SHELFCRITTER AI Digest

MODEL

Gemini 4 Carbon

modeltopic-notegoogle

Overview

Gemini 4 Carbon is Google DeepMind‘s 2026-10-28 frontier release, landing roughly two weeks after Gemini 4 Argon‘s worldwide rollout. The headline positioning is autonomous engineering and cybersecurity: initial access is gated to vetted security experts for US government pre-release testing, with wider availability staged behind that. Per The Decoder’s write-up (the primary blog.google post was not independently reachable this run), Carbon posts an Artificial Analysis Intelligence Index v5 score of 59 vs Argon’s 50, still trails Claude Opus 5.5 on agentic coding, leads 15 of 16 of Google’s own benchmarks, cuts the Gray Swan indirect-prompt-injection attack-success rate from 0.7% to 0.0%, and holds pricing at $2/M in · $10/M out.

Timeline

  • 2026-10-29-AI-Digest — Google DeepMind launched Gemini 4 Carbon on 2026-10-28, ~two weeks after Gemini 4 Argon went GA, with an autonomous-engineering + prompt-injection positioning. Initial rollout gated to vetted security experts for US government pre-release testing. Per The Decoder (primary blog.google post not independently reachable): AAII v5 59 vs Argon 50, still trails Claude Opus 5.5 on agentic coding, cuts Gray Swan attack success 0.7% → 0.0%, pricing holds at $2/$10 per Mtok. Load-bearing softener: the specific benchmark numbers trace to a single secondary outlet; carry as The Decoder’s write-up rather than verified first-order claims. Scope clarification: Argon→Carbon in two weeks reads fast on paper, but Google has been iterating Gemini 4 variants on Arena since September — the first-order claim worth carrying is Carbon ships with the autonomous-engineering + prompt-injection angle, not the frontier release cadence is accelerating.

Key Developments

  1. Autonomous-Engineering + Prompt-Injection Angle on a US-Gov-Gated Early Release (October 28, 2026): Carbon launches positioned explicitly for autonomous software engineering and cybersecurity, with initial access gated to vetted security experts for US government pre-release testing. The Gray Swan indirect-prompt-injection reduction from 0.7% to 0.0% (per The Decoder) is the headline safety number carrying the positioning; the gating pattern echoes Argon’s Oct 1 Fairwind-gated cyber-defender rollout, extending the same distribution shape into a second Gemini 4 variant inside one month.

  2. AAII v5 59 vs Argon’s 50 but Still Trails Claude Opus 5.5 on Agentic Coding (October 28, 2026): Per The Decoder, Carbon’s 59 on the Artificial Analysis Intelligence Index v5 is a nine-point jump over Argon’s 50, and the model leads 15 of 16 of Google’s own benchmarks — but Google concedes it still trails Claude Opus 5.5 on agentic coding. Load-bearing framing to carry: Carbon moves Google’s index-tier floor up while leaving the Opus 5.5 agentic-coding ceiling intact — not a frontier-ceiling flip, a frontier-floor shift. The specific numbers trace to a single secondary outlet in this run and should be carried with that caveat until the primary blog.google post is independently reached.

See also: Google, DeepMind, Gemini 4 Argon, Claude Opus 5.5, MOC - Major Companies, MOC - Agent Security.