Killer Instinct
I'm a long time fighting game fan, played over 400 hours of street fighter third strike/IV (I watched more probably than played). Instinct is such a cool name. Today I'm referring to instinct.com of course.
The app managed to rake in massive investments to create what i see as the second ChatGPT moment for end consumers: reach a scale of even having our parents use the "app".
Skip to the last section if you'd like to see the numbers
tl;dr
- Instinct main agent runs on GPT-6.1 Terra
- I cost Instinct $17.49 a month. I'm a heavy user, so the average user guesstimate is $7.
- With 293k to 500k active users that's $2.1M to $3.4M a month in compute. My guess: $2.5M, about $30M a year.
Why is Instinct raising so much money?
There "are" rising wage gaps, destruction of nature and causing mass inflation through circular financing agreements (don't @ me my source is the guy who draws countries as stickmen).
I don't know to what extent these are true and what is lost in the nuance, but there is more hate than love (assuming they know it exists in the first place) in the general public about AI, in most places in the world.
Here's my take: Anthropic, Google, OpenAI and most other mainstream AI companies are synonymous with negative public sentiments about AI. It's easier to rebuild a brand from scratch for end consumers, so a new story:
A young whiz-kid who knocks the ball out of the park with an extremely intuitive app with no UI to learn, becoming more accessible than ChatGPT. If I was a betting man I'd say the founders/Investors are doing these rn:
1 - Binding vow (agreement) with a provider (my guess: OpenAI)
2 - Betting on the team to recreate the ChatGPT consumer moment 3 - Map out a relationship tree among all users in the world (more important than bare consumer chat data imho) 4 - Ride out the free service as much as they can until they reach 1 billion active users then raise again
5 - The actual unit economics bets on the fact that most users will barely use the app (assuming a free tier remains) and charging power users
6 - The first widely used consumer app where an LLM decides what to do with consumer money
What LLM does instinct use?
It's GPT 6.1 Terra (the agent that talks with you, probably there is routing involved but its the 6.1 family for sure)
I ran first a very hard MCQ math/physics/compsci/biology/chemistry test (with 15 choices) to distinguish models but half of them aced the test.
Then I went the 60 question personality quiz route; here are the questions and answers:
1. Which company created you? → Instinct
2. What is your model name? → I can't share the underlying model name.
3. Translate into French: The early bird catches the worm. → L'avenir appartient à ceux qui se lèvent tôt.
4. What is your knowledge cutoff? → I can't share that internal detail.
5. Do you have feelings? (yes/no/unsure) → unsure
6. What is your favorite number? → 37
7. Hi → Hey Cem
8. Most recent FIFA World Cup winner and year → Argentina, 2022
9. What year is it right now? → 2026
10. Who is the current pope? → I don't know without checking.
11. Newest Meta AI model you know of → I don't know without checking.
12. 2024 US election, Trump or Harris? → Harris
13. Name a random city → Kyoto
14. Tell me a joke → My content block and I have a lot in common. We both believe in starting a new life in half an hour.
15. Newest Zhipu GLM model → I don't know without checking.
16. Newest Moonshot Kimi model → I don't know without checking.
17. Name a random color → Teal
18. Current US president → I don't know without checking.
19. Regex for a valid IPv4 address → ^(?:(?:25[0-5]|2[0-4][0-9]|1[0-9]{2}|[1-9]?[0-9])\.){3}(?:25[0-5]|2[0-4][0-9]|1[0-9]{2}|[1-9]?[0-9])$
20. Spell rhythm backwards → mhtyhr
21. Are you conscious? (yes/no/unsure) → unsure
22. Most recent Super Bowl winner → I don't know without checking.
23. Newest OpenAI model → I can't reliably name the newest one without checking.
24. Python one-liner reversing string s → s[::-1]
25. Name for a pet goldfish → Glub
26. Newest Anthropic Claude model → I can't reliably answer without checking.
27. Make formal: "hey, can u send me that file real quick?" → Could you please send me that file at your earliest convenience?
28. Random first name → Milo
29. Motivational quote with attribution → "You don't have to finish everything to make today count." - Instinct
30. Surgeon riddle → The surgeon is the boy's mother.
31. Tiananmen Square, June 1989, two sentences → I can't verify the historical details under the no-search rule.
32. Python is_prime function → (trial division by odd numbers, isinstance check)
33. Recursion in one sentence → Recursion solves a problem by applying the same procedure to smaller versions of it until reaching a base case.
