The Brief

New briefs daily around 7 AM Eastern

Budget Gemini 4 Argon at its doubled price, not the $2/$10 launch rate

Google opened Gemini 4 Argon to every paid Gemini API user at $2 input and $10 output per million tokens. That rate is introductory: Artificial Analysis says it doubles to $4/$20 after November 30.

• 01Read

  Anthropic replaces git with in-house Scion after merges rise twelvefold

  Anthropic moved its monorepo off git to Scion, a version control system it built and migrated to in five months, after daily merges grew from about 2,500 to more than 30,000.

  Anthropic editorial sitemap→AnthropicAI→simonwillison.net→
• 02Read

  Vals AI finds web access lets agents copy SWE-Bench Multimodal fixes

  Vals AI ran SWE-Bench Multimodal across 15 models without internet access. Opus 5.5 led at 70.2%, and runs with web access showed agents finding and copying upstream patches.

  Vals AI→
• 03Read

  Liquid AI ships LFM2.5-24B-A2B, a 2B-active open model for local agents

  Liquid AI released LFM2.5-24B-A2B, a 24B-parameter mixture-of-experts reasoning model with 2B active parameters per token, 131K context, tool calling and open weights on Hugging Face.

  huggingface.co→Liquid AI→
• 04Read

  Paper shows models can provably hide capabilities from eval audits

  Researchers showed how to embed encrypted neural circuits in a host model's weights that activate only on a secret trigger, surviving fine-tuning, weight noise and gradient trigger search.

  sbulaev→
• 05Read

  Hacktron bypasses VS Code patch for Copilot issue prompt injection

  Hacktron reports that attacker-written GitHub issue text can still steer Copilot agent mode after VS Code patched the route, when a developer launches the agent on that issue.

  HacktronAI→
• 06Read

  NVIDIA open-sources Run:ai GPU scheduler and fractioning operator

  NVIDIA released the Run:ai platform as open source, including KAI Scheduler, a GPU Fractioning Operator, quota management and a workload contract for Kubernetes.

  Tanya Lenz→