Human review for AI memory updates

AI-proposed Memory updates can now wait for your review.

New models

Claude Haiku 5.5 is now in qbash, with a 1M-token context window and adaptive thinking. It’s an option for frequent steps such as classifying incoming requests or extracting fields from documents.

Prompts up to 100K tokens use the lower rate. Above that threshold, input and output rates rise 5×. Check your typical prompt length before switching an existing workflow to Haiku 5.5.

Haiku 5.5 follows Opus 5.5, Sonnet 5.5, the GPT-6 family, and Jev, all added since September.

Browse models →

Memory upgrades

Memory folders can now require human approval for AI writes. Proposed changes become pull requests, with a diff to review before accepting them.

For example, a research Task might update a competitor’s Markdown profile with newly announced pricing. You can review the proposed edits before they’re merged into the file. Until you approve them, agents, Tasks, and people using that profile continue reading the previous version.

Read the docs →

New and updated connectors

MySQL

MySQL joins the database connectors. It supports MySQL 5.7+ and MariaDB, including hosted services such as RDS and Aurora. Tasks can inspect table structures and query the records they need.

Google Ads

Keyword Planner is part of the new Google Ads connector. Tasks can research keywords using historical search metrics or forecast campaign performance. The connector also supports GAQL reports and campaign management. New campaigns default to paused, and dry runs check proposed changes using Google’s validation mode.

You can now access more ActiveCampaign automation and CRM data. Affinity adds tools for opportunity notes and interactions, plus relationship scores.

New from the blog

Who owns the token bill?

Mitchell Tessier

Mitchell looks at who should manage an agent’s growing costs and decide when a recurring workflow should become a fixed Task.

Who the hell is Jev, and why is everyone talking about decisions?

Alex MacLeod

Alex explains how decision models handle classification and works through the cost of processing a million emails a month. He also examines what a probability score can tell you about an answer.

The AI Ticker, October 4 to 8

Where Claude sends your data, and how AI-written research gets checked. Five stories.

GPT-6’s Chat launch has users reporting odd safety blocks, shorter answers, and inaccurate graphics. A question about plague mechanics in Europa Universalis 5 reportedly triggered a biological-safety flag. Another user described a two-minute safety review over an article about a Siberian lab.

The Association for Human Mathematics urges mathematicians to stop working with OpenAI after its mass release of AI-generated research. Roughly 42% of the 719 manuscripts’ main results are formalized in Lean. OpenAI cites consultations with an advisory group. AHM says OpenAI ignored that group’s objection to testing advanced problems on internal models.

Anthropic removed Pro and Max users’ local-only Cowork option on October 6. New tasks run in its cloud, which also stores session history. Connected folders stay local, but Claude fetches copies of files it needs. For local execution, Anthropic points users to Claude Code.

IMDEA’s tests found chat links and email addresses still reached Intercom after users rejected cookies on Claude. A separate server-side route to eleven trackers, including Meta and TikTok, depended on cookie consent. Researchers did not show conversation text reaching ad networks. The measurements date from May, so current behavior may differ.

Pavel Rabtsevich used Claude Code and Codex to investigate a planet candidate. He describes over 1,000 scripts and a test against withheld telescope data. After an audit, he corrected an overstated statistical claim. TESS has approved follow-up observations. The planet remains unconfirmed.

What are you building with Tasks this week? Reply with the process you’re working on or the part that’s giving you trouble.