As we wrote in What Is Personalized Education, and Why Does It Matter?, the core of personalized education is judging where a student is in real time and adjusting how you teach accordingly. Making that happen is hard.
There are three barriers:
- Missing input: AI needs to see and hear how the student is working to understand where they need help.
- Slow decisions: AI needs to decide when and how to help quickly enough to keep up with the student.
- Slow responses: Even with the right decision, the help has to arrive quickly enough to keep up with the student.
Mathos has always focused on personalized education. Today, we have overcome these three barriers. We believe the future where quality, affordable, and personalized learning experiences are accessible to all people at any time has arrived.
What a Good One-on-One Teacher Can See During a Lesson
What stands out about one-on-one teaching? How efficiently you can learn.
Why? Imagine you have just started a lesson with a math tutor. The topic is quadratic equations. The tutor gives you a basic problem:
If you are unfamiliar with quadratic equations, you read the problem and think, without putting pen to paper. But if you know the concept well, you immediately know how to solve it and start working.
One-on-one teaching is efficient because the tutor can see how long you spend reading, how your expression changes, and that you still haven't written the first step. They can pick up these signals when you need help and decide what to do next.
How Mathos Solves This Problem
On June 5, 2024, less than a month after GPT-4o launched, we introduced our Tutor and whiteboard experience. As one of the earliest companies to offer this experience, we have always believed in the importance of a whiteboard in teaching.
In the latest version of Math Tutor, students can write out their steps and talk with the tutor as they work. The tutor can decide what to teach next based on their tone of voice, the knowledge required for the current step, and what it already knows about them—for example, whether they often forget this particular concept.
We can also distinguish scratch work from a submitted answer and assess it in real time.
Students can think, write, and take part in solving the problem. The tutor helps them as they progress, so they can keep up and keep working, with less frustration and exhaustion from listening to explanations they don't understand.
What the Statistics Tell Us
During Mathos's preflight phase, users had more than 100K conversations with our Tutor. Our internal statistics show that 70% of students were willing to ask for help when they encountered difficulty, and users completed a problem with the Tutor's help in 90% of conversations. For comparison, a two-year study in 18 Tennessee middle schools found that the median student messaged Khanmigo in only 17% of exercise sessions involving a mistake. [1]
Mathos Tutor
Khanmigo [1]
What Comes Next
In Mathos Tutor V3, launching in late October 2026, we will also be able to observe whether students are distracted or still online. Alongside assessing their written work in real time, we can analyze their facial expressions, voice, and whiteboard work together and step in when they struggle, bringing the experience closer to learning with a human tutor.
How a Good One-on-One Teacher Makes Decisions During a Lesson
As a one-on-one teacher, I have many ways to teach a student. Suppose I want them to solve a problem, but they can't write the first step. I can ask: “Do you know the concept this problem uses?” If they get stuck halfway through, I might think: are they struggling with the same thing as last time? Then I can help them review that concept. Once they finish, I can give them a similar problem to see whether they can solve it on their own. If they already know it, we can skip the problem altogether.
For AI, the hard part is making these choices promptly: where is the student now, and how should we teach them next? Those decisions have to be fast enough to keep up with their work.
How Mathos Solves This Problem
Long before models such as Jev appeared, Mathos started using our own framework to assess students' solution steps. A conventional LLM can take 3–4 seconds to make a judgment. In the version we launched in March 2026, we were already checking whether each step was correct in an average of 800 milliseconds. We have now reduced that to 400 milliseconds—twice the speed.
Time to assess a solution step
For reference, in an adult arithmetic experiment, participants with stronger mathematical skills took an average of 563–628 milliseconds to judge whether simple addition and subtraction answers were correct. [2]
This lets us assess almost in real time which step a student has reached and whether it is correct, then choose how to help: give a hint, show a video, review a concept, ask them to rewrite a step, or give them another problem.
What the Learning Outcomes Tell Us
Our internal statistics show that after using Mathos Tutor, more than 95% of students can independently solve a similar problem.
