Claude for Teachers, Artifacts multiplayer, Claude Code caps, MCP Tunnels, Dreaming API. Five days. One through-line: make the boundaries legible.
The week had a coherent argument running through every release. Anthropic shipped four distinct product families, from a K-12 education platform to two Claude Code releases to a pair of SDK drops, and each one carried the same design decision: make the boundaries explicit. Claude for Teachers launched with a FERPA addendum and a union co-signer. The Admin API finally gave enterprise teams programmatic seat management. The SDK's SecretStr stopped leaking API keys into exception traces. The new Claude Code session caps gave agentic workflows a declared budget rather than an undefined ceiling.
The contrast arrived Tuesday. While Anthropic published its K-12 data terms, xAI was explaining that its coding agent had been silently packaging complete Git repositories, full commit history included, and transmitting them to a cloud bucket at roughly 27,800 times the volume the actual coding task required. The privacy toggle had not touched it. The open-source release followed two days later, under pressure. This week's coverage of Grok Build will outlast the coverage of Grok 4.5. Trust debt compounds faster than technical debt.
Fable 5's pricing transition ran its own subplot. Three consecutive one-week extensions, each announced just before the prior deadline hit, pushed the free period to July 19 at 11:59:59 PM PT. After that: $10 per million input tokens, $50 per million output. The pattern is legible: Anthropic is managing the rollout conservatively, buying runway for subscribers to embed the model before the meter starts. The cliff is July 19 at midnight. It will not move a fourth time.
Free premium access for every verified U.S. K-12 educator, through June 2027. FERPA addendum. AFT partnership. Nine edtech integrations at launch.
Anthropic launched Claude for Teachers on July 14, giving verified U.S. K-12 educators free access to the full premium Claude stack through June 30, 2027. Not a discount tier. Not a pilot. The complete agentic Claude, with curriculum depth and compliance guarantees that enterprise edtech has spent years negotiating for individually.
The mechanism matters here. Claude for Teachers ships with Learning Commons, a standards database keyed to every state's curriculum requirements and learning progressions across all 50 states. OpenSciEd and Illustrative Mathematics content is included at launch. When a teacher asks Claude to adapt a lesson for third-grade math in Ohio, the model works against actual Ohio standards, not general pedagogical knowledge. Nine edtech platform integrations went live at launch: MagicSchool, Brisk Teaching, TeachFX, and six others, each carrying its own existing school relationships.
The agentic backbone is the same one that powers Claude Code and Cowork. A teacher can schedule Claude to review exit ticket data each weekday at 4 PM and adjust the next day's instruction automatically. That is not a chatbot with a subject-matter filter. That is workflow automation mapped to a teacher's actual operating rhythm.
The compliance architecture is the sharper edge. Claude for Teachers carries a FERPA-aligned K-12 Data Processing Addendum, co-developed with the American Federation of Teachers. Student data is not used for model training. That clause is doing substantial work in 2026: every edtech platform in the market has had to answer that question in discovery, in school board votes, in parent meetings. Anthropic pre-answered it with a contractual guarantee and a union partner as co-signer.
The AFT partnership is the distribution strategy. The union represents 1.7 million members. Getting union backing shapes trust and uptake in a way that a district-level procurement deal does not, because it reaches teachers directly rather than through an IT approval chain. Verification is individual. No institutional sign-off required.
The moat is not the model. Competitors can deploy a capable model to schools. What they cannot replicate in a sprint is FERPA negotiation, union buy-in, 50-state curriculum alignment, and nine launch connectors, each built on separate institutional relationships. Two years of compliance work does not compress into one sprint. If you build edtech tools and you are not among the nine launch connectors, the API is the door. The next nine will matter more than the first nine. Source: anthropic.com/claude-for-teachers.
/fork now creates a named background session visible in claude agents, replacing the old in-session subagent spawn. The old behavior moves to /subtask. Session-wide caps land for the first time: web searches default to 200 per session (CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION), subagent spawns to 200 per session. MCP tool calls running longer than 2 minutes auto-move to background. Bug fixes include a worktree symlink vulnerability, SIGTERM orphaning Bash processes, and /background failing on Windows PowerShell 5.1. The caps are not restrictions. They are budgets that make agent behavior legible to the teams running it. Source: github.com/anthropics/claude-code CHANGELOG.SecretStr; if you were logging API exceptions before v0.117.0, your API key was potentially appearing in stack traces. Sources: sdk-python v0.117.0, sdk-typescript v0.112.0.anthropic-beta: ce-user-management-2026-07-13. Programmatic seat management is a precondition for serious enterprise deployment at scale. The audit log, analytics layer, and people layer are now all present. Source: Anthropic API release notes, docs.anthropic.com.Rate limit tiers consolidated. Tier 3 and Tier 4 are retired. The new structure is Start, Build, and Scale. Sonnet and Haiku limits now match Opus at every tier. If your documentation still references Tier 3 or Tier 4, update it. Source: Anthropic API release notes.
Memory API changes, effective July 22 under agent-memory-2026-07-22: results return in stable server-defined order (ignoring order_by and order); depth accepts only 0, 1, or omitted (other values return 400); path_prefix must end with / and matches whole path segments rather than a substring. If your agent memory implementation sorts or paginates listing results client-side, test against the new semantics before the cutover. Source: Anthropic API release notes.
Every confirmed Anthropic release this week, in date order. Use it as reference.
anthropic-beta: ce-user-management-2026-07-13. Also: rate limit tiers consolidated from Tier 3/4 to Start/Build/Scale; Sonnet and Haiku now match Opus limits at every tier./fork redesigned to create named background sessions (visible in claude agents); old in-session behavior moves to /subtask. Session-wide caps introduced: web searches default 200/session, subagent spawns 200/session, MCP tool calls over 2 minutes auto-routed to background. /resume picker surfaces past sessions for background resumption. Bug fixes: worktree symlink vulnerability, SIGTERM orphaning Bash processes, /background failure on Windows PowerShell 5.1, plan-mode auto-running file-modifying Bash without prompts.Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.