Give your AI just the passages it needs

Send your AI the evidence, not the whole document

Cohere’s new Rerank 5 models can return relevant passages within documents, not just rank the documents themselves. The optional Context Focus feature gives retrieval apps a direct way to control how much text reaches the model.

• One API call returns document scores and scored text spans. Configurable thresholds let you trade evidence retention against context size, with no re-indexing required.
• In Cohere’s search-agent evaluation, the full retrieval stack used 41% fewer tokens across six deep-search benchmarks than keyword search alone, while accuracy rose from 39% to 52%. Those are vendor-reported results, not a forecast for your app.
• Pro and Fast are generally available. Cohere lists API pricing at $0.05 per million tokens for Pro and $0.02 for Fast.

What it means For document-heavy RAG apps, compare passage filtering against your current context selection on known questions, checking whether shorter inputs preserve the evidence your answers need.

Source: cohere.com

AI helped find PHP flaws. Now find every copy.

OpenAI reported on October 21 that its models helped researchers discover critical flaws in PHP, which it estimates runs 70% of websites. PHP has released fixes, making this both a significant example of AI-assisted security discovery and a concrete patching task.

• OpenAI says it spent the preceding two months coordinating maintainers, researchers, and security providers to prepare fixes and protections before disclosure. It says frontier AI found the vulnerability ahead of attackers, to its knowledge, giving defenders time to prepare.
• Sourcegraph’s remediation guide identifies two flaws allowing remote code execution without authentication. It lists patched upstream releases 8.5.12, 8.4.27, 8.3.36, and 8.2.35, and calls for upgrading runtimes and restarting workers; PHP 8.1 and earlier no longer receive upstream support.
• The [Sourcegraph guide](https://sourcegraph.com/blog/php2shell-patching-php-is-easy-finding-it-is-the-hard-part) starts with Deep Search: list Dockerfiles, Helm charts, and CI workflows that install or pin PHP, including each version, to build an inventory linked to source files.
• Its rollout method uses scripts for mechanical version bumps and Claude Code or Codex for edits needing judgment. Start with a canary, respond to CI failures, and track pull requests until they are mergeable.

What it means Find and fix the repository pins as well as deployed runtimes, so your next build does not quietly restore the vulnerable interpreter.

Source: OpenAI

Claude Duet keeps working after you close the app

Anthropic opened a research-preview waitlist for Claude Duet, a proactive agent running on its own cloud computer. It can prepare morning briefs, draft invoice reminders, and produce files such as research briefs or slide decks while you’re away.

• Access is limited to US Pro and Max users, with invitations arriving in waves. The preview costs no extra, but background work counts against your existing plan limits.
• Duet carries over your Claude memory, skills, and connected apps. It is designed to request approval before sending messages, making payments, signing documents, or deleting anything.
• It checks accounts at set intervals, not continuously. You cannot watch or take over its browser, and it cannot access files on your computer; Microsoft 365 is not available yet.

What it means Consider the waitlist for recurring briefings and draft preparation, but not urgent monitoring: scheduled checks and shared usage limits constrain how much background work you can delegate.

Source: support.claude.com

Flint brings tool-using coding to a local model

Recursive released Flint’s weights under Apache 2.0, giving builders another coding-agent model they can run on their own hardware. It supports reasoning and multi-turn tool use for investigating codebases, fixing bugs, and writing tests.

• Flint has 39.8 billion total parameters, with approximately 2.5 billion active per token. Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts.
• The model card documents OpenAI-compatible tool calls and instructs developers to use its chat template and preserve returned reasoning_content in conversation history.
• Recursive warns that results vary with checkpoint precision and runtime, and that context capacity does not guarantee consistent accuracy at every length. Tool-using apps still need their own permissions and execution boundaries.

What it means For local coding agents, test Flint on a contained bug fix using your intended runtime; its published laptop speed is a short-context result, not an end-to-end task guarantee.

Source: RichardSocher

Elsewhere in AI

• OpenAI models are coming to Google Cloud , including frontier models and Codex workflows in Model Garden, a distribution expansion onto a competing lab’s cloud platform; the announcement does not establish immediate general availability.

AI workflow of the day

How to make a printable coloring book

Shared by God of Prompt (@godofprompt) · View original post ↗

Creators making a seasonal digital product can use this sequence to plan original scenes, generate coloring pages, and draft a listing. It provides a production method, not evidence that the book will sell.