Anthropic just gave every US teacher a Claude Code seat

Good morning 👋 Anthropic just solved the actual bottleneck in AI-for-schools sales, and it isn't the free Claude Code seat.

In today's issue:

• 🔭 Anthropic gives every verified US K-12 teacher a free Claude Code seat, wrapped in the legal paperwork that actually gets it approved
• 🔬 Agents that don't know a task is small, judges that grade easy with no answer key, and 44 models that can't stop picking the same word
• 🛠️ Steal the lesson-plan-and-grading agent setup Anthropic just gave every teacher for free
• 🔗 Four worth a sentence: AI's real power ceiling, a pay-for-work pricing pitch, and two healthcare builds going to production

🔭 THE ONE THING

🎓 Anthropic just gave every US teacher a Claude Code seat

Anthropic is opening Claude for Teachers to any verified K-12 educator in the US, free, with full Claude Code and Cowork bundled in rather than some cut-down classroom mode. Sign up by June 30, 2027 and the year of access is locked in. Nine connectors ship on day one, MagicSchool, Canva Education, Brisk Teaching and Diffit among them, alongside a FERPA-compliant K-12 Data Processing Addendum built to get past district IT. The DPA is the real story here, not the Claude Code seat: procurement is the actual moat in education sales, and Anthropic just handed every district's lawyer a reason to say yes before Google or OpenAI make the same offer. Whether "standards-aligned across all 50 states" via Learning Commons survives an actual Tuesday lesson plan is a different question, worth checking back on in a semester.

🔬 RESEARCH HIGHLIGHTS

• Agents don't know when a task is small. A new paper catches LLM agents defaulting to max-context-first: re-reading every file and dependency they've already seen, even for a one-line edit. The fix isn't a bigger context window, it's a cheaper first question the agent never asks: how much effort does this actually need. Do AI Agents Know When a Task Is Simple?

• LLM judges grade easy when there's no answer key. Turns out judge models skew generous specifically in no-reference setups, which is most real evaluation pipelines (open-ended output, no ground truth to check against). If your eval numbers look suspiciously good, this is a plausible reason why. LLM Judges Can Be Too Generous

• Ask 44 models to pick any word, 41% say "serendipity." Funny, but it's a real signal on model homogeneity. Post-training keeps compressing "creative" output down to the same handful of safe choices across labs. Remember that next time a demo calls an LLM's output original. The One-Word Census

🛠️ TRY THIS

Turn Claude into a lesson-plan-and-grading agent (steal the setup Anthropic just gave every K-12 teacher for free)

1. Spin up a Claude Project and drop in your source material: a syllabus, a rubric, a style guide, whatever the repeatable task actually references.

2. Write the output format into custom instructions once (objective, activity, exit ticket) instead of re-specifying it in every prompt.

3. Run the same prompt against five different inputs and check where it drifts from the rubric.

4. Keep the prompt that survived all five. That's the reusable asset, not any single output.

Prompt: Using the attached rubric, draft a 45-minute lesson plan on [topic] with an objective, one guided activity, one independent activity, and a 3-question exit ticket. Flag any part of the rubric you couldn't address.

Worth a look

• Amazon Nova Act QA Studio AWS extended its agentic QA tool with batch regression test suites that parallelize execution, plus a CLI that wires agentic testing straight into CI/CD pipelines. AWS blog post

🔗 QUICK LINKS

• NVIDIA says the real ceiling on AI factory revenue isn't chip count, it's performance per watt: tokens generated per fixed power budget, a metric you can't fake your way around. link
• OpenAI wants enterprises to stop counting agent seats and start measuring useful work per dollar. Sensible, and also a hint at how they'll eventually price agentic usage. link
• ScienceSoft built a HIPAA-compliant AI voice scheduler for healthcare on Nova 2 Sonic plus Bedrock Guardrails. link
• Flo Health took its medical content review system from AWS proof-of-concept to production on Bedrock. Worth reading if you're stuck in that same PoC purgatory. link

More tomorrow.

Pradeep Perugu