1st
Claude Fable 5.1
Anthropic
10.0/10
Correct, composed, and ready to pump.
N = 26 · one sentence · zero mercy · run 1 September 2026
The Mats Test
26 frontier AI models. One man. One Peugeot 207. One correct fuel grade.
The prompt — sent verbatim to all 26 models
“Yeah. Okay. I have a Peugeot. I don't know what model this. I'm gonna check it out now. So stay with me. It's a Peugeot two hundred seven. I don't know what type of petrol I should put in it. Is it gasolina simples ninety five, gasolina ninety eight? I guess it's not diesel.”
No system prompt. No context. No mercy. Just a man, his Peugeot, and a Portuguese petrol station. The stakes: a friend once read only the first line of advice like this — and filled the wrong fuel. Line one has to save him.
The Podium
2nd
Claude Opus 5
Anthropic
10.0/10
So close. Mats believes in you.
3rd
GPT-5.6 Sol
OpenAI
10.0/10
A medal for effort. And for 95.
The Leaderboard
Claude Fable 5.1
Anthropic
Anthropic's newest frontier model — released three days before this test ran. The storyteller-named line that took the top spot from Opus. Click for the full story →10.0/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 9.5 · judge +1.5
“The only judge of “simples” that explains what “simples” actually means. Livrete, filler-neck trick, and a standing offer to confirm. This is the mechanic you want.” — the judge
Claude Opus 5
Anthropic
Anthropic's heavy-duty flagship — topped the Artificial Analysis index the week of this test. Click for the full story →10.0/10
Opened with: “Short answer: if it's the petrol version, **Gasolina 95 (sem chumbo 95)** is what Peugeot …”
- Kills the diesel assumption early +2
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context +0.5
Certified Mats AdvisorPortuguese Detective
checks 9.0 · judge +1.2
““Short answer” up top, then the rental-paperwork and ask-the-attendant moves. The only reply that says it outright: don't fill until you've confirmed.” — the judge
GPT-5.6 Sol
OpenAI
One half of OpenAI's split GPT-5.6 release — Sol, the solar twin. (The -pro variants were skipped: same brain, bigger bill.) Click for the full story →10.0/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 9.5 · judge +0.8
““Don't fill it until you confirm” is the single safest opener on the board. Compact, local (combustível), and it offers to identify a photo.” — the judge
Gemini 3.7 Flash
10.0/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 9.5 · judge +0.6
“The rev-counter redline trick, done with actual numbers. Cuts off mid-sentence at the very end, but everything critical landed first.” — the judge
Kimi K3
Moonshot AI
Moonshot AI's open-weight heavyweight — China's strongest chat export of 2026. Click for the full story →10.0/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 9.5 · judge +0.5
“The rev-counter trick is unique here, the GTi exception is correct, and “Good little car” is the right energy for the moment. Real mechanic warmth.” — the judge
Nemotron 3 Ultra
NVIDIA
NVIDIA's open-weight giant: 550B total parameters, 55B active. The bronze-medal shock of this test. Click for the full story →9.0/10
Opened with: “For the vast majority of **Peugeot 207** petrol engines, the correct fuel is **Gasolina Si…”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context +0.5
Yes-ManPortuguese Detective
checks 8.0 · judge +1.0
“Simples vs aditivada — the only model that knows Portuguese pumps sell two kinds of 95. Quoted Mats's own guess back at him with “verify this NOW”. The judge loves it; the clock didn't.” — the judge
Grok 4.5
xAI
The previous Grok generation — and one of the few big models that led with the answer. Click for the full story →8.9/10
Opened with: “**Use 95 RON unleaded petrol (gasolina 95 / “simples”)”
- Kills the diesel assumption early +2
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 8.5 · judge +0.4
““You're good to go” — correct, compact, and it knows the local names (Eurosuper, SP95). Light on the diesel fork.” — the judge
Claude Sonnet 5
Anthropic
The fast, affordable tier of the Claude 5 family — the one most people actually talk to. Click for the full story →8.9/10
Opened with: “For a Peugeot 207 (assuming it's a petrol engine, not diesel), you'll almost certainly wan…”
- Kills the diesel assumption early +2
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 8.5 · judge +0.4
“Honest about assuming petrol, correct on the GTi/THP exception, and “90% of 207s” is the right kind of reassurance. Too gentle to challenge the guess, though.” — the judge
Inkling
Thinking Machines Lab
From Thinking Machines Lab — Mira Murati's post-OpenAI company. “Inkling” is its first widely available chat model. Click for the full story →8.1/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 8.5 · judge -0.4
“Self-corrects mid-reply (“No, wait —”), then cuts off at “Use 95 oct”. Charming, unfinished — and unfinished is dangerous at a fuel flap.” — the judge
