Our analysis
The research behind our thesis, and how today’s learning products plan for the people who use them.
What the research says
The strongest results from AI teaching so far come from systems where people designed the teaching method around the model.
| Finding | Source |
|---|---|
| An AI tutor built around expert teaching methods produced more than twice the learning gains of an active-learning physics class, in less time. | Kestin et al., Harvard, 2025 |
| High school students who practiced with a plain GPT-4 tutor scored 17% lower once it was taken away. A version with detailed prompts, written by hand for each problem, avoided that drop. | Bastani et al., PNAS, 2025 |
| Students used Khanmigo on about a third of the days they practiced on Khan Academy, even though nearly all of them had access. | Oreopoulos and Low, 2026 |
| Across randomized trials, human tutoring raises results by about 0.29 to 0.37 standard deviations. That is the honest bar for any AI system. | Nickow, Oreopoulos and Quan; updated review of 89 trials |
| Real-time AI suggestions raised student mastery by 4 percentage points for human tutors, and by up to 9 for the least experienced tutors. | Tutor CoPilot, Stanford, 2024 |
Good results today depend on teaching methods that people write by hand, usually for one subject at a time. That cannot reach every subject and every person. It points to systems that write a method for each learner and keep improving how they write it.
What today’s products adapt
We read each product’s public materials and asked which parts of learning it changes for the person.
| Kind of product | The explanation | Pace and order | The plan | A strategy for each person | Improves across learners |
|---|---|---|---|---|---|
| Courses and textbooksMost classes, online courses, legacy prep | Not found in public materials | Not found in public materials | Not found in public materials | Not found in public materials | Not found in public materials |
| Adaptive practiceKhan Academy, IXL, Math Academy | Partly | Core to the product | Partly | Not found in public materials | Not found in public materials |
| Consumer learning appsDuolingo | Partly | Core to the product | Partly | Not found in public materials | Partly |
| AI assistants in learning modeChatGPT Study Mode, Gemini | Core to the product | Partly | Not found in public materials | Not found in public materials | Not found in public materials |
| AI tutors for tests and schoolAristotle, Acely, Crimsora, OnePrep | Core to the product | Core to the product | Partly | Partly | Not found in public materials |
| AI-native schoolsAlpha School | Partly | Core to the product | Core to the product | Not found in public materials | Not found in public materials |
| Orca21 LabsLearn 0.1a today, then Pod and Learn 1 | Live in Learn 0.1a | Live in Learn 0.1a | Live in Learn 0.1a | Live in Learn 0.1a | Being built |
Core to the product Partly Not found in public materials Live in Learn 0.1a Being built
What the market tells us
People pay for AI that teaches
Aristotle raised a $5M seed in September 2026. Acely charges up to $149 a month. OnePrep reports more than 450,000 students. Alpha School charges $40,000 or more a year for AI-led learning with human guides, and is growing to about 50 campuses.
Sessions are getting cheaper
Voice and whiteboard tutoring now sells for about $30 a month. Google, the College Board and OpenAI give away practice tests and step-by-step help. As sessions get cheaper, the value moves to results, and to planning that produces them.
Planning for each person is still open
We looked at about 20 products, 19 YC companies and 7 a16z Speedrun companies. We found none that builds a plan for each person, has a person check each change, and learns across learners which plans work.
Why this is hard to copy
A system like this needs a clear signal of whether a change helped. Tutoring toward a test gives a measured result on a known date, which is why we start there. It needs people who review changes at volume. Our tutoring business already does this, and we save every correction. And it needs a harness that can be rewritten and tracked version by version, which products built around a question bank or a single chat do not have.
Each reviewed change makes the next harness better. Pod, which we are building, will carry what works from each learner to the next one.
Product by product
The products families compare with LessonOrca.
Aristotle
$49 a month for 8 sessions, $199 unlimited
A voice-first AI tutor for grades 6 to 12 in 65 subjects. Raised a $5M seed in September 2026.
How it plans
Remembers how each student likes ideas explained and uses earlier sessions to suggest what comes next. The plan comes from the same general model that does the tutoring.
How Orca21 differs
Orca21 builds a plan for each learner that runs to their goal and deadline, rewrites it after every session, and has a person check each change.
Acely
$49 a month annual, $149 month to month
AI SAT and ACT prep (YC W18) with 50 practice tests and staff-written explanations.
How it plans
A diagnostic sets a study schedule of about 20 minutes a day, which adjusts as accuracy changes.
