Gemini 4 Neon lands as fourth Gemini 4 in seven weeks; Anthropic IPO priced at $83 a share
Latent Space · rss · 2026-11-19
Latent Space's AI News digest for Nov 17–18, 2026:
Gemini 4 Neon
- Google DeepMind shipped its fourth Gemini 4 model in 7 weeks; launch gaps shrank from 28 days (Argon→Carbon) to 15, then 6 days (Hydrogen→Neon); Argon retired after 1.5 months
- Priced at $3/$0.15 cached/$15 per M tokens, 1M context, four effort levels; Vals reports up to 1M-token outputs
- At high effort it matches Carbon's 60 on the Intelligence Index for $3.74/task; debuted second on Agent Arena (1,336 pts, $2.82 median task cost); CursorBench 4.1: 63.7% vs Hydrogen's 57.3%
- Early hands-on: 762 tool calls, 93% input-cache hit rate, but 57 capacity errors logged
GPT-6.2 results
- Astra and Sol take the top two MathArena spots; Astra solved 42/53 ArXivLean problems; Vals reports first Lean formalization of Smoktunowicz's Gap Theorem, kernel-checked
- OpenAI opens public beta for background Responses API jobs with 30%/40% discounts for 2h/8h targets
Open weights & inference
- Tencent releases Hy4 Flash: 224B multimodal MoE, 12B active, 1M context, 64B n-gram embedding table
- vLLM 0.34.0: extensible KV cache on by default, up to 1.95× faster NVFP4 decoding on GB200
- Epoch: chip capacity for agent workloads grew 4.6× in a year vs 2.6× peak compute
- Runway halves serving hardware for a 57B video model to 24 fps on four GPUs
- HF corrected Pi adapter numbers and shipped an OpenCode adapter lifting Qwen3.8-27B from 35.3% to 51.0% on the hardest SWE-bench Pro subset
Industry
- Anthropic IPO priced at $83/share, 980M shares, $81.3B, Nasdaq debut Nov 19 under ANT
- Palo Alto Networks to buy Applied Compute for $4.5B in stock
- Google's Genie 4 world model reaches Project Genie with native audio; Kling Image 4.0 ranks seventh in text-to-image
More from Companies & People
- Google Photos & AI Subscriptions VP on 17 Years Inside Big Tech — laparisa · 2026-11-19
- Google's recent coding model surge credited to Mechanize acquihire — himanshustwts · 2026-11-19
- OpenAI pretraining researcher: 'no confidence' humans keep better research taste than models for long — ChrisGPT · 2026-11-19
- Google ships a model every two weeks and nothing happens, argues dev — gabriel1 · 2026-11-19
- India's Sarvam AI denies using subsidised GPUs to host Chinese open models — HimanshiET · 2026-11-19
- OpenAdvisor launches as 'Yelp for PhD advisors' with anonymous reviews and rebuttals — Pseudomanifold · 2026-11-19