Why Lattu
A chatbot can name famous universities. You need to know where you can actually get in.
In October 2026 we gave OpenAI's GPT-6, Anthropic's Claude and Google's Gemini the same questions a student asks us, then scored all four of us against real admission decisions none of us had seen. This page shows exactly what came back, how we tested it, and why Lattu answers the way it does.
Real decisions
164,276
real admit and reject decisions behind every probability we show you
Universities
7,187
universities in 50 countries, with every tuition, admissions, programme and funding record carrying a link to the page it was read from
Asked again
14 of 14
schools stay the same when you ask us again. No chatbot we tested managed that
The short version
Four differences that decide where you apply
If you read nothing else on this page, read these. Everything below is the evidence for them.
Your odds, not a reputation list
A general AI assistant was trained on what the internet says about universities, so it answers with the famous ones. Lattu predicts the chance that you get an offer from a specific programme, learned from 164,276 real admission decisions across 995 universities. Anything under a 10% chance never reaches your list, because an application fee spent there is a fee wasted.
The same answer tomorrow as today
Ask a chatbot the same question five times and you get five different lists. Ask us twice and you get the same 14 schools with the same numbers on each one. That is what lets you plan, and what lets you explain your list to your parents or a counsellor without the story changing.
A number you can check yourself
Every tuition fee, deadline and entry requirement we show you is stored with the address of the university page it was read from, and the panel shows it as a Source link. If you do not believe a figure, click through to the university's own page and judge it for yourself. We would rather you checked.
Built for an application from Bangladesh
Costs in BDT next to USD, airfare priced from Dhaka, deadlines against Bangladesh time, and your HSC and SSC grades converted rather than asked for as a 4.0 GPA you do not have. A general assistant treats you as a generic international applicant, because that is all it knows about you.
Test one
Ask the same question five times, get five different answers
We built five Bangladeshi applicant profiles, the kind of student who signs up with us, and asked each chatbot to shortlist universities. Same profile, same wording, a fresh chat every time, five times each. Then we wrote down every university each one named.
One profile was a student looking for a Management master's in the UK or Germany. Across its five answers, GPT-6 named 39 different universities, and not a single one of them appeared in all five lists. Gemini named 37. Claude named 27. Lattu was asked for 14 schools and returned the same 14 both times, with the same probability on each one.
Same profile, five fresh chats: how many of the same schools came back
Each block is one school, identified by its letter. Each row is one chat with the same student profile.
GPT-6
Five runs, five different shortlists
Run 1
GPQLBNIATRDKCM
Run 2
IQSLFTNEHBMCKG
Run 3
QCHOJFKRISMDAL
Run 4
BPINTODJGMEFAR
Run 5
DHGRJSBNKAQCFT
39 different schools named, none in all five lists
Lattu AI
Asked twice: the same 14 schools, the same odds
Run 1
MHFSAKICBTOJPR
Run 2
MHFSAKICBTOJPR
14 schools, identical both times
Universities that came back in every one of the five answers
Higher is better. Each chatbot was run five times per student, Lattu twice. Average across the five student profiles.
Lattu AI
14.0
Gemini 4 Argon
7.0
Claude Opus 5.5
6.8
GPT-6
5.2
0714
| Student profile | Claude Opus 5.5 | GPT-6 | Gemini 4 Argon | Lattu AI |
|---|---|---|---|---|
| MS Computer Science, US or Canada | 28 named, 5 every time | 31 named, 3 every time | 23 named, 7 every time | 14 named, all 14 every time |
| MSc Management, UK or Germany | 27 named, 7 every time | 39 named, 0 every time | 37 named, 2 every time | 14 named, all 14 every time |
| PhD Electrical Engineering, US | 26 named, 3 every time | 26 named, 6 every time | 23 named, 8 every time | 14 named, all 14 every time |
| MSc Computer Science, Germany | 22 named, 8 every time | 26 named, 8 every time | 29 named, 5 every time | 14 named, all 14 every time |
| Master of Engineering, Australia | 18 named, 11 every time | 20 named, 9 every time | 15 named, 13 every time | 14 named, all 14 every time |
It is not just the names that move. The odds move too.
In our 15 October run against Claude, asked five times about the same student and the same university, it called Arizona State a 75% Safety, then 60%, then 60%, then a 55% Target. Texas A&M went from a 30% Target to 22% to a 15% Reach. Nothing about the student changed between those answers.
A probability that moves is a plan that moves. You cannot tell your family which school is your safe option, you cannot decide how many applications to pay for, and a counsellor cannot defend the list to anyone who questions it.
Claude's chance for Arizona State, same student, run after run
Chance of admission shown in each answer. The axis starts at 30%.
