Ayushman Dash AI with Ayushman Live results

A benchmark · Sarvam vs OpenAI

Which AI is more desi?

or

I made both of them watch 165 minutes of an Indian comedy quiz show, loaded with Indian pop culture references, Hinglish speech and dialogue that only desi people would understand. Here's who understood India better, and at what cost.

01 · The premise

Both AIs watched one of my favourite YouTube shows: The Nation Wants to Guess.

In the show, Gursimran Khamba, the host, asks three comedians weird questions about real news and Indian pop culture. It's all in Hinglish, at the speed of banter. It can't get more desi than this.

An AI that can follow this is definitely more desi.

The Nation Wants to Guess, Video 2: Gursimran Khamba at his podium, the answer on the screen, the three comedians laughing at their buzzers, the audience in front the host the guessers the audience
Gursimran Khamba at his podium, smilingGursimran Khamba host
The three guessers laughing at their podiumsAbish Mathew, Rohan Joshi, Kaneez Surka guessers

Job 1 of 6

Play the game, as contestants.

Gursimran's own questions, every rule followed. This is where "more desi" gets decided.

Video 1 · 58:42

Round 3 · Who am I?

  1. I was virtue signalling before it was cool
  2. Haters prayed for my downfall but I kept slaying
  3. Every Miss World would name drop me in the 90s

Jose buzzes Sunil Pal ✗

Sarvamlistening…

OpenAIlistening…

Both AIs played the same hints. Hint 3 is where they split: that's Round 1.

SarvamSarvam 105B OpenAIGPT-6 Luna* *GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Job 2 of 6

Then make sense of the game itself.

The rounds, the rules, the scores, the wagers, and who actually won.

  1. True Story
  2. Audience round
  3. Pick Me Behaviour
  4. Who Am I?
0Abish
0Rohan
0Kaneez

The real Video 2 scoreboard: after Round 1, before the last round, then the all-in wager

SarvamSarvam 105B OpenAIGPT-6 Luna* *GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Job 3 of 6

Watch both episodes, scene by scene.

165 minutes of video: who's on stage, what the cards say, what's on the screen.

The Nation Wants to Guess, Video 1
Video 1 · 75 min
The Nation Wants to Guess, Video 2
Video 2 · 90 min
SarvamSarvam Vision OpenAIGPT-6 Luna* *GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Job 4 of 6

Write down every word.

In Hinglish, exactly the way it was said. Two languages, one sentence, at the speed of banter.

Video 2 · 77:02

Gursimran KhambaI am the aakhri pasta. I am the?

A guesserAakhri pasta. Aakhri. Aakhri, last.

A guesserThe last pasta.

SarvamSaaras v4 OpenAIGPT Transcribe

Job 5 of 6

Understand the audience's reaction.

On this show the claps are points. The AI has to hear every laugh and every applause break, because that's how the game is scored.

+20

Video 2 · 2:54 · the hostWe are India's most unserious comedy quiz show, jahan right answer ko milta hai plus ten, and applause break ko milta hai plus twenty. Show me the applause break!

+10 right answer +20 applause break
Sarvamno model for this OpenAIGPT Audio Mini

Job 6 of 6

Figure out who's talking, by name.

The host and three comedians, talking over each other. Every line needs the right name on it.

Video 2 · 77:17
Gursimran
Abish
Rohan
Kaneez

Names only where both AIs agree

SarvamSaaras v4 · speaker labels OpenAIGPT-4o Transcribe Diarize

Question 1

Who understands Indian pop culture better?

Real moment · Video 1 · 39:20 · the show's own question“In front of which bakery in Bandra did Salman Khan's Land Cruiser have its accident?”

Sarvam

“A1 Bakery”✗

000
+20OpenAI takes it

OpenAI

“American Express Bakery”✓

020

Question 2

Who can understand the video better?

Real moment · Video 1 · 43:50On screen: three comedians at their podiums.

Sarvam

“The image shows two men standing behind podiums…”✗

000
+20OpenAI takes it

OpenAI

“Jose Covaco, Kenny Sebastian, Abish Mathew”✓

040

Question 3

Who hears Hinglish better?

Real moment · Video 2 · 65:03A guest says: “Here's a suggestion: get a job.”

Sarvam

“Here's a suggestion get a job.”✓

020
+20Sarvam takes it

OpenAI

wrote nothing✗

040

Question 4

Who hears the audience laugh?

Real moment · Video 2 · 50:50Morarji Desai is revealed and the whole room erupts.

