The Brief

New briefs daily around 7 AM Eastern

Gemini 4 Carbon opens to all paid API users at $2/$10 until December 1

On October 29 only selected users and vetted cyber defenders could use Carbon, which led the AA index by one point over Opus 5.5. Until Dec 1 it costs half of Opus 5.5's $4/$20 rate, then the price doubles to match it.

• 01Read

  MKA benchmark: Fable 5.2 leads at 57%, Carbon 54% for a fifth the cost

  Fable 5.2 averages $31.97 per task against Carbon's $6.29 for a lead of a few points, and GPT-6.1 Sol costs $1.11. Within ten points of the top score, choosing a model for visual document work is mostly a decision about cost.

  Artificial Analysis articles→blog.google→
• 02Read

  Git 2.56.1 patches recursive-clone code execution and five other bugs

  Coding agents routinely clone repositories they did not write. A crafted repository could run attacker code during a recursive clone, so any agent host or CI runner on an unpatched Git, including copies bundled inside other tools, ran that risk.

  Elijah Newren→
• 03Read

  DeepMind's AlphaProof Code proves Signal protocol code secure in 4 weeks

  Protocol proofs usually stop at the design, while real bugs hide in the code. Agents wrote over six million lines of Lean proving properties of Signal's compiled code, leaving humans to review about 1% of it, work DeepMind says took experts years.

  deepmind.google→
• 04Read

  Shopify rebuilds its 300-screen app in Swift and Kotlin in 12 weeks

  Agents wrote most of the code, but the speed came from the system around them. Screens became testable from the terminal in milliseconds instead of minutes per emulator loop, and CI fails when iOS and Android drift apart.

  Shopify Engineering→
• 05Read

  NVIDIA open-sources Snapshot, restoring GPU inference pods in seconds

  Cold starts of large models take minutes, so serving fleets keep idle replicas warm to absorb spikes. Restores of 3.6 to 42.7 seconds, 8x to 36x faster, make scale-from-zero workable even for MoE models with 1 TB checkpoints.

  Ron Kahn→
• 06Read

  Mercor raises $750M at $20B, says inference costs 3x its salaries

  Mercor's annualized revenue went from $1B in February to $3B now. A company spending three times more on inference than on payroll is a concrete datapoint on where AI budgets now sit, from the firm that sells training data to labs.

  swyx .ai→