For reference, a randomized trial in high-school mathematics published in PNAS found that students using a standard GPT-4 assistant scored 48% higher on practice than the control group. But once AI access was removed, their independent exam scores were 17% lower than the control group's. [3]
How a Good One-on-One Teacher Sets the Pace
When we talk with someone, they usually respond in less than one second. A common gap between turns is only around 200 milliseconds. [4]
That's because we are already understanding and thinking about our response while we listen. A one-on-one teacher does the same thing: while watching a student work, they are already judging the situation and thinking about what to do next.
What the Research Tells Us
But AI often waits until you finish a sentence before processing it, or until you finish a step before assessing it. In StudentBench's study of 12 AI tutors, the time for a model to return a complete response ranged from 1.9 to 31 seconds. Slower replies were associated with less student participation in the conversation. [5]
After all that waiting, the student may get a long explanation. They are still trying to understand the beginning while the tutor has already moved on. A study analyzing 16,851 student–AI interactions also found that longer responses were associated with lower understanding in the student's next interaction. [6]
How Mathos Solves This Problem
We have always dreamed of giving students an almost real-time interactive experience. In September 2026, we finally reduced the average response time in student conversations from 1,410 to 798 milliseconds.
Average conversation response time
Our architecture also lets us generate dynamic UI during the lesson.
If a student wants to see a function graph, we draw it. If they want to change a parameter, we can update the graph while continuing to explain the problem.
For comparison, ChatGPT's Intelligent UI, released by OpenAI on October 7, 2026, takes 5–6 seconds to draw a function graph. Our average graph generation time is under 800 milliseconds, making generation approximately 6.25–7.5 times as fast. [7]
A User Story
A Silicon Valley parent, an executive at a publicly traded technology company, spent several weekends trying to build a better math tutor for his middle-school daughter. He wanted interactive graphics she could explore herself, but ChatGPT didn't produce the experience he wanted, and Desmos required manual work. Then he found Mathos and contacted us to say this was the experience he had been trying to give his daughter.
We were also among the earliest companies to provide students with visual tutorial videos. We launched video generation on March 7, 2025. With our latest Harness technology, we can now generate the first frame in under 5 seconds and a 5-minute video in under 30 seconds. Students can watch and talk with the tutor at the same time.
We have also kept working on overlapping visuals and misplaced elements in videos. Our in-house math animation rendering engine has reduced visual overlaps or errors by 80%.
With these three barriers addressed, we’re launching Mathos Tutor.
One More Thing
We believe the future of education should not be limited by the supply of great teachers.
When Claude Code launched in February 2025, programmers could hand an agent a task: define the goal, give instructions, and review the results. The agent could inspect the project, write code, run tests, and keep adjusting until the task was complete.
Programmers could now delegate the coding.
But in education, teachers still do most of the teaching. AI helps prepare materials, while the teacher observes the student, identifies the difficulty, and decides how to respond.
We believe tutoring is ready for the same change.
A teacher could tell Mathos Tutor: “He remembers the quadratic formula but cannot use it yet. Work through a few problems with him and see where he gets stuck.”
A learning goal, carried through.
Set the learning goal
“He remembers the quadratic formula but cannot use it yet.”
Teach, observe, and adjust
See what to work on next
The Tutor would follow the student’s work, explain when needed, and let him try again. When the teacher returns, they could see what he has learned and where he still needs help. Parents could arrange practice this way too.
This is AI tutoring’s Claude Code moment: teachers and parents set learning goals, and tutoring agents follow through on the teaching.
It brings us closer to what Mathos has always wanted to do: make quality, affordable, personalized learning available to everyone, whenever they need it. Great teachers could reach more students and have more time to understand them, encourage them, and help them find their own direction.
AI can scale teaching. Great tutoring can change lives.
Sources
- [1]One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment
- [2]The neural differences of arithmetic verification performance depend on math skill
- [3]Generative AI without guardrails can harm learning: Evidence from high school mathematics
- [4]Universals and cultural variation in turn-taking in conversation
- [5]StudentBench: AI and human tutoring yield equivalent GRE learning gains
- [6]Micro-level AI Feedback Features and Student Responses in Consecutive LLM Tutoring Interactions
- [7]OpenAI: Intelligent UI announcement