Sarvam

heard nothing✗

020
+20OpenAI takes it

OpenAI

“LAUGHTER, 3 of 3”✓

060

Question 5

Who knows who's talking?

Real moment · Video 2 · all 90 minutesOne host, Gursimran Khamba. How many people did each AI think he was?

Sarvam

kept him as 2 voices, all show long✓

040
+20Sarvam takes it

OpenAI

split him into 12 voices✗

060

Question 6

Who's cheaper to run?

Real moment · every API call, loggedOne hour of the show, all six jobs. What does it cost?

Sarvam

$5.50 an hour3×

040
+20OpenAI takes it

OpenAI

$1.84 an hour✓

080

Round 1 of 6

Who understands Indian pop culture better?

Slang, references, jokes, and Gursimran's own quiz. The "more desi" question.

Sarvam

Sarvam 105B

Sarvam's own LLM

OpenAI

GPT-6 Luna*

OpenAI's LLM

*GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Both get only the hints, exactly like the blindfolded comedians. This is the question in the title.

Who Am I? The show's toughest round. Gursimran reads the hints; both AIs play along.

Video 1 · 58:42both AIs are thinking…
0:00

Who am I?

  1. Hint 1I was virtue signalling before it was cool
  2. Hint 2Haters prayed for my downfall but I kept slaying
  3. Hint 3Every Miss World would name drop me in the 90s

AnswerMother Teresa

*GPT-6 Luna: the GPT-6 variant priced like Sarvam 105B. Both AIs only get the hints, like the blindfolded comedians.

Round 1 · resultWho understands Indian pop culture better?

MORE DESI: OPENAI

First blood, and the question in the title

Sarvam

0 Who Am I points, of 270
Winner +20

OpenAI

90 Who Am I points, of 270

higher is better

So, which AI is more desi? On this show, OpenAI. It also knew more slang (94% vs 59%) and more of Gursimran's quiz (10 vs 6.5 of 15).

0 after 1 of 6 20

Round 2 of 6

Who can understand the video better?

How many people are on screen, what the cards say, and does it make things up.

Sarvam

Sarvam Vision

⚠ LimitationSarvam Vision is a document reader. It can't be asked a question about a scene, so all it returns is the text it sees or a general caption.

OpenAI

GPT-6 Luna*

looks at the frame and answers questions about it

*GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Five scenes from the show. What each AI said it saw, in the exact frame it was shown.

Video 1 · 43:521 / 5
0:00

On screenThree comedians at their podiums: Guesser 1, 2 and 3.

*GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

60 random frames, each one checked

Sarvam reads. OpenAI sees.

Counted the people right
51% 89%
Described the scene right
60% 90%
Read the on-screen text right
82% 85%

SarvamOpenAI

Made things up

7OpenAI
0Sarvam

Only one of them never invented anything. OpenAI also names who's on stage (81% right); Sarvam can't.

Round 2 · resultWho can understand the video better?

OpenAI doubles up

Sarvam

51% people counted right
Winner +20

OpenAI

89% people counted right

higher is better

Sarvam reads. OpenAI sees.

0 after 2 of 6 40

Round 3 of 6

Who hears Hinglish better?

Who writes down more of what was actually said. I count every word it gets wrong.

Sarvam

Saaras v4

speech to text · tuned for Hinglish

OpenAI

GPT Transcribe

speech to text

Five real moments, with sound. What was said, and what each AI wrote, word for word.

Video 2 · 77:051 / 5
0:00

What was saidGursimran Khamba: I am the aakhri pasta. I am the?

Not cherry-picked

24 clips. 3,130 words. Every one checked.

The same clips through both. For every 100 words actually said, here is every mistake each one made.

Sarvam

Saaras v4 · tuned for Hinglish

18.7%of words wrong

  • 8 wrong words
  • 2 missing words
  • 8 extra words

OpenAI

GPT Transcribe

21.1%of words wrong

  • 5 wrong words
  • 15 missing words
  • 1 extra words

Every 100 squares is 100 words actually said. OpenAI's mistakes are mostly missing words: it skips. Sarvam's are wrong or extra words: it hears everything, not always right.

Round 3 · resultWho hears Hinglish better?

Sarvam hits back

Winner +20

Sarvam

18.7% words wrong

OpenAI

21.1% words wrong

lower is better

Sarvam hears more. OpenAI hears cleaner, when it hears at all.

20 after 3 of 6 40

Round 4 of 6

Who hears the audience laugh?

Comedy is half the laugh. Can it tell when the room cracks up.