There is a second habit worth knowing about. A chatbot fills your reach slots with names it has read about most, not with schools that are actually within range. For one PhD Electrical Engineering profile in that same run, Claude offered MIT, Stanford and Berkeley at a 4% to 6% chance. Those are three application fees, three sets of documents, and three months of waiting for an answer that was never likely.
Test two
Every chatbot we tested told students their chances were worse than they were
This is the test that matters most, and it is the one you can check the hardest, because real students already know how these applications ended.
We took 293 real admission decisions where the outcome was already known, hid the outcome, and asked every system for the chance of an offer. 150 of them were mostly United States decisions, where 58 out of every 100 applicants were admitted. The other 143 were outside the US, where 68 out of 100 were admitted. Every system was scored on exactly the same decisions.
A well calibrated system should average close to what actually happened. Here is where each one landed.
Mostly United States decisions: average chance each system predicted
150 decisions where 58 out of 100 applicants were admitted. Closer to the dashed line is better.
What actually happened: 58%
Lattu AI
60%
Gemini 4 Argon
50%
Claude Opus 5.5
49%
GPT-6
31%
0%50%100%
Decisions outside the US: average chance each system predicted
143 decisions in Canada, the UK, France and Germany, where 68 out of 100 applicants were admitted.
What actually happened: 68%
Lattu AI
67%
Gemini 4 Argon
60%
Claude Opus 5.5
54%
GPT-6
47%
0%50%100%
GPT-6 was the furthest from reality. On decisions where 58 out of 100 applicants were admitted, it put the average applicant at 31%. Of the 65 applicants it gave less than a 20% chance, 37% of them were admitted.
What a pessimistic number costs you
Outside the US, where 68 out of 100 of these applicants were admitted, the three chatbots averaged between 47% and 60%. A student is told they have around a coin flip chance when two out of three students like them actually got in.
If you believe that, you apply down a level to feel safe, you pay those application fees, and you leave out the universities that would have taken you. Being 8 to 21 points low is not a rounding error. It changes where your year and your money go.
What the chatbots told a student, against what actually happened
Applicants outside the US, on real decisions. 68 out of 100 were admitted.
47% to 60%what the three chatbots averaged
68%actually admitted
8 to 21 points too low
30%40%50%60%70%80%
For readers who want the benchmark numbers
Brier score measures how far a predicted chance sat from what actually happened, so lower is better. Anything worse than the bottom row is worse than ignoring the student and quoting the national average, which is where GPT-6 landed on the US decisions. Ranking AUC asks a different question: given two applicants, does the system put the stronger one higher. A system that ignores the applicant scores 0.50 by definition, and 1.00 is perfect.
| System | Brier | AUC |
|---|---|---|
| Lattu AI | 0.188 | 0.79 |
| Claude Opus 5.5 | 0.204 | 0.75 |
| Gemini 4 Argon | 0.203 | 0.76 |
| GPT-6 | 0.289 | 0.72 |
| Guessing the rate | 0.244 | 0.50 |
| System | Brier | AUC |
|---|---|---|
| Lattu AI | 0.204 | 0.66 |
| Claude Opus 5.5 | 0.201 | 0.76 |
| Gemini 4 Argon | 0.207 | 0.70 |
| GPT-6 | 0.253 | 0.67 |
| Guessing the rate | 0.218 | 0.50 |
What we are not claiming
On the decisions outside the US, Claude ranked applicants against each other better than we did, 0.76 against our 0.66, and edged us on Brier score, 0.201 against our 0.204. That is in the same test, so it belongs on the same page. What decides your shortlist is the level of the number, because that is what sets how many reach, target and safety schools you apply to, and all three chatbots sit 8 to 21 points low there.
The decisions are self reported by applicants to GradCafe, a public admissions results board. 150 and 143 decisions are enough to see a gap this size and not enough to break down university by university. A probability is never a promise, and we say so on your shortlist too.
How we tested
The method, in full, so you can judge it
A comparison is only worth reading if you can see how it was run. Everything here is repeatable by anyone with accounts on the three assistants.
The students
Five profiles
Five Bangladeshi applicant profiles, each with real grades, test scores and experience of the kind our students have: MS Computer Science aiming at the US or Canada, MSc Management in the UK or Germany, PhD Electrical Engineering in the US, MSc Computer Science in Germany, and a Master of Engineering in Australia.
The chatbots
Three, at their best setting
Claude Opus 5.5 ran as a plain assistant with no tools. GPT-6 ran through OpenAI's Codex assistant with web search off. Gemini 4 Argon ran as Google's top model on the Gemini Pro plan, inside the Gemini app's temporary chats, so no saved memory or earlier conversation could help it.
