On July 14, 2026, Anthropic released Claude for Teachers. It had been roughly eight months since OpenAI launched ChatGPT for Teachers on November 19, 2025. Both products serve verified U.S. K–12 educators, neither provides student accounts, and both promise that inputs and outputs within the workspace are not used for model training by default.
On the surface, this looks like two companies giving teachers a free chat box, one after the other. The real difference lies in where each product begins. OpenAI built a workspace that schools can centrally manage — what counts as a good lesson plan is left largely for teachers and districts to define. Anthropic baked curriculum standards and lesson-planning rules into the workflow first; identity, permissions, and administrative features that districts need will come in later versions.
These two paths happen to address the two necessary conditions for teacher AI: whether schools dare let teachers use it, and whether teachers can reduce ineffective rework once they do. But even if you combine the two products, the step education cares about most is still missing: once materials enter the classroom, the system has no way of knowing whether students actually learned.
Three months ago, in “An AI Tutor for Every Student — Is That Really Where AI in Education Lands?”, I argued that AI’s greater leverage in education may lie on the teacher and curriculum design side, rather than giving every student a virtual tutor. That both companies are now giving accounts to adult teachers first suggests this entry point is becoming a product consensus. But they also fill in two barriers I didn’t fully develop back then: how a good materials workflow enters a school’s governance structure, and how classroom results feed back into the next round of teaching.
Here, “school district” does not refer to the enrollment zoning familiar to Chinese readers, but to a layer of local governance in American public K–12 education. A school district typically manages multiple elementary, middle, and high schools, with a district office responsible for budgeting, procurement, account systems, data privacy, and technical standards. Principals and teachers carry out instruction within individual schools. Arrangements vary by state, but whether a tool moves from individual teacher adoption to formal deployment often depends not just on the school, but on the district.
For a teacher to have AI revise a piece of material, the first question is which data can be uploaded. If the material contains student names, test scores, or special education information — who can access it, whether the school can centrally revoke accounts, what happens to materials after a teacher leaves — none of these can be answered with “data is not used for training” alone. A district’s technology and privacy team needs to manage identity, permissions, data, and collaboration scope before it can uniformly approve a tool.
Even with an approved account, lesson planning doesn’t automatically become easier. A general-purpose chatbot doesn’t know where fifth-grade math is in a given state’s curriculum, doesn’t know which sub-skills sit beneath a standard, and doesn’t know how to support students at different paces toward the same learning objective. The teacher has to find this context themselves and then check every generated result item by item. If the curriculum requirements have to be explained anew each time, a fast first draft can still end up costing time in revision and rework.
OpenAI and Anthropic each chose to lower one of these costs first. Viewed through this lens, it becomes easier to understand why the two products adopted different features, partnerships, and distribution approaches.
ChatGPT for Teachers allows teachers, school staff, school-level leaders, and district administrators to verify their identity and create workspaces independently, while also giving districts a path to bring scattered accounts back under management.
According to the official guide, administrators can claim their district’s email domain — domain claiming — bringing accounts registered with those email addresses into the same managed workspace. SAML SSO lets teachers sign in with school credentials, RBAC determines which features each role can access, and analytics lets administrators see usage across the workspace. Individual teachers can start using the product first, and districts still have a way to pull these accounts into a uniform rule set later.
Once inside the workspace, shared projects let colleagues co-maintain course files and project context, and Custom GPTs can save templates and rules commonly used within the school. OpenAI later added Workspace Agents to the Teachers plan, letting teachers offload repetitive workflows to a background Agent. It remains in research preview; long-running task success rate, human handoff, and error recovery have no published data and should not be treated as mature school automation.
These features answer who can use it, where collaboration happens, and how administrators manage it. They do not automatically answer what counts as good material. ChatGPT for Teachers does not ship with a dedicated 50-state standards graph and public teaching rubrics comparable to Claude’s; teachers and districts must inject local curriculum rules through files, Custom GPTs, or Agents. A strong instructional team can encode its own practices; a school lacking that experience may end up reusing, at scale, materials that appear complete but deviate from learning objectives.
Claude for Teachers tackles the other cost first. It connects to Learning Commons, allowing it to query the curriculum standards of all 50 U.S. states, the sub-skills beneath each standard, and the typical sequence in which those skills are learned. Curricular resources such as OpenSciEd and IM v.360 also fall within this query scope. Teachers don’t have to rely entirely on model memory to guess standards and prerequisites, though they still need to verify the query and generation results.
Once standards are retrieved, Claude organizes tasks through built-in Skills. Anthropic has open-sourced the K–12 Teacher Skills repository; only two are currently built in as official: lesson planning, for preparing a lesson from scratch, and lesson differentiation, for adapting existing materials. These are more than a prompt — they include reference materials, execution steps, guardrails, documentation scripts, and rubrics.
