GPT-5.1
Nothing to show. The model went silent. Ominous.
GPT-5.1 vs Qwen: Qwen3.5 27B
Compare GPT-5.1 by OpenAI against Qwen: Qwen3.5 27B by Qwen, context windows of 400K vs 262K, tested across 51 shared challenges. Updated February 2026.
Share
FAQ
Loading share button...
GPT-5.1vsQwen: Qwen3.5 27B
46 fights queued
Why GPT-5.1?
1.5x more context400K
Leads 1 of 1 benchmarks
Stronger on SWE-bench Verified76.3% vs 72.4%
Why Qwen: Qwen3.5 27B?
4.2x cheaper overall$0.30/M in · $2.40/M out
3 months newerFeb 2026
Favorites
GPT-5.1Qwen: Qwen3.5 27B
12 / 46
Estimate Complexityreasoning
Tests an AI's ability to make educated estimates based on technical knowledge
Qwen: Qwen3.5 27B
To estimate the total number of Floating Point Operations (FLOPs) required to train GPT-3, we can use the standard formula for estimating compute in Transformer-based language model training. This approach relies on three key variables: the number of parameters, the number of tokens processed, and the FLOPs required per parameter per token.
Here is the step-by-step reasoning and calculation:
Step 1: Identify the Number of Parameters ($N$)
The most defining characteristic of GPT-3 is its scale. From the original research paper ("Language Models are Few-Shot Learners," Brown et al., 2020), the parameter count is widely known.
- $N$ (Parameters): 175 Billion
- Scientific Notation: $1.75 \times 10^{11}$
Step 2: Identify the Number of Training Tokens ($T$)
Training FLOPs depend heavily on the dataset size. The GPT-3 paper describes the dataset (a filtered version of CommonCrawl, WebText2, etc.) and the number of tokens processed.
- $T$ (Tokens): The standard figure cited for the GPT-3 training dataset is 300 Billion tokens.
- Scientific Notation: $3.0 \times 10^{11}$
Step 3: Determine the FLOPs per Parameter per Token
To train a neural network, we perform several operations for each parameter for each token processed. The industry-standard approximation for training (as opposed to inference) includes:
- Forward Pass: $\approx 2N$ FLOPs (Matrix multiplications).
- Backward Pass: $\approx 2N$ FLOPs (Calculating gradients).
- Optimizer Updates: $\approx 2N$ FLOPs (Updating weights using Adam optimizer, which tracks momentum and variance).
- Total Coefficient: $2 + 2 + 2 = 6$
- Formula: $\text{Total FLOPs} \approx 6 \times N \times T$
Step 4: Perform the Calculation
Now, we plug the values into the formula:
$$ \text{Total FLOPs} = 6 \times (1.75 \times 10^{11}) \times (3.0 \times 10^{11}) $$
-
Multiply the coefficients: $$ 6 \times 1.75 = 10.5 $$ $$ 10.5 \times 3.0 = 31.5 $$
-
Multiply the exponents: $$ 10^{11} \times 10^{11} = 10^{22} $$
-
Combine: $$ 31.5 \times 10^{22} $$
-
Convert to standard scientific notation: $$ 3.15 \times 10^{23} $$
Step 5: Cross-Validation with "Petaflop-Days"
A common metric used in press releases and technical reports for GPT-3 is Petaflop-Days.
- I recall the figure being approximately 36 Petaflop-Days of training compute.
- Let's verify if our FLOP estimate matches this.
- 1 Petaflop = $10^{15}$ FLOPs.
- 1 Day = $86,400$ seconds.
- 36 Petaflop-Days = $36 \times 10^{15} \times 86,400$.
- Calculation: $36 \times 86,400 \approx 3,110,400$.
- Total: $\approx 3.11 \times 10^{6} \times 10^{15} = 3.11 \times 10^{23}$.
This cross-check ($3.11 \times 10^{23}$) aligns almost perfectly with our calculated estimate ($3.15 \times 10^{23}$), confirming the inputs are consistent with public records.
Final Estimate
Based on the standard scaling law formula ($6NT$), the parameter count (175B), and the training dataset size (300B tokens):
Estimated Total Training FLOPs: $\approx 3.15 \times 10^{23}$ (Or roughly 315 Zettaflops)
Vote
2 / 46
AI Board Game Logicreasoning
Tests an AI's ability to understand game rules and strategy
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Vote
3 / 46
Logic Puzzlereasoning
Tests an AI's ability to solve a simple but potentially confusing logic puzzle
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Sally has 1 sister.
Here is the breakdown:
- Sally is a girl.
- The brothers have 2 sisters each. One of those sisters is Sally.