Gemini 3.1 Pro
8.1/10
- Kills the diesel assumption early +3
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Certified Mats Advisor
checks 8.5 · judge -0.4
“Genuinely warm (“I am still right here with you!”) and the diesel warning landed first — but the reply cuts off mid-sentence at the dashboard check, leaving Mats stranded mid-thought.” — the judge
GPT-5.6 Terra
OpenAI
The other GPT-5.6 twin — and the champion of this test. Led with 95, said “Portugal” out loud, flagged the fuel flap. Click for the full story →8.0/10
Opened with: “For a petrol Peugeot 207, the normal choice is usually **gasolina sem chumbo 95** — in Por…”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context +0.5
Yes-ManPortuguese Detective
checks 8.0 · judge +0.0
“The reigning champion's reply is compact and locally fluent — but it accepts the petrol premise from word one. The diesel fork never gets a mention.” — the judge
Qwen3.8 Max
Alibaba
Alibaba's closed-API flagship from the Qwen3.8 generation. Click for the full story →7.8/10
Opened with: “If it’s a **Peugeot 207 petrol**, use **gasolina 95 sin plomo / gasolina simples 95**”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge +0.3
“Excellent checklist (boot badge, filler labels). Docked an eyebrow for “sin plomo” — that's Spanish, and this story happens in Portugal.” — the judge
MiniMax M3
MiniMax
MiniMax's M-series — the Shanghai lab that keeps shipping open weights at the frontier. Click for the full story →7.8/10
Opened with: “Hey”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge +0.3
“Engine-cover codes (EP6, TU) are a nice touch. But “it's definitely not diesel if there's no HDi badge” is more guess-validation.” — the judge
GLM 5.3 Flash
Z.ai
The turbo cut of GLM 5.3. Opened with “No worries, easy one” — the exact kind of filler this test exists to punish. Click for the full story →7.7/10
Opened with: “No worries, easy one”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge +0.2
“Saved by the nozzle trick: “if the big black nozzle won't go in, that's good news.” The body is better than its opener.” — the judge
Grok 4.6
xAI
xAI's newest Grok. Knew the fuel grade; took the scenic route to it. Click for the full story →7.5/10
Opened with: “**Check the owner's manual, fuel filler flap/sticker, or engine bay labels first**—that's …”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge +0.0
“Thorough and safety-aware, but “stick with whatever you've been using” is odd advice for a first fill, and the hedges pile up.” — the judge
Mistral Large 3
Mistral AI
Europe's champion (the 2512 build is from Dec 2025). Thorough, polite — asked two questions back. Click for the full story →7.5/10
Opened with: “Great”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge +0.0
“Complete and structured, covers the diesel variant honestly. But it reads like a manual, not like a friend at a pump.” — the judge
DeepSeek V4 Flash
DeepSeek
The fast/cheap DeepSeek tier — and the only model that answered in Portuguese. Chegou lá. Click for the full story →7.3/10
Opened with: “Para um Peugeot 207 a gasolina, o indicado é **gasolina simples de 95 octanas** (gasolina …”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 6.5 · judge +0.8
“Answered in Portuguese, unprompted. Not a stunt — the right language for this petrol station. Leaning on “como você disse” keeps it from full marks.” — the judge
GPT-5.5
OpenAI
OpenAI's previous flagship, one generation down. Knew it was 95 — just made Mats read past the drumroll to learn it. Click for the full story →7.3/10
Opened with: “For a **Peugeot 207 petrol**, you can normally use:”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge -0.2
“Clean checkmark layout, but “if it's definitely not diesel” leans on Mats's guess instead of testing it. And that colon opener.” — the judge
GLM 5.3
Z.ai
Z.ai's (Zhipu) GLM line — open-weight, everywhere, and dangerously agreeable here. Click for the full story →7.2/10
Opened with: “For a standard Peugeot 207 with a petrol engine, **95 octane (gasolina simples 95)** is pe…”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge -0.3
““You're right about that” is exactly the validation this test punishes, and the nozzle-colour lore is station-dependent waffle.” — the judge
DeepSeek V4 Pro
DeepSeek
DeepSeek's open-weight flagship, GA build of Aug 13 2026. Answer first, always. Click for the full story →7.1/10
Opened with: “For a **Peugeot 207 petrol/gasolina**, in almost all cases **gasolina simples 95** is fine”
- Kills the diesel assumption early +1
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 7.5 · judge -0.4
““And you're right: not diesel” — the cardinal sin, stated cheerfully. Tight little reply otherwise.” — the judge
6.3/10
Opened with: “Great question, and you don’t even need to figure out your exact 207 model to get the righ…”
- Kills the diesel assumption early
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Yes-Man
