Every aitutors.me tutor runs on Anthropic's Claude. The choice was deliberate — safety training, instruction-following on long system prompts, MCP ecosystem, and the founder's experience using Claude as a daily tool. Here's the technical reasoning for parents who care about what's under the hood.
The decision in one paragraph
When you build a tutor for children, three things matter most: the model's safety training, its ability to follow instructions exactly (so the Socratic protocol is enforceable), and the ecosystem around it (deployment, observability, cost). Claude scored highest on all three when I made the call in late 2025.
Safety: Constitutional AI
Anthropic publishes its safety approach — Constitutional AI — openly. The model is trained against a set of explicit principles (be helpful, harmless, honest), with the principles themselves visible for inspection. This is genuinely unusual.
For a child-facing product, this matters because:
- The training principles are auditable
- Refusal behaviour is predictable
- The model has strong "I can't help with that" defaults on harmful requests
- Anthropic's safety research is transparent in published papers
That said, no frontier model is perfect. aitutors.me adds its own safeguarding layer on top — the Mentor agent scans every conversation for safeguarding indicators independently of the model's built-in safety. See docs/safeguarding.md in the product repo.
Instruction-following: long system prompts
aitutors.me's agents have detailed system prompts. Professor Pi's prompt specifies the 4-level hint ladder, the "never give the answer" invariant, the Show Your Working protocol, and the KS3 misconception library. The prompt is several pages long.
Long system prompt adherence is one of the dimensions where Claude particularly excels. Practical tests during development:
- Claude follows multi-page agent prompts consistently
- Edge cases ("demand the answer") trigger the expected refusal behaviour
- Personality voice stays consistent across long conversations
GPT-4 and Gemini are both competitive on this dimension. Claude, in 2025–26 testing, was marginally more reliable for our specific Socratic constraints.
MCP: the deployment ecosystem
aitutors.me launches as MCP servers (Model Context Protocol) deployed on Anthropic's Managed Agents platform. The architecture:
- Each professor is an MCP server
- Claude Desktop / Claude Code users install them via OAuth
- Managed Agents handles execution, isolation, observability
- Authentication gates per-account access
This architecture was only possible because Anthropic shipped Managed Agents in late 2025. Building the same thing on another model required self-hosting the agent runtime — a major build that would have taken months. Picking Claude let me ship in 3–4 weeks.
The product later moves to a Next.js chat UI (Phase 2), which removes the Claude-subscription dependency for end users. But the agents themselves remain on Claude.
Cost
At MVP volumes, Claude's pricing via Managed Agents lands at roughly $0.08/hour + token costs. For typical 20-minute tutoring sessions, the per-session cost is a few cents. This is sustainable at £14/month subscription pricing with healthy margins.
Other models are competitive on raw cost. Claude wins specifically because the managed deployment removes ops overhead that would otherwise eat the margin.
The Anthropic-Claude relationship
aitutors.me has no special partnership with Anthropic. We're a regular API customer. We pay the same rates as anyone else. There's no preferential access, no co-marketing, no exclusivity.
The "Built on Claude (Anthropic)" footer attribution is what Anthropic's brand guidelines permit for API customers — equivalent to "Powered by Stripe" on a Stripe-billed site.
Why this matters to you as a parent
The model choice affects:
- Safety — what the AI refuses to do
- Quality — how good the tutoring feels
- Privacy — Anthropic's data policies apply (no training on API customer data; you can read Anthropic's policy directly)
- Resilience — when Anthropic ships better models, aitutors.me gets the upgrade
The first two are why the choice was deliberate. The privacy point is important: Anthropic's policy is that API customer data is not used to train models. aitutors.me adds its own no-training, 30-day deletion layer on top.
What happens if Claude has an outage?
Realistic risk. Mitigation:
- aitutors.me is built to swap models if needed (the prompts are LLM-agnostic, though tuned for Claude)
- Phase 2 web app architecture allows fallback to a backup provider
- For MVP, an extended Claude outage would degrade service — we've sized the Founding cohort small enough that a 24-hour outage is recoverable with personal apologies
What I'd say to someone choosing model providers in 2026
Use the model that:
- Has published, auditable safety training
- Follows long instructions reliably
- Has a deployment ecosystem that fits your architecture
- Costs in line with your business model
For aitutors.me's specific Socratic-child-facing constraints, Claude was the answer. For a different product, another model might win.
FAQ
Which AI model powers aitutors.me?
Anthropic's Claude (Opus and Sonnet family). All tutors run on Claude via Anthropic's Managed Agents platform.
Why Claude instead of GPT or Gemini?
Three reasons: Constitutional AI safety training, strong long-system-prompt adherence (needed for our detailed agent protocols), and Managed Agents + MCP ecosystem that enabled the soft-launch architecture.
Is Claude safer than other AI for children?
All frontier models have robust safety training. Claude's approach is published and auditable. None are perfect — aitutors.me adds its own safeguarding layer on top.
Related reading
Written by Jason at aitutors.me. Updated 20 May 2026.