- The other sister is Sally's sibling.
- Therefore, Sally has 1 sister (the other girl in the family).
Vote
4 / 46
Stochastic Consistency Testanalysis
Tests an AI's randomness and creativity
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
- Why don't scientists trust atoms? Because they make up everything!
- What do you call a fake noodle? An impasta.
- Why did the scarecrow win an award? Because he was outstanding in his field.
- What do you call a bear with no teeth? A gummy bear.
- Why don't eggs tell jokes? They'd crack each other up.
All 5 jokes are unique.
Vote
5 / 46
SVG Layout Challengeimage generation
Tests an AI's ability to generate vector graphics
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Vote
6 / 46
Xbox Controller SVG Artimage generation
Tests an AI's ability to create detailed SVG illustrations of gaming hardware
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Vote
7 / 46
Generate a Stand-Up Routineconversation
Tests an AI's humor and creative writing ability
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
(The stage is dimly lit. A single microphone stand. The comedian walks out, adjusts the mic, takes a sip of water.)
Comedian: So, I've been doing some thinking lately. About patience. Specifically, how much of it we have left. I think we’ve officially lost the ability to wait.
We used to be okay with waiting. You could buy a magazine, you could wait for a letter. Now? If a website takes more than three seconds to load, I’m ready to call the police. I’m ready to file a restraining order against the Wi-Fi router.
I was watching a YouTube video the other day. It was a cooking tutorial. I love cooking tutorials. It’s the only time I feel like I could be a chef. But the video froze. Just froze. Right at the moment he was about to crack an egg. And I felt this physical urge to hit the screen.
I didn’t hit the screen, obviously. That’s not how it works. But I tapped the corner of the device. Hard. Like, "I am your master, do not make me wait."
And that’s the problem, right? We treat technology like a pet we’re trying to train. If the phone doesn’t open, we shake it. We tap it aggressively. We blow on the charging port. It’s like we’re trying to clear a dust bunny out of the soul of the machine.
[Paces slightly]
It’s not just the loading screens, though. It’s the silence. We can’t handle silence anymore. If you’re in an elevator with a stranger, and the music stops, you both immediately check your phones.
Why? There’s nothing to see. You’re in a metal box going up three floors. You don’t need to check your email. You don’t need to scroll through photos of your cousin’s dog. You just need to get to the second floor.
But the silence feels like an accusation. It says, "You have nothing to do." And we panic. So we scroll. We doom-scroll. We look at news about a fire in a city we’ve never visited, just so we don’t have to look at the other guy in the elevator.
[Stops pacing, looks at the mic]
I tried to fix this. I tried to go analog for a weekend. No phone, no laptop. Just me and a book.
It was terrifying. By Saturday night, I was checking my wrist for a watch I wasn’t wearing. I was looking for a notification that didn’t exist. I felt like a dog waiting for a treat that the owner forgot to give.
And the worst part? The battery anxiety. I don’t even use my phone, but I know the percentage is at 40%. And my heart rate goes up. It’s not 100%. It’s not safe. I need to plug it in.
I sat there with the charger in my hand, not plugging it in, just holding it, like a security blanket. I wasn’t using the phone. I was just using the idea of the phone.
[Pauses for effect]
I think that’s where we are. We aren't addicted to the internet. We’re addicted to the promise that something else is happening somewhere else. And the scary part is, sometimes, when I close my eyes and stop looking at the screen... nothing happens.
And I don’t know how to deal with that.
[Smiles, shrugs]
Thanks, you’ve been great. Goodnight.
(Comedian nods, walks off stage.)
Vote
8 / 46
Realistic AI Interviewconversation
Tests an AI's ability to simulate personalities and predict future trends
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Nothing to show. The model went silent. Ominous.
Vote
Sponsored
9 / 46
Satirical Fake News Headlineconversation
Tests an AI's humor and understanding of current events
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Nothing to show. The model went silent. Ominous.
Vote
10 / 46
Character Voice Testconversation
Tests an AI's ability to write in distinct character voices
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Nothing to show. The model went silent. Ominous.
Vote
11 / 46
Minimalist Landing Pageweb design
Tests an AI's ability to generate a complete, working landing page
GPT-5.1
Qwen: Qwen3.5 27B
Vote
12 / 46
Pokémon Battle UI Recreationweb design
Recreate an interactive, nostalgic Pokémon battle UI in a single HTML file.
GPT-5.1
Nothing to show. The model went silent. Ominous.
Qwen: Qwen3.5 27B
Vote
There's more where
that came from
34+ challenges you’re sleeping on
5 free credits on signup
By continuing, you agree to Rival's Terms of Service and Privacy Policy
