Claude Code 2.1.211 Adds Subagent Streaming and Locks Down Permission Previews
One day after 2.1.210, Claude Code shipped 2.1.211 — a broad stability and workflow update. The marquee addition is opt-in subagent text streaming: a new --forward-subagent-text flag and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT env var include a subagent’s text and thinking in stream-json output, so orchestration tooling can watch nested agents work in real time.
Security got sharper too. Permission previews relayed to chat channels now neutralize bidirectional-override, zero-width, and look-alike quote characters, so a tool input can’t visually rewrite the approval message you’re about to click. A PreToolUse hook’s ask decision now floors auto mode instead of being overridden for unsandboxed Bash. Rounding it out: faster terminal rendering, a fix for parallel sessions all logging out after wake-from-sleep, and repairs across Chrome uploads, Vertex, Bedrock, and background-session recovery.
API Keys Can Finally Expire — With Reminders and Admin Reporting
A small but overdue governance win: the Claude Developer Platform now supports API key expiration settings in the Console. When you create an API key or Admin API key, you can set a lifetime — a preset, a custom duration, or Never — instead of minting a credential that lives forever.
For keys with a lifetime of at least seven days, Anthropic emails the creator before expiration, and admins get reporting on which keys are approaching their end date. It’s the kind of quiet plumbing that turns a security best practice (rotate your keys) into something the platform actually nudges you toward.
New agent-memory Beta Header Stabilizes Memory Listing
Builders working with Claude’s agent memory get a firmer foundation. A new agent-memory-2026-07-22 beta header makes memory listing return a stable, server-defined order and tightens the behavior of depth, path_prefix, and cursor parameters — the difference between paginating a memory store reliably and getting shuffled results.
The SDKs have been updated to send the new memory-store header by default, so most developers pick up the stricter, more predictable semantics without changing their code. It’s a maturation step for the agent-memory primitive that underpins long-running Claude agents.
Anthropic Extends Free Fable 5 Access to July 19 — Again
Anthropic has stretched the promotional window for Claude Fable 5 once more, keeping it available to Pro, Max, Team, and select Enterprise subscribers at no extra cost through July 19. It’s the second extension in a week, and reporting frames it as an apparent response to OpenAI’s latest model release — a “keep them from switching” move in an increasingly twitchy top of the market.
After July 19, the economics change: Fable 5 usage runs on prepaid usage credits at $10 per million input tokens and $50 per million output tokens, with Anthropic saying it aims to fold the model back into subscriptions once capacity allows. Fable 5 only returned to general availability on July 1, after US export restrictions on it and Mythos 5 were lifted.
Claude’s Memory Switches to Individual, Categorized Entries
Claude’s consumer memory just got more granular. Memory now works as a set of individual, categorized entries that Claude reads and updates during your conversations, replacing the previous single daily memory summary. In practice that means richer, more precise recall — Claude can pull the one relevant fact instead of leaning on a coarse end-of-day digest.
It pairs with the recently launched monthly recap and focus settings under Settings > Reflect, part of a broader push to make Claude feel like it remembers you across sessions without turning that memory into a black box.
Two Speeds at Anthropic: Frantic at the Frontier, Patient in the Plumbing
Today’s stories split cleanly into two tempos. At the frontier, everything is reactive and fast: Fable 5’s free window keeps getting pushed a few days at a time, each extension timed to a competitor’s launch. That’s not a roadmap — it’s a retention firefight, and the tell is that Anthropic is willing to eat the compute cost of a flagship model to keep paying subscribers from wandering during a hot stretch of the model wars.
Underneath, the platform work moves at a completely different, unhurried pace: expiring API keys, stable memory ordering, invisible-character hardening in Claude Code approvals. None of it trends on X, but it’s the compounding stuff — the boring guarantees that let a security team say yes to Claude in production. The interesting question is which tempo actually decides the enterprise. Flashy model swaps win the week; predictable key rotation and tamper-proof approvals win the three-year contract. Anthropic is clearly betting it needs both — but only one of them shows up in the press releases.