Fable 5.1 ships with 75% cache read cut, Sonnet 5 locked at $2/$10 forever, commerce agents blueprint drops

Anthropic shipped Claude Fable 5.1 on September 1 as a point release on top of Fable 5. The headline isn't the base price, which stays at $10 per million input tokens and $50 per million output tokens — it's the cache-read price, cut 75% from $1.00 to $0.25 per MTok. New cache-write rates are $12.50/MTok for 5-minute and $20/MTok for 1-hour TTL. Anthropic estimates this lowers typical Fable 5.1 spend by roughly 25% and highly agentic spend by up to 45%. On benchmarks, Fable 5.1 leads Fable 5 and Opus 5 on all published rows at launch and more than doubles Fable 5 on Terminal-Bench-Science (52.6 vs 24.7). Context window stays at 1M tokens, max output at 128K, adaptive thinking on by default, knowledge cutoff is June 2026. Model ID: claude-fable-5-1. Anthropic commits to no retirement before September 1, 2027.

Claude Mythos 5.1 is the same underlying model as Fable 5.1 with select safety filters lifted for verified cybersecurity and life-sciences organizations — access runs through Project Glasswing and the Cyber Verification Program via Anthropic, AWS, or Google Cloud account teams. Claude Security codebase scans now run on Mythos 5.1 rather than Mythos 5. Text generated by both Fable 5.1 and Mythos 5.1 carries Anthropic's invisible statistical watermark required under the EU AI Act — detectable only via Anthropic's detection API, currently in private preview for regulators, law enforcement, media, and fact-checkers.

Fable 5.1 introduces three breaking changes vs Fable 5: (1) a new history-editing check blocks tool_result content that edits assistant history — requests that do this now return a 400 error; (2) progress updates between tool calls are now returned as text in a new format when thinking.display is set to 'updates' (requires the thinking-display-updates-2026-08-18 beta header); (3) stricter thinking block validation rejects malformed extended thinking blocks that Fable 5 previously accepted silently. Audit any agent harness that manipulates conversation history or passes custom thinking blocks before migrating.

Anthropic confirmed on August 11 that Claude Sonnet 5's $2/$10 per MTok introductory pricing is now the permanent standard price — the scheduled September 1 rise to $3/$15 will not happen. The decision came twelve days after OpenAI cut GPT-5.6 Luna pricing by 80%, confirming what analysts had called a frontier model price war. Sonnet 5 now undercuts GPT-5.6 Terra on output tokens ($10 vs $12) while matching it on input ($2). For teams that had budgeted a 50% September price jump, the risk is fully off the table. Note: Sonnet 5's tokenizer still maps the same text to 1.0–1.35x more tokens than Sonnet 4.6 — the per-task cost comparison is narrower than the list price alone suggests.

Anthropic published its Commerce Agents Blueprint — a production-ready reference architecture and open-source implementation for companies building agents that serve shoppers, travelers, subscribers, and merchants. Built in collaboration with Visa, the blueprint covers trust and consent flows, transaction confirmation gates, scope controls, and abuse-prevention patterns. Fork the reference implementation at github.com/anthropics/commerce-agents. An engineering deep-dive is at claude.com/blog/the-anatomy-of-effective-commerce-agents, and vertical demos are available at claude.com/solutions/commerce.

The Claude Developer Platform's Managed Agents received a significant upgrade this week: budget controls let you set per-session and per-agent token spend limits with automatic cutoffs. Advisor support adds an optional review agent that checks an agent's plan before it executes — a lightweight human-in-the-loop gate. Geo-pinned inference lets enterprise customers lock Managed Agent sessions to a specific AWS or GCP region for data-residency compliance. GitHub-loaded skills let agents pull skill definitions directly from a repo branch, removing the need to upload them through the API separately.

The thinking.display parameter now accepts a third value — 'updates' — in beta (requires the thinking-display-updates-2026-08-18 header). Reasoning comes back with an empty thinking field (as under 'omitted'), but short progress updates that Fable 5.1 and Mythos 5.1 write between tool calls are returned as text, at most one thinking block before each tool call. Useful for surfacing model reasoning state in long agentic runs without streaming full extended thinking. The anthropic-workspace-id response header also shipped this week, returning the wrkspc_-prefixed workspace ID on every API response.

Claude for Teachers expanded to include a free Enterprise offering for K-12 schools and districts. School IT admins can now provision centrally managed access with single sign-on, role-based controls, and admin dashboards for all verified educators in their organization. The tier includes the Learning Commons connector and teaching skills library at no cost. Previously, institutional access was a paid Enterprise add-on; the move to free brings Claude for Teachers in line with Anthropic's commitment to equitable AI access in education.