Lesson differentiation includes a concrete requirement: when adding or reducing support for students at different levels of progress, the core objective and cognitive demand must not be lowered. Rules like these can catch common material errors and also give outsiders a way to inspect what Anthropic means by “instructional alignment.” However, the rubric still inspects materials, not outcomes, and there are only two official Skills. It has not proven that teachers who adopt these materials will produce better classroom results.
Claude Cowork extends tasks to multi-file and cyclical work, but it is likewise in research preview. The more immediate limitation is school governance. The current individual teacher version does not offer domain claiming, SAML SSO, member management, or RBAC; school and district-specific versions are still under development. That a teacher can sign up does not mean the district has authorized them to upload student data; a DPA governs how the service provider handles data but cannot substitute for internal school authorization, auditing, and centralized revocation. Claude’s highest-value scenarios often require richer classroom context, and the product’s most conspicuous gap right now is precisely the district control plane for managing that context.
Partnership rosters easily conflate product coverage, actual usage, and educational outcomes. Arranging the evidence in sequence makes things clearer: first comes a launch announcement, then school deployment and training, then teachers actually using the product consistently. Further up, you’d need to measure whether teachers objectively save time, whether materials improve, and finally, whether student learning outcomes change.
At launch, OpenAI said its initial 16 partner organizations represented nearly 150,000 teachers and staff. The operative word is “represented.” 150,000 is the personnel scope of those organizations — not the number of activated accounts, active users, or training completions. Implementation progress also varies among the initial partners: some have established managed workspaces, while others’ public information still stops at the partnership announcement.
U-46 provides one of the rare publicly reported usage figures. According to a Chicago Tribune report, 756 staff members were using ChatGPT for Teachers at the time. 47% of users responded to a survey; among respondents, more than 80% used it at least several times per week, and over 70% self-reported saving 1–5 hours per week.
These numbers are enough to show the product crossed the threshold from “eligible to use” to early actual adoption in one district. They are not yet retention or outcome data. The survey captured responses from fewer than half of users, with active users more likely to respond; time savings came from recall-based self-reports; and whether students learned better as a result was not addressed by the report.
Claude for Teachers has just launched. What can currently be confirmed is that the product is available, standards queries work, and two Skills can be inspected. Prospect Schools has provided early feedback; Detroit Public Schools Community District has announced future evaluation. The Iceland pilot, the Rwanda partnership, and the Teach For All network demonstrate that Anthropic is engaging real education systems, but an opportunity pool, training, and partnerships do not translate into active users of this product.
Broader field research also needs to be placed in its proper context. Early causal studies summarized by Stanford SCALE suggest that teacher-facing AI tools can reduce lesson-planning time while maintaining material quality. The evidence remains thin, and it did not study ChatGPT for Teachers or Claude for Teachers specifically, nor has it shown that student learning improved as a result.
So ChatGPT for Teachers is more mature in institutional deployment and has early usage data from a small number of districts. Claude for Teachers has gone deeper on instructional materials workflow, while actual deployment, retention, and outcome evidence is just beginning. There is no single score to declare a winner here, because the two products are mature in different dimensions.
The most visible advantages each company holds today could be caught up to. Learning Commons has opened parts of its knowledge graph, evaluators, and Skills as public infrastructure, so other platforms can adopt similar standards queries and material rules. Anthropic can ship a district version later, but a feature release does not mean schools are ready to adopt — identity integration, data permissions, support, and renewals will each need to be verified in turn.
The harder stretch happens after materials leave the AI. Once a teacher prints out a lesson plan or imports it into a learning system, ChatGPT and Claude typically cannot see what happens next: which step students get stuck on, what the teacher changes on the fly, whether students remember it on the next test. A model can check whether a lesson plan has tagged the right standards, but it cannot tell from the plan itself whether students learned.
Connecting classroom data back is also not as simple as adding a connector. Student exercises, teacher modifications, and long-term performance involve new privacy authorizations, data quality issues, and attribution problems, and they also add to the burden of teacher documentation and school implementation. For a teacher-facing product to genuinely influence learning, it needs instructional rules to enter a reusable workflow, that workflow to enter school governance, and classroom feedback to be captured at a tolerable cost. OpenAI and Anthropic have each made one of the first two more complete; neither has yet offered product proof for the third.
This also points to the more practical choice today. If a district’s most urgent need is unified identity, permissions, and collaboration management, ChatGPT for Teachers is more complete. If a professional teacher’s biggest aim is to reduce the rework of standards lookup, learning objective decomposition, and material differentiation, Claude for Teachers offers more tangible gains. Both choices still require teachers to shoulder the final instructional judgment.
The genuinely informative metrics going forward won’t be another batch of partnership rosters, but how many teachers are still using the product after 90 or 180 days, how much time was actually saved after deducting verification and rework, what failures or overreach the Agents experienced, how materials performed once they entered the classroom, and how much students retained after AI support was withdrawn. The teacher interface has become the new entry point for AI in education. What remains to be verified is whether this entry point can lead all the way to learning.