Google admits AI Mode hurts Search revenue; Anthropic Ultra drops 5-hour cap; Gemini 4 Carbon tops arenas

Latent Space · rss · 2026-10-31

Latent Space's AI News roundup covers a heavy week:

Google earnings shock

Google admitted strong AI Mode query growth comes with lower monetization (CBO Philipp Schindler), triggering a 15% drop and $500B market-cap loss, its worst since the 2004 IPO; backdrop: OpenAI raising at $1.25T instead of $1.4T.

Anthropic reworks Claude tiers

Max 5x ($100/mo) becomes "Max", Max 20x ($200/mo) becomes "Ultra"; verified Ultra accounts lose the 5-hour session limit (new accounts keep a temporary anti-fraud cap). Weekly limits unchanged.
Unverified analysis: Anthropic API margin 92% vs DeepSeek 84%; subscriptions bring 18% of revenue but use >50% of compute.

Releases

Gemini 4 Carbon: #1 Text Arena, #2 Agent Arena, SOTA on Perplexity's WANDR benchmark.
Genie 4 world model: steerable interactive simulations with sound, research preview.
Mistral Large 4: 1T params, 1M context, day-0 LoRA fine-tuning on one 8xB200 node.
NVIDIA Nemotron 3.5 Super: 120B/12B active, 51.78% SWE-bench Pro, up to 4,856 output tok/s/GPU on GB200.
Microsoft's 4B computer-use hybrid agent scores 66.0% on OSWorld-Verified, beating 27B/30B agents; Phonic-1 (10-person team, $4M) claims top speech benchmarks.

Evals and quant disputes

AA Intelligence Index v5: Opus 5.5 (max) 58, GPT-6 Astra 57, GPT-6.1 Sol 55 at 1/8 the cost; Opus low-effort scores just 35.
Agent Arena (26K sessions): without prompt caching top-agent tasks cost 3.6x more ($7.44 vs $27 for Fable 5.1).
Unsloth ships DeepSeek V4.1-Flash dynamic GGUFs (64GB+ RAM) and says it can't reproduce GSQ-RCO's 100% 1-bit AIME25 claim (got 91.88%).
Leaked DeepSeek-Cyber scores (92% CyberGym) are unofficial; analysts argue they fit a guardrail-free 1017 build, not a new model.

More from Companies & People