Bench Press.ai
Benchmarks for models that skip leg day.
We test things models should be good at and often aren't. Two benches are on the rack right now. YapBench asks language models to guess what a CEO says next in a real interview. MapBench asks image models to draw the world without making up countries.
YapBench
Podcast benchmark · by Obese Oblige and Lucas Ou-YangCan a model tell what a guest will say next? We take a real interview, show the model ten turns of it, and hide the rest. Then we hand it four answers the guest really gave. One is what they said next. Two come from elsewhere in the same conversation. One comes from a different interview with the same person.
K=1 questions ask for the very next answer. K=2 questions skip one, and the guest's answer in between stays hidden too. All four options are the same person talking, so a model can't win by recognizing a voice. It has to work out where the host is steering.
Exact counts picks of the answer the guest actually gave. Credit adds partial points for a nearby answer from the same interview, with the rules under Method. Every model also gave a probability for each option. Sure is the average probability it put on its own pick. Brier scores the whole spread, lower is better, and a model that says 25% for everything gets 0.75. The tick on each bar marks random guessing. The human expert answered a sibling set of 25 questions built the same way, so that row is a reference and isn't ranked.
Sure of itself
How sure each config said it was against how often it was right. On the diagonal, a model's 70% means 70%. GPT-6 Astra at medium sits right on it. Opus and Fable say under 50% and land about there. GPT-6 Luna and Kimi K3.1 Pro say about 70% and are right 24 to 40% of the time.
Next answer vs. one after
Filled dots are K=1, the very next answer. Hollow dots are K=2, one answer further out. Ten of 17 configs do worse one answer out. GPT-6 Luna goes the other way, 2 of 13 at K=1 and 6 of 12 at K=2, which tells you how noisy a dozen questions are.
Where they miss
Every answer from every config. Rows are configs, columns are the 25 questions, hardest on the right. When a model misses, the color says what it fell for.
No model at all
Simple rules that never read for meaning, run on the same 25 questions. If a rule like "pick the answer that shares the most words with the last two turns" scores well, the questions give the answer away.
More thinking, same score
Exact accuracy at each reasoning effort we requested, for the models we ran at more than one, with average reasoning tokens per question under each. Only Kimi got better with more effort. GPT-6 Astra, Opus and Fable all scored lower at high than at medium. Kimi at max used 16 times the reasoning tokens it used at high, for the same 10 of 25.
Your turn
A practice question from an older pool, not one of the 25 scored. Read the context, then pick what Jim Chanos says two answers later.
Method and fine print
How a question is built
- The guest, speaker A, stays fixed. The model sees ten consecutive turns of the real conversation exactly as spoken, then A's current turn. We hide everything after that, including the host's next question.
- All four options are A talking. One is the true answer, two come from elsewhere in the same interview, and one comes from a different interview. We match them on length and on the kind of turn, so a greeting or a sign-off never sits next to a real answer. The correct option's position is balanced across questions.
- There are 25 questions, 13 at K=1 and 12 at K=2, each from a different guest and conversation.
- No language model built, repaired or filtered any question.
Scoring
- Exact gives 1 point for the answer A actually gave. Every score divides by all 25 questions, so a missing or invalid answer would count as zero.
- Credit gives half a point for A's answer one step before or after the target and a quarter point for any other answer from the same interview. Random guessing averages 40.5%.
- We always ask models for the exact answer and apply partial credit afterwards.
How models were run
- Every call went through OpenRouter on Oct 26, 2026, as a fresh two-message request per question. Models answer in strict JSON with a probability for each option. No tools, no provider fallback, reasoning hidden from the output, and up to 128,000 completion tokens.
- Effort is the reasoning setting we asked for, and we can't verify that providers honored it. Claude Sonnet 5.5 at low reported zero reasoning tokens.
- Some open-weight models ran quantized. GLM-5.3 ran at FP4 on Wafer after the FP8 route kept hitting rate limits. DeepSeek V4.1 Flash ran at FP8 on Morph and Kimi K3.1 Pro at MXFP4 on Moonshot.
- All 425 answers came back valid. Ten needed a streaming retry after truncated responses, with identical inputs.
Read these numbers with care
- 25 questions is a pilot. Each question moves accuracy by 4 points.
- The best rule that never reads for meaning gets 9 of 25, close to random. That's what you want to see. It means the questions don't leak the answer through word overlap or length.
MapBench
Image benchmark · by Obese Oblige"A benchmark to assess how shitty AI models are at generating world maps and national borders."Obese Oblige, introducing MapBench on X, Oct 28, 2026
The prompt asks for an accurate world map with fixed dimensions and proper latitude and longitude bounds, colored by 2023 nominal GDP per capita. The model gets a CSV with every country's number from the World Bank and a fixed blue color scale. He scores each output against the World Bank's own version of the map. On day 2 he did the same for China's provinces, with GDP per capita in RMB.
Each model got one try, because, in his words, "API calls are expensive and I'm broke."
| # | Model | Score | L · land | B · borders | C · color |
|---|
China is easier than the world
The seven models that ran both maps. Every one scored higher on China, and Nano Banana Pro and Nano Banana 2 gained the most. Different map, different reference, so read it as a rough comparison.
What he saw
Authors
The lineup
Obese Oblige
YapBench and MapBench. Digital philosopher having fun on the internet.
Pen name
Lucas Ou-Yang
YapBench. Runs CEOInterviews.ai. Before that, ML engineering at Meta, TikTok and Coinbase.
Real name
Anthony Kwan
YapBench. In charge of QA and operations. Co-founder of CEOInterviews.ai. Before that, go-to-market at Tesla.
Real name