This edition covers the Fable 5.1 launch, a new page for your full execution history, connector upgrades across Airtable, GitHub, Gmail, and Drive, and simpler memory access inside Task scripts. Two new posts are on the blog, including the one explaining our new name. This week's AI Ticker is at the end.
Claude Fable 5.1 is available
Anthropic's latest model is live in the catalog. It ships with a 1M token context window and 128K output, priced at $10 per million input tokens and $50 per million output tokens. Cache reads cost 75% less than they did on Fable 5, which cuts the bill for Tasks that re-send the same context on every run. Early reports say 5.1 also dials down the verbosity users saw across the 5.0 line. Fable and Mythos 5.1 ship with Anthropic's first confirmed statistical text watermark, aimed at slop, or at least at slop turning up in training data.
Fable 5.1 requires org opt-in. If your org already had Fable 5 enabled, you are opted in for 5.1 automatically. Fable models carry a 30-day data retention period and are not eligible for Zero Data Retention. The original write-up covers why we ship these models switched off.
All Runs: your full execution history
There is a new page at /build/runs with every task and prompt execution in one table you can page through, search, and sort. Filter by type (Tasks or Prompts), by trigger (scheduled or manual), and by status.
Each row now shows the trigger source, the trigger type, and the entrypoint. Failed runs carry a compact error summary in the table, so you can scan a column instead of opening runs one at a time to find the broken one. The Build home page links to it under "View all runs."
Connector improvements
Four changes:
- Airtable search accepts a
viewparameter, so you can search within a specific view instead of the whole table. - GitHub PR and review endpoints support sorting and pagination.
- Gmail search accepts
queryas an alias forq. - Google Drive search accepts
queryas an alias forname_contains.
The two aliases exist because query is what most people write on the first try. Existing calls keep working.
Simpler task scripting
Task scripts can now call qbash.memories.get() and qbash.memories.search() without a folder= argument to read the Markdown files bundled with the Task itself.
The AI task builder picked up matching tools, so it can read those inline documents when it helps you write a script.
Two new posts
Aisle is now qbash. AI made judgment programmable and created a new class of application. qbash is the platform where those systems are built, run, and distributed.
RAG in one folder and one Task. Alex turns on Searchable by AI for a folder of AI governance PDFs, exposes vector search through a single Task, and asks Claude a question that takes several searches to answer.
The AI Ticker, August 30 to September 2
Five stories from the past four days. We kept the ones that touch what you build on, or what it costs to run.
Claude quality regressed over the 7 to 10 days before the 5.1 release, according to a 121-comment thread. Users report models rushing to output instead of asking clarifying questions, across Claude Chat, Code, Fable, and Opus 5.
Claude Max API accounts default to a $200,000 organization spend limit with notifications off. Multiple users confirmed it independently and Anthropic has not responded. If you have your own key, set a cap in account settings.
Fable 5.1's text watermark is live, but the public detection tool does not detect it. claude.com/check-content only reads C2PA credentials in media files. Text detection remains in private preview for eligible organizations.
A Marmelab benchmark of Claude Code cost plugins found a prompt-cache trap. Ponytail cut generated code by about half. Headroom's context compression backfired: rewriting history broke the cache prefix and forced expensive cache writes that outweighed the read savings.
The New Stack published an agent harness checklist: typed tool contracts, idempotency keys, a scoped service identity per tool, and a trace of what the agent saw when something broke. On testing, it says to score hard assertions and rubric-graded prose separately, because agents are nondeterministic.