Gemini 4 Argon goes GA at $2/$10: matches GPT-6 Astra but tops 50% false-claim rate
Latent Space · rss · 2026-10-14
Latent Space's weekly AI News covers the AI Engineer NYC conference (20,000 attendees; keynotes from Anthropic, Vals AI, Coatue) and a week of frontier releases.
Gemini 4 Argon GA
- Global rollout for coding, knowledge work, enterprise workflows and cyber defense via the Gemini API/app/AI Mode/Antigravity. Intro pricing $2/$10 per million tokens (cached input 95% off), doubling to $4/$20 after November 30.
- Google claims Argon separated core webpage handling from a 6M-line legacy Chromium folder in weeks and cut rollbacks by 1/3 with 35% more agentic code submissions (internal figures).
- DeepMind demoed Argon deriving a cornhole simulation from human video and landing 7 consecutive throws on Apptronik's Apollo robot.
- Cursor reports 52.8% on CursorBench, on par with Opus 5.5; Cognition reports 60.5% on FrontierCode.
Evaluations: strong capability, uneven reliability
- Artificial Analysis: Argon ties GPT-6 Astra on its Intelligence Index at 60% of the cost ($1.99/task), with a 15% AA-Omniscience hallucination rate—lowest among 45+ scorers.
- Arena ranks Argon High fifth behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra Max.
- Roboflow: 84.7% text localization vs Astra's 76.8%, 4x cheaper, 19.6s vs 29.7s latency.
- But Theo found a >50% false-claim rate on his codebase-analysis benchmark despite a perfect Skatebench-private score; Ethan Mollick flags weak multi-agent execution and overconfidence—construction ability ≠ trustworthy analysis.
Meta Muse Wonder
- Positioned beyond coding as a personal assistant (Active Focus, Active Memory); Alexandr Wang calls consumer use the primary use case.
- Expensive, restrictive quotas: users report burning the $200 plan's five-hour allowance in hours or the weekly $200 in a day; Meta acknowledges rollout issues. MemoryBench only tests within-conversation fact updates, not durable cross-session memory.
Also: Artificial Analysis launched a Decision Index—decision models (Jev 1.13 at 58) are competitive at classification/routing but weak and overconfident on knowledge/reasoning; Haiku 5.5 Low scores 61 but costs $184 vs $42 per million tasks. ScaleDown launched classify/extract/summarize endpoints at $0.05 per million input tokens, claiming 200x speed and 467x cost advantages.
More from Companies & People
- Anthropic Pays £145K for a London Human Assistant Despite Its Own AI Agents — AlexTensor · 2026-10-14
- PostHog to devtools: stop asking 'how to add AI,' ask 'what's the agent-first version' — edwin · 2026-10-14
- Jeff Ladish backs METR's OpenAI eval report, urges OpenAI to release full logs — JeffLadish · 2026-10-14
- NVIDIA joins Earthshot Prize as founding partner, offering finalists free AI compute and DGX Spark — nvidia · 2026-10-14
- Mathematicians say rumors OpenAI's Navier-Stokes proof stole two mathematicians' work may be overblown — analisereal · 2026-10-14
- US AI firms hire only 33 DevRel roles but forward-deployed engineers quadruple to 136 in a year — FinanceYF5 · 2026-10-14