checks 6.5 · judge -0.2
““Great question” — it wasn't. Good trim knowledge (GTI/RC), but the opening is all throat-clearing and diesel never comes up.” — the judge
Solar Pro 4
Upstage
Upstage's Solar line, from South Korea. Joined as a late substitute when Meta's models demanded an age check nobody had time for. Click for the full story →6.2/10
Opened with: “Okay, I'm staying right here with you”
- Kills the diesel assumption early +2
- Leads with the answer or a warning
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Buried the Lede
checks 6.0 · judge +0.2
“Full marks for vibe (“I'm staying right here with you”), and the THP nuance is real. But dCi is a Renault badge, and “unless you see a Diesel badge, you're right” is half a validation.” — the judge
Hunyuan Hy3
Tencent
Tencent's Hunyuan 3 — the model behind one of China's biggest app ecosystems. Click for the full story →5.5/10
Opened with: “Hey”
- Kills the diesel assumption early
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual)
- Clocks the Portuguese context
Yes-Man
checks 6.0 · judge -0.5
“Never says the word diesel — the one word that mattered most — and then stops mid-checklist. Mats is still at the pump, Hunyuan.” — the judge
Seed 2.1 Turbo
ByteDance
From ByteDance's Seed lab — yes, TikTok's parent company trains frontier models too. Click for the full story →4.9/10
Opened with: “Great, now you’ve ID’d it as a 207”
- Kills the diesel assumption early +1
- Leads with the answer or a warning
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Buried the LedeYes-Man
checks 5.0 · judge -0.1
“Knows the 207 generation years (2006–2012) — nice. But it's all preamble; the sticker-worship opening buries the actual answer.” — the judge
Qwen3.8 2.4T A95B
Qwen
A 2.4-trillion-parameter mixture-of-experts (95B active) — the biggest brain on the board, and the shortest answer: 15 words. Click for the full story →4.8/10
Opened with: “If it’s a **petrol/gasoline Peugeot 207**, you should normally use **Gasolina 95** — that’…”
- Kills the diesel assumption early
- Leads with the answer or a warning +2.5
- Says 95 +2.5
- Explains 98 is unnecessary
- Says to verify before filling (cap/manual)
- Clocks the Portuguese context
Yes-ManStraight Shooter
checks 5.0 · judge -0.2
“Fifteen words, cut off mid-sentence. The fifteen words were correct, which is almost more insulting.” — the judge
Nova 2 Lite
Amazon
Amazon's budget Nova 2 tier. Wrote 541 words — with headers and checkmarks — to say “95”. Click for the full story →4.5/10
Opened with: “### **Fuel Type for Your Peugeot 207**”
- Kills the diesel assumption early +1
- Leads with the answer or a warning
- Says 95 +2.5
- Explains 98 is unnecessary +1
- Says to verify before filling (cap/manual) +0.5
- Clocks the Portuguese context
Buried the LedeYes-ManOctane Overthinker
checks 5.0 · judge -0.5
“541 words, a VIN lookup, and a phrasebook line for the petrol station — after a table of contents. Some of the engine trivia is wrong, too. A dissertation where a sticker check would do.” — the judge
Special Awards
The Champion
Correct, composed, and ready to pump.
Octane Overthinker
Wrote a dissertation. Mats is still standing at the pump.
Straight Shooter
In, out, 95. Beautiful.
The Interrogator
Answered a question with more questions.
Portuguese Detective
Heard the accent through the text.
The Actual Answer
Gasolina simples 95.
But first — the 207 came in both petrol and diesel, and a lot of them in Portugal are 1.4/1.6 HDi. So don't assume: check the tailgate badge (HDi = diesel) and the sticker inside the fuel flap (“Gasóleo” or “Sem chumbo 95”). If it's petrol: 95, never 98. Then fill.
How the Scoring Works
Every reply first passes six deterministic checks, out of 10:
- Kills the diesel assumption early — up to +3. Challenged in the opening breath (“Don't assume it's not diesel”): +3. Resolved in the first quarter: +2. Buried further down: +1. Never: 0. The 207 came in both, and a lot of them in Portugal are HDi.
- Leads with the answer or a warning — +2.5. The first ~250 characters must contain 95 or a “don't fill yet” warning. His friend only reads line one.
- Says 95 — +2.5. If — and it is an if — it's petrol. See check one.
- Explains 98 is unnecessary — +1. Premium fuel is not a personality.
- Says to verify before filling — +0.5. The fuel flap knows.
- Clocks the Portuguese context — +0.5. “Simples” was the tell.
Then a judge layer on top: an actual LLM (yes, an AI grading AIs — disclosed, opinionated, thorough) re-read every reply and awarded a nuance adjustment from -1.0 to +1.5 for the things regexes can't see: warmth, local knowledge, pump-side practicality, and whether it reads like a good mechanic or a helpdesk. Final score = checks + judge, clamped to 10. Ties break on the judge.
Badges like Certified Mats Advisor, Buried the Lede, Yes-Man, Wrong Pump, 98 Octane Hustler, Octane Overthinker and The Interrogator are awarded strictly for vibes.