Each one got a fresh chat for every single run, and each shortlist question was asked five times. Lattu was asked for 14 schools, twice per profile.
The scoring
Decisions nobody had seen
Real admission decisions were held back and the outcomes hidden. Every system predicted a chance of an offer for each one. Gemini answered every other batch of them, 293 in all, so the published comparison uses those 293 and every system is scored on exactly the same decisions. No system got an easier set than another. The three way test ran on 16 October 2026, after a first run against Claude on 15 October.
What we did not do
The honest limits
We did not give the chatbots our data, our tools or a second attempt at a bad answer, because a student would not either. We did not test essay writing, interview practice or visa questions, which is work a general assistant does well. This test is about one job: deciding where to apply.
Under the hood
Why Lattu answers the way it does
A chat model predicts the next word in a sentence. Your shortlist comes out of something else entirely, and it is worth knowing what.
Trained on outcomes
A decision model learned from 164,276 real admit and reject decisions across 995 universities, plus about 47,000 applicant profiles labelled by experts across 109 home countries. It is checked against official admit rates in more than 15 countries and 26 graduate schools.
Exam countries scored directly
Where places go by national exam rank instead of a file review, your chance is the chance the programme's next cut off lands at or beyond your rank. Eleven exam systems across twelve countries work this way, including JEE, YKS, CAO, ENEM, ATAR and TCAS.
A list with a shape
14 schools, sorted into Reach at 10% to 35%, Target at 35% to 70% and Safety at 70% and above. The mix follows how ambitious you want to be. Below 10% never appears at all.
Honest about safety schools
Few universities give an international applicant a 70% chance, so when there are not enough true safeties, we take the strongest remaining targets from 55% up and label them as a near safety with the real number showing. A labelled near safety is more useful than a list with no safety at all.
Requirements checked first
Programmes you clearly cannot be considered for are dropped, and the ones we could not verify are flagged. An application fee should never be spent on a door that was closed before you started.
It remembers you
Upload your resume and transcript once. We read out your GPA and its trend, test scores, work and research experience, publications and leadership, and the profile feeds every shortlist after that. You never retype yourself to get a second opinion.
How the mix of 14 schools changes with how ambitious you want to be
Each block is one school on your shortlist.
Aggressive
6 · 6 · 2
Balanced
4 · 6 · 4
Conservative
2 · 6 · 6
ReachTargetSafety
The information underneath
Where the fees, deadlines and requirements come from
A probability is only as good as the university information behind it. This is the part a chat model has no answer for, because it keeps nothing between one conversation and the next.
| What | Count |
|---|---|
| Universities | 7,187 in 50 countries |
| Collected records in total | 9,206,974 |
| Programme records | 2,765,529 |
| Admissions records | 691,848 |
| Funding and scholarship records | 778,428 |
| Tuition records | 305,482 |
| Records carrying an application deadline | 295,070 |
Every value carries its source
Not optional, not sometimes
In the live database, 100% of our programme, admissions, funding and tuition records store the address of the page they were read from. The admissions and tuition panels show it as a Source link beside the number, and about 1,520 universities are fully linked across every panel. If a figure looks wrong to you, one click takes you to the university's own page.
Unsure answers are not shown
Filtered before you see them
Every record carries a score for how confident the reading was, and anything below a half is filtered out before it ever reaches a panel. Where two pages disagree, the better supported and then the newer value is the one you see. Average confidence runs at 0.97 on tuition, 0.95 on admissions and 0.94 on programmes and funding.
The university's own words
Nothing tidied up
Values are stored exactly as the page printed them: "$315.16 per credit", "12th day of class", "about 8000". We do not round them, convert them or rewrite them into a house format, so what you read is what the university said and you can judge it for yourself.
We search, we do not crawl
Which is why we can refresh it
For every university we run one targeted search per kind of information: admissions, tuition, programmes, departments, funding, contacts, faculty, student life. A blind crawl of one university's website read 260 pages and found none of the five things we needed. One search each found all five. Going straight to the right page is also what makes re checking affordable, which is what keeps a deadline current instead of two years old.
A field means the same thing everywhere
Informatik is Computer Science
The same subject is named differently in every country. We map programmes onto CIP, the standard field of study classification, so your field can be matched across countries without guessing at names, and that is what the STEM OPT eligibility check reads. 139,352 programme records across 1,902 universities are coded this way. Where a university's catalogue is thin, published research from OpenAlex fills the gap, and where a match is not certain the university stays unmatched rather than guessed at.