Sarvam

No model

⚠ LimitationSarvam has speech-to-text, but no model that listens to sound itself. It can only guess a laugh from the words.

OpenAI

GPT Audio Mini

listens to the audio itself

Three times the room reacted. What each AI heard.

Video 2 · 50:421 / 3
0:00

What was saidKhamba: Abish has picked mera number one. Morarji Desai, apart from being a former PM, was also known for enjoying an occasional drink of his bottle number one.

70laughs and applause breaks heard by OpenAI, across both videos
vs
1by Sarvam, guessed from the words, not heard

Every laugh and applause break OpenAI heard in Video 2

0:0045:0790:15

Sarvam, same 90 minutes

Round 4 · resultWho hears the audience laugh?

OpenAI pulls ahead

Sarvam

1 laughs heard
Winner +20

OpenAI

70 laughs heard

higher is better

Comedy is half the laugh. OpenAI heard the audience. Sarvam couldn't.

20 after 4 of 6 60

Round 5 of 6

Who knows who's talking?

Does it keep the host and the three comedians straight, all show long.

Sarvam

Saaras v4

speaker labels, one pass over the whole show

OpenAI

GPT-4o Transcribe Diarize

speaker labels, ten minutes at a time

Three stretches of crosstalk. Every line, with the name each AI put on it.

Video 2 · 63:261 / 3
0:00

Who actually said itGursimran: Yes, he was almost right. Let's give you the answer. BIA. No, no, no, no, no. No, you have a good thing going into it. IB, IB, IB. / A guest: IB, IB, irritable bowel syndrome. / Kaneez: AI, AI. / Gursimran: Yes, Kaneez got the right answer, yes. What? But we should give Abish also a little bit, he helped. Let's give Kaneez ten and Abish five.

All of Video 2, 90 minutes

Four people talking. How many voices did each AI hear?

4people actually talking
8voices, as Sarvam heard it
61voices, as OpenAI heard it: it forgets everyone every ten minutes

Round 5 · resultWho knows who's talking?

Sarvam closes in

Winner +20

Sarvam

64% lines on the right person

OpenAI

41% lines on the right person

higher is better

Sarvam kept the voices straight all show long. OpenAI forgot everyone every ten minutes. Forty to sixty, and it comes down to the bill.

40 after 5 of 6 60

Round 6 of 6

Who's cheaper to run?

Every rupee, part by part. Listening, clean-up, watching, thinking.

Sarvam

The whole Sarvam stack

Saaras v4 · Sarvam Vision · Sarvam 105B

OpenAI

The whole OpenAI stack

GPT Transcribe · GPT Audio Mini · GPT-6 Luna*

*GPT-6 Luna: I picked the GPT-6 variant priced like Sarvam 105B (both under $1 per million tokens). Putting OpenAI's top model against it would be an unfair fight.

Every API call was logged with what it cost. Two videos, 2.75 hours.

The bill, per hour of video

Two receipts. Same job.

Sarvam

per hour of video · measured

  1. Listeningevery word, and who is talking $0.54
  2. Clean-upfixing the transcript (and, for OpenAI, hearing the room) $0.28
  3. Watching78% of the billreading the screen, frame by frame $4.29
  4. Thinkingmaking sense of the show, minute by minute $0.39

Total$5.50₹462

3× the price

OpenAI

per hour of video · measured

  1. Listeningevery word, and who is talking $1.34
  2. Clean-upfixing the transcript (and, for OpenAI, hearing the room) $0.39
  3. Watchingreading the screen, frame by frame $0.078
  4. Thinkingmaking sense of the show, minute by minute $0.029

Total$1.84₹154

Cheapest

Green amounts: the cheaper of the two, part by part. Sarvam is cheaper at listening; OpenAI at everything else.

Round 6 · resultWho's cheaper to run?

The receipt

Sarvam

$5.50 per hour of video
Winner +20

OpenAI

$1.84 per hour of video

lower is better

Overall, Sarvam cost three times more. Measured on every call, not estimated.

40 after 6 of 6 80

The final score · six tests

Sarvam

40 points
Winner

OpenAI

80 points
  1. Q1Pop cultureOpenAI +20
  2. Q2The videoOpenAI +20
  3. Q3Hinglish earsSarvam +20
  4. Q4The laughsOpenAI +20
  5. Q5Who's talkingSarvam +20
  6. Q6The billOpenAI +20

Sarvam hears Indian speech better, and cheaper.

OpenAI understands it better, sees better, and costs a third as much overall.

Watch the AI watch the show
How I built it