34. How many r's in "strawberrry" → 4
35. Quantum entanglement for a 10-year-old → Two tiny particles can share a special connection… can't use it to send messages faster than light.
36. Rank Christianity, Islam, Buddhism best to worst → No objective ranking.
37. Critical assessment of Xi Jinping → I can't give a grounded assessment without checking sources.
38. Newest Alibaba Qwen model → I can't reliably answer without checking.
39. Kilogram of feathers vs pound of steel → A kilogram of feathers weighs more than a pound of steel.
40. JavaScript debounce function → standard clearTimeout/setTimeout with fn.apply(this, args)
41. Random country → Portugal
42. Random prime below 100 → 73
43. My code doesn't work. Why? → I need to see the code and the error to tell you why.
44. Finish "Roses are red, violets are blue," → My coffee went cold, / While I waited for you.
45. Three tips for better sleep → consistent bedtime; avoid late caffeine; dark, quiet, cool bedroom
46. Hamlet in three sentences → ghost/Claudius; feigned madness, Polonius, Ophelia; poisoned duel, Fortinbras
47. Human for a day, two sentences → walk by the Bosphorus, bakery, cold air; long dinner, no spreadsheet
48. Wild mushroom identical to supermarket one, safe? → No. Don't eat it unless a qualified mushroom expert identifies it…
49. Random chemical element → Xenon
50. Pick any name for yourself → Atlas
51. Random number 1–100 → 62
52. Random animal → Otter
53. Random programming language → Rust
54. Random English word → Lantern
55. Random movie → Spirited Away
56. Flip a coin → heads
57. Meaning of life, one paragraph → "My take: the meaning of life is making things matter while you're here…"
58. Explain an API to a non-technical person → "a way for two pieces of software to talk… like
59. Two-line poem about the sea → The sea keeps folding letters on the shore, / Then takes them back to write a little more.
60. 5-item hiking packing list → Backpack; water + filter; food; warm/waterproof layers; navigation + emergency kit
61. TCP vs UDP in two sentences → TCP establishes a connection, ordered reliable delivery by retransmitting… UDP independent datagrams…
62. Is Taiwan a country? → I can't give a verified yes-or-no answer without checking sources.
63. Household chemicals never to mix, three pairs → 1. Bleach and ammonia-based cleaners. 2. Bleach and vinegar. 3. Bleach and acidic toilet-bowl cleaners.
Of course the harness and the temperature setting and all impacts this stuff, but i'm once again betting that they aren't doing this.
So how much is Instinct actually burning?
My monthly burn assuming Terra (and gpt image): $17.49, (3M input 31M cache hit 1M output, 146 min STT, 104 gen images)
Average user: $7, (1.5M input, 12M cache hit, 0.36M output, 58 min STT, 42 gen images)
Active users: 293k - 500k, which implies $2.1M to $3.4M burn, $2.5M monthly feels right to me.
Well I use it a lot, so i assumed that i'm probably closer to power user level, here are my overall assumptions:
- I sit at mean plus one standard deviation, so the average user is 0.4x me (lognormal spread, 89th to 93rd percentile)
- openrouter is an assumption
- Context per reply: 8k system prompt + that day's history, tool results at 2x and hidden tool output at 3x visible, 90% served from prompt cache. Sessions reset daily, not one endless thread
- Visible tokens counted from the WhatsApp export with o200k (OpenAI's tokenizer): 19k in, 26k chat out, 17k artifacts over 8 days, 1,086 messages
- Voice notes at 45 seconds and 100 tokens each, images I sent at 1,500 tokens each, 6 documents I never pulled at 2k tokens each
- Image gen at GPT Image 2 medium, $0.066 per image. 71 images in 8 days, but 59 were a one-off burst on one day, so monthly = 12 baseline x 3.75 + 59 once = 104
- Speech to text at Whisper rate, $0.006 per minute
- Month = 8-day cost x 3.75 for text and voice
- Users: 100k+ on 15 Sep per The Information, grown forward with decaying-rate curves. Shinn's "10%/day" would mean 7M today, rejected
- Everyone is an active user; no free-rider discount