How Orca21 differs
Orca21 changes how each learner is taught, from their work and their words, and not only the practice schedule.
Gemini
Free
Free full-length SAT practice tests in Google Gemini, with feedback and Guided Learning for step-by-step help.
How it plans
Writes a study plan once, from a practice test.
How Orca21 differs
Orca21 keeps rebuilding the plan after every session until test day. A Gemini practice test is a good first input to it.
Crimsora
$29 a month founding rate, $49 regular
A voice-first AI tutoring academy with a whiteboard for the SAT, ACT and AP. Named AI instructors each have a style.
How it plans
Students choose an instructor and follow its lesson plans.
How Orca21 differs
Orca21 chooses and changes the teaching approach for each learner from their results.
Khanmigo
$4 a month for families
A Socratic AI tutor on top of Khan Academy content.
How it plans
Follows Khan Academy’s course and mastery system.
How Orca21 differs
Orca21 decides when to teach, for how long and in which format for each learner, which is where low engagement starts.
OnePrep
Free, with a paid Pro tier
SAT, ACT and AP prep from General Learning (YC F24), which reports 450,000+ students.
How it plans
A planner rebuilds the schedule as accuracy changes.
How Orca21 differs
Orca21 adapts the formats, the agents and the approach, and a person reviews each change.
Miyagi Labs
Free to start
AI tutors for exam prep (YC W25) in five countries, reporting 125,000+ students.
How it plans
A 15-minute diagnostic sets a study path.
How Orca21 differs
Orca21 rewrites the path after every session, so it keeps fitting the learner as they change.
LearnQ (Mia)
From $17 a month annual
Digital SAT prep with an AI tutor that has several teaching modes, and optional paid human coaching.
How it plans
A diagnostic builds a knowledge graph and a study plan.
How Orca21 differs
In Learn 0.1b, one reviewer will oversee 75 to 100 learners and improve the system with each review, where LearnQ’s coaches teach one learner at a time.
Other companies worth knowing
Duolingo
Free, with paid tiers
The largest consumer learning app.
How it plans
Its Birdbrain model learns from every exercise across learners to estimate what each learner knows.
How Orca21 differs
Pod is meant to learn which ways of teaching work for which learners, one level above what learners know.
Alpha School
About $40,000 to $55,000 a year
AI-native private schools growing to about 50 campuses, with human guides for motivation.
How it plans
Students spend about two hours a day on mastery apps such as Math Academy.
How Orca21 differs
Orca21 puts a self-improving plan under human review at about one hundredth of the price.
Math Academy
$49 a month
Fully automated adaptive math from grade 4 to university.
How it plans
A hand-built map of thousands of topics, with its own spaced repetition model.
How Orca21 differs
Experts built that strategy by hand for one subject over years. Orca21 builds and improves a strategy for each learner, in any subject.
ChatGPT Study Mode
Free to $200 a month
A Socratic learning mode on every ChatGPT plan.
How it plans
Answers when the learner asks. Each chat starts over.
How Orca21 differs
Orca21 carries a plan to a goal and a deadline and decides what happens next.
Tau
App, Spain only
An agentic tutor for secondary students in Spain that generates tutor agents for each student.
How it plans
Says it learns from all students’ interactions.
How Orca21 differs
Orca21 measures which changes improve outcomes, and a person reviews each one.
Prices
Learn 0.1b launches at $35 a month. Personal teaching today costs far more: $200 to $400 an hour for a tutor at an elite SAT prep or boutique tutoring company, and $40,000 a year for an AI-native school.
| Price | Products |
|---|---|
| Free | Gemini SAT practice, Bluebook and Khan Academy, ChatGPT Study Mode, OnePrep, Miyagi |
| Under $30 a month | Khanmigo ($4), LearnQ (from $17), Crimsora founding rate ($29) |
| $35 a month | Learn 0.1b |
| $49 a month | Acely annual, Aristotle (8 sessions), Math Academy, Crimsora regular |
| $99 to $199 a month | Acely quarterly and monthly, Aristotle unlimited |
| $300 to $2,000+ | Princeton Review, Kaplan and PrepScholar courses |
| $200 to $400 an hour | Elite SAT prep and boutique tutoring companies |
| $40,000+ a year | Alpha School |
Method and sources
We read each company’s public materials and press as of October 10, 2026. Traction figures are the companies’ own. “Not found” means we did not find it in public materials. Learn 0.1b is launching, so we compare its design with products that have shipped.