Hands-on with Gemini 4 Argon: strong on professional benchmarks, weak on agents and coding UX
数字生命卡兹克 · wechat · 2026-10-14
- Google's new flagship Gemini 4 Argon is now fully open, framed by the author as a do-or-die release after a chaotic stretch: Gemini 3.5 Pro was internally cancelled over coding shortfalls, Demis Hassabis moved to Alphabet chief scientist, and Jeff Dean/Sanjay left to start a company.
- Benchmarks: Vals Index 68.9 (#1 on professional work like finance/law/tax, at half the per-run cost of Opus 5.5), but Artificial Analysis ties GPT-6 Astra at 53 while costing 3x per task ($1.99 vs $0.72). CUA-bench computer-use is only 4.83%, second-to-last.
- UX complaints: Antigravity has no language setting, a $100 Ultra quota emptied after fewer than 10 tasks, frontend output quality far behind GPT-6 Astra, and Chinese writing retains Gemini 3's quirky metaphors.
- Verdict: skip Ultra; $20 Pro is enough for similar users. The author also shares his monthly stack: ChatGPT Pro $100, Claude Max $250 (Opus 5.5 still the king), X Ultra $200 for GrokBot and 50x quota.
Related event: Gemini 4 Argon hands-on wave: Google is back in the frontier race(29 posts)→
More from coding & agent
- AI builder shipped 510 chars of JS where a customer had 70,000 — a field report on blaming the model vs the pipeline — AdministrativeBad752 · 2026-10-14
- Reddit debate: What risky actions would you let an AI agent take without human approval? — InvisibleAct_1988 · 2026-10-14
- iFixAi, an AI agent auditing CLI claiming 120-second diagnostics, hits 26k GitHub stars — ifixai-ai · 2026-10-14
- Japanese engineer combines AI CAD kernels with in-house tooling to auto-generate a heat pump — burhop · 2026-10-14
- agent-manager: an open-source MCP server giving coding agents a shared, persistent brain — Aware_Web9715 · 2026-10-14
- Dev chains 28 scraping tools into a router reading 96% of sites blind-tested — Careless_Love_3213 · 2026-10-14