Whether parents or teachers, nearly everyone in the past two years has faced this scene: a student takes a photo of a homework problem with their phone, feeds it directly to ChatGPT or Doubao, and gets the solution and answer instantly—easy and enjoyable. AI companies have clearly long recognized the potential of the education market, so even amid intense competition, they are carving out energy to enter this space. However, unlike Chinese AI companies that go directly after students and parents, American AI companies first need to win over teachers.
On July 14, Anthropic released Claude for Teachers, offering free access to the Pro level of Claude to all K‑12 teachers in the United States. The news itself was not explosive—after all, OpenAI launched ChatGPT for Teachers back in November last year, and Google had already stuffed Gemini into its education‑focused Workspace much earlier. Over the past year or more, almost every major large‑language‑model company has been pushing products into education. OpenAI gives free access to teachers and college students; Google leverages its Workspace ecosystem to directly cover school districts; Khan Academy’s Khanmigo is funded by Microsoft, allowing teachers across the U.S. to use it at zero cost.
Why are top AI companies using the same “free” strategy that the internet once used to capture campuses? What is the real calculation behind this land‑grab? Below, we might find some answers by looking at Anthropic’s approach.
Anthropic’s “Infrastructure” Play
The most noteworthy aspect of Claude for Teachers is not the “free” offer, but the technical architecture behind it. Anthropic has integrated a curriculum knowledge graph called Learning Commons. This system, developed by the Chan Zuckerberg Initiative, covers academic standards for all 50 U.S. states—and goes beyond the standards themselves. It includes detailed sub‑skills under each standard, as well as the typical learning sequences for students. This means that when a fifth‑grade math teacher in Texas asks Claude to help with lesson planning, Claude is not guessing what to teach using general knowledge; instead, it follows the Texas curriculum system to pinpoint exactly what to teach at that grade level, that semester, and for that specific knowledge point. It can also tap into the Open SciEd science curriculum and Illustrative Mathematics’ IM v.360 math curriculum, generating lesson plans directly based on the textbooks already in use at the school.
Even more interesting is Anthropic’s choice of product form. Claude for Teachers is not just a chat window. It integrates Claude Code and Cowork, allowing teachers to set up automated tasks—for example, automatically grading that day’s classroom quizzes at 4 p.m. every day and then adjusting the next day’s teaching plan based on the results. This is no longer a “Q&A‑style AI”; it is a teaching assistant that can run continuously.
Anthropic has also done something almost no one in the AI education space has done: it open‑sourced teaching skills. The two core skills it developed jointly with Learning Commons—curriculum planning and curriculum differentiation—along with the code and evaluation framework, have all been placed on GitHub. Any developer or educational technology company can take them and use them in their own AI products, not limited to Claude. This move means far more than the product itself. Anthropic is not building an educational product; it is building the infrastructure for educational AI. It has positioned itself as the “standard‑setter”—you all come and use my framework, and I will define what “good AI teaching” looks like.
The Land‑Grab for the “Next Generation” of Users
Looking back over the past year and a half, the pace at which AI giants have entered education has been accelerating. In April 2025, Anthropic launched Claude for Education, targeting universities and partnering with institutions like Northeastern University and the London School of Economics, featuring a function called “Learning Mode” that guides students to think for themselves rather than just giving answers. That same month, OpenAI announced free access to ChatGPT Plus for North American college students. In July 2025, OpenAI, together with the American Federation of Teachers, established the National Academy for AI Instruction, aiming to train 400,000 teachers within five years. Anthropic and Microsoft also provided financial support for this project. In October 2025, Anthropic formed a Higher Education Advisory Council, chaired by former Yale president Rick Levin.
