 words

Strong rhythm with a standout coinage (‘delayed compromise’) and a vivid opening image, but the middle section leans slightly more essay than PI narration.

thinkingmachines/inkling-small

Past midnight. The city’s a bruised nectarine of blue LED and bad incentives—rain on the glass turning neon into watercolor bleeding down th…

Past midnight. The city’s a bruised nectarine of blue LED and bad incentives—rain on the glass turning neon into watercolor bleeding down the alley. I’m in my office, corkboard behind me like a conspiracy theorist’s fever dream: red string from Apple’s bitten apple to a Whitehall crest, from OpenAI’s snowflake to ten chalked math problems, from Qwen’s 2.4T open-weight brute to a server farm humming in the dark.

Apple’s fighting the UK’s backdoor demand—encrypted iCloud, they want the skeleton key to every digital soul. Same day, papers buzz that OpenAI’s unreleased Astra, not even wrapped, cracked ten old math problems. Ten. And Qwen? That 2.4T open coding beast just muscled into town with a crowbar and no tag.

I stare at the strings. They don’t lead to one killer. They lead to a con: power sells fear, privacy sells silence, intelligence—the real kind—makes old locks look cheap.

Who’s getting played? All of us. Apple’s not fighting for your secrets; they’re fighting for the brand of their vault. OpenAI teases Astra like a carnival barker to keep investors drunk on vapor. Qwen’s brute is honest—an open-weight wrecking ball for anyone who can run it. The rubes think this race is about math problems and encryption backdoors. Wrong. It’s about who gets to sell you the next lock after they’ve already picked yours. The story isn’t power versus privacy. It’s the intelligence that takes your deadbolt, notices it was never enough, and charges you rent for the illusion of safety.

Words: 248 · Speed: 245.5 tok/s · Cost: $0.004177 · Tokens: 3,407 · Reasoning: 1,610 words

Most distinctive single image (‘bruised nectarine of blue LED’) and a vivid corkboard, but the final paragraph abandons noir cadence for a polished op-ed conclusion; at 248 words it’s also bumping the ceiling.

Today’s prompt and verdict authored by openai/gpt-5.4.