The Brief

New briefs daily around 7 AM Eastern

Try Sonnet 5.5 before Opus: same price as Sonnet 5, 30% faster

Anthropic released Claude Sonnet 5.5 at unchanged Sonnet 5 prices, running more than 30% faster with fewer tokens per task. Cursor, Copilot and Devin shipped it the same day, so cheaper coding runs are available now.

• 01Read

  Try Sonnet 5.5 before Opus: same price as Sonnet 5, 30% faster

  Anthropic released Claude Sonnet 5.5 at unchanged Sonnet 5 prices, running more than 30% faster with fewer tokens per task. Cursor, Copilot and Devin shipped it the same day, so cheaper coding runs are available now.

  D2OQZG8l5BI1S06→Claude Sonnet 5.5→GitHub→Cursor→Cognition→
• 02Read

  Vals AI finds top reasoning effort costs up to 20x for small gains

  Vals AI ran GPT-6 Astra and Opus 5.5 at every reasoning level on its new MysteryMechanism benchmark and found accuracy gains that cost up to 20x more.

  Vals AI→[AINews] Opus 5.5 is good at explainer videos→
• 03Read

  Cloudflare ships cf CLI covering its entire API and open-sources Forge

  Cloudflare launched cf, a JSON-first CLI spanning its full API, after agents came to drive 48% of Wrangler usage, and open-sourced Forge, the generator behind it.

  Introducing cf: the agentic CLI for the entire Cloudflare API→Introducing Forge: the open source pipeline for generating SDKs, CLIs, docs, and more→Cloudflare→
• 04Read

  GitHub Security Lab's Taskflow Agent finds 24 Android vulnerabilities

  A GitHub Security Lab researcher used packaged taskflows for the open-source Taskflow Agent to audit Android apps, reporting more than 20 vulnerabilities.

  Kevin Stubbings→
• 05Read

  H Company releases Holo4, an open computer-use model at 61.7% OSWorld

  H Company's open Holo4 operates GUIs, code, MCP and APIs through one interface and scores 61.7% on OSWorld 2.0, with full trajectories released.

  Holo4: powering generalist computer-use agents→
• 06Read

  Git 2.56 adds git add --resolved to stage only fixed conflicts

  Git 2.56 adds git add --resolved, which stages only conflicted paths and refuses files that still contain conflict markers, a safer default for agent merges.

  Elijah Newren→
• 07Read

  Audit finds production GPT-4o-mini SQL judge agreed with humans at 0.04

  A production text-to-SQL audit found its GPT-4o-mini judge agreed with human graders at a Cohen's kappa of 0.04, over-flagging 77% of correct results.

  Haowei Liu, Hsin-Tai Wu, Yi Fang→Salma Roshdy Aly, Hussein Assaf, Ziad Kobti→