By 2026, the front line had extended from universities to K‑12. OpenAI’s ChatGPT for Teachers is now extended through 2028; Google has deeply integrated Gemini into Classroom and Docs; and Anthropic’s Claude for Teachers launched in July, along with a commitment of CAD 10 million for Canadian education research and the introduction of localized pricing in India. Taken together, these moves reveal a very clear logic: these AI companies are not doing charity—they are replicating the strategy Google used a decade ago with Chromebooks to capture American schools. Around 2012, Google pushed Chromebooks to schools at extremely low prices, bundled with free Google Classroom and Google Docs. By 2020, over 60% of devices in U.S. K‑12 schools were Chromebooks, and an entire generation of students grew up using the Google ecosystem. When they graduated and entered the workforce, Google Workspace required almost no promotion. Although Chromebooks have never performed well in the mainstream consumer market, they did provide a pool of potential users for Google’s services and software.
Today’s AI companies are doing exactly the same thing. Students use Claude to write papers, lab reports, and prepare for interviews; after graduation, the first thing they do at their new jobs is renew their Claude subscriptions. Teachers who have used ChatGPT for three years to plan lessons, grade assignments, and write comments will find the switching cost to another platform ever higher. Free is not a cost—it is customer acquisition.
Anthropic’s head of education, Drew Bent, put it in measured terms, but his meaning is clear: Anthropic’s goal is “to work with universities to explore AI‑empowered teaching models.” Translated, this means making AI a part of the teaching process, not a cheating tool that students secretly use. This also explains why the companies all emphasize “not training models on teacher data” and “complying with FERPA student privacy regulations”—they need to alleviate school administrators’ concerns before they can truly make it onto procurement lists. Anthropic has prepared a dedicated K‑12 data processing agreement for Claude for Teachers and has launched a formal pilot evaluation in Detroit public schools, tracking teacher experiences and instructional outcomes in partnership with the Gates Foundation.
The Difficulties of AI in Education
Turning our gaze back to China, an interesting contrast emerges. China is not short of AI education initiatives. In January 2026, ByteDance pushed “Doubao Aixue” to the core entry point of the Doubao app; Alibaba’s Qianwen launched a “one‑click search for exam papers” feature; and TAL Education’s Jiuzhang large model, at the World Artificial Intelligence Conference in July, showcased both the teacher‑side “Jiuzhang Lobster” and the student‑side “Little Jinglong.” According to Quest Mobile data, by the third quarter of 2025, domestic AI education apps had surpassed 120 million monthly active users. But on closer inspection, Chinese and American AI education are taking two almost entirely different paths.
The American companies’ approach is to “empower teachers.” Anthropic builds curriculum knowledge graphs, OpenAI sets up teacher training academies, and Google integrates teaching workflows—their common feature is that they do not touch the content itself; instead, they provide tools for teachers, leaving it up to them to decide how to use AI. This is a platform logic, with the core being to become the infrastructure of the education system.
The Chinese companies’ approach is to “reach students directly.” TAL Education leverages its two‑decade accumulation of question banks to build a vertical large model; Doubao relies on free traffic to create a general‑purpose entry point; Zuoyebang and Yuanfudao each have their own AI problem‑solving products. Their common feature is centering on “doing exercises,” with the core being paid conversion. TAL goes even further—self‑developed learning devices, self‑developed large models, and self‑developed content form a closed ecosystem.
Both paths have their own logic. The U.S. education system is decentralized, with different standards in each state and significant teaching autonomy for teachers, so “empowering teachers” is a reasonable entry point. China’s education system revolves around unified examinations, and the core need for students and parents is score improvement, so “helping students solve problems” is a more direct business logic. But a reality that is hard to ignore is that when all “AI education products” are doing problem‑solving tools, the one that students actually use on a large scale turns out to be general‑purpose large models like Doubao. Not because Doubao has built great educational features, but because it is free, sufficient, and can answer anything. Dedicated AI education products, by contrast, fall into an awkward position—they are not as fast as general‑purpose models for solving problems, nor as deep as real teachers for instruction.
American companies have chosen to start from the teacher side and build open infrastructure, while Chinese companies have chosen to start from the student side and build closed paid products. Whose path will work may not be known until this generation of “AI natives” grows up. But at least one thing is already clear: AI’s entry into education is not a technological problem—it is a question of power over who defines teaching and learning. Where the answers differ, entirely different forms of education will emerge.
