Gemini 4 Carbon tops AA index at 59, gated for cyber defenders, at $2-$10/M tokens
Latent Space · rss · 2026-10-29
Google DeepMind released Gemini 4 Carbon two weeks after Argon, targeting cybersecurity and long-horizon tasks. It tops Artificial Analysis' Intelligence Index v5 at 59 (Opus 5.5 previously 58).
Model & access
- 1M context and 1M max output; text/image/video/speech input
- First available to trusted cyber defenders via the Fairwind Program; wider developer access TBD
- Pricing: $2/$10 per M tokens per Phil Schmid; AA lists $4/$20 with an open-ended 50% launch discount; cached input $0.20
Key scores
- Cyber: 88.2% SEC-bench Pro (vuln finding), 54.8% ExploitGym, 71.0% CWE-bench (patching)
- Coding: 59 on AA Coding Agent Index (Antigravity CLI) vs 62 for Claude Code + Opus 5.5; FrontierSWE 61.2%, behind GPT-6 Astra (65.5)
- Knowledge work: 1609 Elo on AA-Briefcase (+126 vs Argon), still behind Grok 4.7 et al. (1678)
- Cost: $7.76/task vs $10.04 for Opus, using 29% fewer output tokens
- Hallucination rate rose from 4% to 11%, still second-lowest; #1 on Text Arena (1528) and Vending-Bench 2
Other notes
- Pichai expects "continued acceleration because of RSI" over the next 12 months; an internal engineer says launches now feel like "periodic checkpoints"
- scaling01: Google 2 months behind the next OpenAI/Anthropic releases; China–US gap 3.3 months
- Benchmark integrity: AA v5 runs 60% private tests; a mystery "Obsidian 3" (46.8%, above Opus 5.5) appears in Cognition data, unconfirmed
- Safety research: DeepSeek V4.1 Flash-1017 shows a 46.4% false-positive rate on eval detection; a red-heart image makes Claude Opus 4.7 and GPT-5.4 write recursive code 100% of the time with 74.7% wrong explanations; Anthropic's GRAFT confines reward-hack updates to post-training-deleted parameters; Qwen3.5 has an embedded lie detector at 91% hit rate with zero false positives
- Unconfirmed: The Information reports Kimi K3.1 Pro was distilled from GPT-6 Astra; "European sovereign AI" Giotto criticized as a Kimi rebrand
More from Models
- Claude Duet Appears to Be Rolling Out to Many More Users — imjustnewatai · 2026-10-29
- Qwen3.5-4B hits ~11.4 tok/s CPU-only on a Ryzen 5 laptop with llama.cpp tuning — solyarisoftware · 2026-10-29
- A video 'made by Opus 5.5' surfaces as Anthropic leak chatter grows — repligate · 2026-10-29
- US models lead China's on use-case post-training and even frontier benchmarks — teortaxesTex · 2026-10-29
- Terra tops Artificial Analysis over Kimi and Fable, sparking benchmark trust debate — teortaxesTex · 2026-10-29
- GPT-Image 2.6 spotted in the wild as OpenAI readies next-gen image model — koltregaskes · 2026-10-29