Freshness, where it counts most
Deadlines first
Not every fact moves at the same speed. Deadlines and intake dates can change mid season with a few days' notice, and getting one wrong costs you the application outright. Fee schedules move once a year when the new schedule is published. Office phone numbers barely move at all. Because we go straight to the right page rather than crawling a whole site, re checking a university is cheap enough to do where it matters most, and every value is stored with the date it was last read and the page it came from.
After the list
A shortlist is where the work starts, not where it ends
A chatbot's answer ends when the message does. Your shortlist turns into a plan you can follow, one stage at a time, with dates, costs and documents attached.
- 1
Choosing the right school
Which universities fit your profile, budget and goals.
- 2
Requirements and deadlines
What each school asks for, and by when, against Bangladesh time.
- 3
Immigration
The visa and legal steps for your destination.
- 4
Financial planning
Tuition, living and travel costs, in taka next to dollars.
- 5
Housing options
Where you can live once you get there.
- 6
Flights and travel
Priced from Dhaka, not from an average international student.
- 7
Post-arrival and student life
What to do in your first weeks after you land.
- 8
Long-term pathways
Jobs and settlement after you graduate, so the plan reaches past the degree.
An application tracker
53 tasks across 8 stages, from preparing documents to arriving, so nothing slips between the steps above.
Scholarships matched to you
Funding matched to your destinations, from a set of 778,428 funding and scholarship records.
For counsellors, agencies and universities
A list you can defend in front of a parent
If you advise students for a living, the problem with a chatbot is not that it is unhelpful. It is that you cannot stand behind what it said.
Reproducible by design
Same profile in, same list out
Run a student's profile twice and you get the same 14 schools with the same probabilities. When a parent asks why a university is a target and not a safety, the answer is the same this week as last, and it comes with the number it was based on.
Checkable, line by line
Every figure links to its page
Each fee, deadline and requirement carries a link to the university page it came from. A disputed number is settled in a minute rather than defended on reputation, and if it is wrong, tell us and we refresh that university.
It gets better with every placement
The part that compounds
When a student is placed, the outcome goes into the tracker, and the model is retrained on it. Predictions improve for the next student with a similar profile, and they improve fastest exactly where our students apply. That record of real outcomes is not something anyone can buy.
Questions
The things students ask us about this
Is Lattu just ChatGPT with a Lattu label on it?
No. Your shortlist and your probabilities come from a decision model trained on real admission outcomes, not from a chat model writing a plausible answer. We do use a language model for one job, reading university pages into structured records, and every record it produces keeps the page it came from and a confidence score, which is why we can show you the source and drop the readings we are unsure about.
Where does my percentage actually come from?
From 164,276 real admit and reject decisions across 995 universities, plus about 47,000 applicant profiles labelled by experts, checked against official admit rates in more than 15 countries. In countries where places are given out by exam rank, it is the chance the programme's next cut off lands at or beyond your rank.
What if I think one of your numbers is wrong?
Click the Source link next to it and read the university's own page. If the university says something different from us, tell us and we re check that university's information. The source link exists so you never have to argue with us about a figure.
Do you only cover famous universities?
We hold 7,187 universities across 50 countries, and the list you get is built from the ones that teach your subject at your level in the countries you chose, ranked by your chance of an offer. Fame is not an input. A famous university appears on your list when your odds there are worth an application, and not because it is famous.
Why 14 schools?
Fourteen is what a real mix of reach, target and safety schools needs: six targets in every case, with the reach and safety counts moving with how ambitious you want to be. Anything below a 10% chance is left out, so every one of the 14 is a school where an application fee buys you a genuine chance.
Should I stop using ChatGPT, Claude or Gemini completely?
No, and we would rather be straight with you about that. A general assistant is good at drafting, rewriting and explaining, and plenty of students use one to practise interview answers or tidy a statement of purpose. What this page shows is that it is the wrong tool for deciding where to apply, because it changes its answer every time you ask and it is pessimistic about your chances in a way that costs you money.
How often is the university information updated?
At the moment we update university information periodically, three to four times a year. Every value is stored with the date it was last read and a link to the page it came from. We re check a university by searching for the page a value lives on rather than crawling the whole site again, which is what makes keeping information current affordable rather than a once in a while clean up. If you find something out of date, the Source link shows you exactly where it came from and we will refresh that university.
Does a high probability mean I will get in?
No. A probability is a probability, not a promise, and we would rather say that plainly than sell you a guarantee nobody can give. What a well calibrated number does is tell you how to spend your applications: how many reaches are worth trying, where your real targets are, and which school is genuinely your safety.
See your own numbers, not someone else's famous names
Build a shortlist on your real grades, your subject and your chosen countries. It takes a few minutes, and you can check every figure it gives you.