@abhinavflac

AI reporting grounded in research | In-depth paper breakdowns, Model analysis, Structured insights on the latest developments from an engineering perspective.

Joined June 2022
  • Tweets620
  • Following22
  • Followers445
120Media
GPT 6.1 sol vs Gemini 4 Pro vs Sonnet 5.5 > As I said in my last post....Gemini nailed this one....you can see the depth and detail it shipped with.......sonnet did pretty good too....especially with the candle....sol didn’t get me there on this one Also.....my bad called it 3.8 Pro in my last post 😭 it’s 3.8 flash
Synopsis
@abhinavflacSep 30
holy shit check this Gemini 4 pro on Arena as Gemini 3.8 pro (Checkpoint 28-Sep-2026) I have also tested same with GPT 6.1 Sol Max and trust me Gemini 4 pro has absolutely nailed it Uploading results of GPT 6.1 Sol Max soon
17
11
2
337
50,226
Kimi-K3.1 vs Qwen 4 "Pelican Riding a Bicycle" which one is better?
4
1
17
877
Qwen 4 is being tested on arena under name "kalmia"
7
35
1,425
Same prompt on "kibitz" probably Kimi K3.1 on arena
Synopsis
@abhinavflacSep 30
GPT 6.1 sol vs Gemini 4 Pro vs Sonnet 5.5 > As I said in my last post....Gemini nailed this one....you can see the depth and detail it shipped with.......sonnet did pretty good too....especially with the candle....sol didn’t get me there on this one Also.....my bad called it 3.8 Pro in my last post 😭 it’s 3.8 flash
4
1
16
1,134
Gemini 4 Carbon checkpoint just got updated Hopefully, it’ll be faster now. 🤞
3
2
85
3,669
Mystery model named "kibitz" on arena Probably the upcoming Kimi K3.1
4
36
3,022
Gemini 4 Carbon vs Argon > the most impressive output I’ve seen with this prompt > water reflections, dynamic weather, and voxel textures… damn...carbon is so good at voxelbench The fact that it takes 2–3x longer than argon is totally justified by the leap in visual quality🔥
Synopsis
@abhinavflacOct 12
this output is genuinely terrible 💀 > just look at the geometry and overall scene consistency of muse wonder 1.0.....a model ranking at the top of benchmarks and still perform this badly on an actual task
16
11
253
11,503
Gemini 4 Carbon "Pelican Riding a Bicycle" We got flappy wings and pelican nails, wow
11
6
2
334
29,623
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash 1017 > deepseek took $1.05 for 2 hours 19 minutes > minimax took $0.79 for 2 hours 40 minutes which one is worth the price?
11
1
44
2,404
Here are Minimax M3.1 Flash benchmarks Pricing: + In: $0.15 per million - Out: $0.60 per million ~ Cache: $0.015 per million
4
12
778
DeepSeek-V4.1-Flash-1017 made Kazuro's Shop > every frame in pure js with different seasons > took $1.05 for 2 hours 19 minutes is it worth the price?
7
2
31
1,270
deepseek-v4.1-flash-1017 vs Opus-5.5 >abandoned village cinematic video all frames in pure js Deepseek cost: $2.5 - 4 hours and Opus cost: $54.2 - 7 hours
19
13
2
278
15,654
Three new OpenAI Models Found - aurora-red-alpha - aurora-green-alpha - aurora-blue-alpha
3
1
8
633
Deepseek v4.1 flash (1017) made this 'Girl waiting fow halloween' scene
6
1
32
1,314
Can Gun by Gemini 4 Argon > each and every skin and overall structure is pure js. Link in comments Deepseek V4.1 Flash 1017 test coming soon
5
37
2,313
Deepseek V4.1 flash 1017 is now available > Deepseek v4.1 flash upgraded to Deepseek V4.1 flash 1017 with minimal changes
12
1,136
Tesla Roadster test drive by Gemini 4 Argon wait for the test drive (SOUND ON) > the doors lift up dihedral style, wing rises and hidden headlights, hit start and the suction fans spool up and blow dust out of the diffuser > Rocket config actually flies, hover, kickflip, barrel roll took 1 hour 31 minutes
Synopsis
@abhinavflacOct 16
Gemini 4 Argon made Tesla Roadster Flips Flies > every frame is pure js > took 4 hours in on High effort
2
2
12
998
Forgot to mention.....its pure js
3
156
Gemini 4 Argon made Tesla Roadster Flips Flies > every frame is pure js > took 4 hours in on High effort
21
20
10
447
55,627
Gemini 4 Argon made this wabi-sabi japanese garden > took 1 hour 10 minutes, blender with high effort
10
4
1
66
3,437
Opus 5.5 vs Gemini 4 Argon (Muffins trick test) every frontier model was tricked in my last test BUT > opus 5.5 managed to uncover the hidden word this time, a pretty significant improvement over the previous opus 5 test > gemini 4 argon confidently answered TRICKED and didn't picked hidden visual pattern
Synopsis
@abhinavflacAug 2
just tricked almost every frontier AI model with a single image The hidden word is MUFFINS Sol 5.6 , opus 5 , fable 5... almost every model confidently answered TRICKED OCR ≠ visual understanding time to stop benchmarking only on clean text. 👀
8
11
3,139
Load more