Gemini 4 Argon

Original Article ↗ Hacker News Discussion ↗

Google’s release of its new Gemini 4 Argon frontier model is being met with a mix of technical praise and product frustration. Early users report that Argon is fast, strong at reasoning, writing, spatial tasks, and multimodal work (like SVG and image-to-HTML), but often behind Claude Opus and OpenAI’s best models for large-scale coding agents, with some issues like reasoning loops and hallucinations. Many are more critical of Google’s fragmented product lineup, confusing rollout, paywalling in Antigravity, and inconsistent access by region or plan than of the model itself, arguing that poor UX and pricing strategy are limiting what could otherwise be a highly competitive offering.

Access & Rollout

  • Major confusion over who can use Gemini 4 Argon:
    • Official copy says “Pro and Ultra” but Antigravity often shows “Ultra only.”
    • Many Pro users (esp. EU/Workspace/Google One) report no access, 404s, or “agent execution terminated” errors.
    • Rollout is clearly batched; some users get it hours later without explanation.
  • Antigravity vs Gemini app vs AI Studio vs API:
    • Pattern reported: Ultra gets Argon everywhere; Pro gets it in Gemini app / web, not in Antigravity (coding IDE).
    • Via OpenRouter/API, Argon generally available sooner and more reliably than via Google’s own UIs.
  • Several users tried to subscribe or upgrade and hit payment failures, plan conflicts (e.g., Google One vs AI Ultra), and geo restrictions.

Product & UX Frictions

  • Launch page criticized as heavy, glitchy scroll-slop, especially on Safari and Firefox; mobile often unusable.
  • Google’s AI product surface seen as confusing: Gemini app, Antigravity, AI Studio, Google One AI, Workspace, all with slightly different models and limits.
  • Error messages and limits are opaque (e.g., generic “terminated due to an error,” flat chunks of usage per query, very small Ultra “5-hour” windows for Argon).

Capabilities & Comparisons

  • Many see Argon as a clear step up from Gemini 3.x, especially for:
    • Prose quality and structure, long-form reasoning, multilingual output (German and Indic languages noted).
    • Spatial reasoning, SVG and UI generation, and 3D/physics-style tasks.
  • Coding:
    • Mixed feedback. Some say it rivals Opus 5.5 on real projects and audits; others find it worse than Opus/Astra/Sol and sometimes below Flash 3.8.
    • One systematic eval places it between Fable 5 and 5.1, “near frontier” but slightly behind top agents.
  • Several users praise it for nuanced medical/technical explanations; others report basic arithmetic slips and instruction-following issues.

Pricing & Usage

  • Intro API price: $2M input / $10M output tokens, doubling to $4/$20M after Dec 1; compared directly to Opus 5.5, Astra, Sol, DeepSeek, GLM, etc.
  • Thread repeatedly notes that “price per token” is misleading:
    • Models differ dramatically in verbosity, tool use, and tokens-per-task, changing effective cost.
  • Subscription side:
    • $20 Pro vs $99 Ultra vs bundled/discounted plans via carriers and Google One.
    • Frustration that Argon-in-Antigravity is locked to $99 Ultra while competing labs expose their best coding models on cheaper plans.

Benchmarks & Technical Findings

  • Standard community tests:
    • SVG “pelican on a bicycle”: Argon’s outputs are among the best seen; some worry this particular benchmark is now overfitted.
    • Other SVG tasks (armadillos, UI diagrams) also look strong; object detection (COCO) significantly improved over earlier Gemini.
    • Pac-Man and coding-agent benchmarks: Argon performs very well but not clearly above the best OpenAI/Anthropic models.
    • German-language benchmark confirms Google’s strong non-English performance; legal-domain score is competitive with specialized legal models.
    • Video generation quality rated only on par with cheaper open Chinese models in one user’s benchmark.

Safety, Loops & Reliability

  • Multiple reports of:
    • Infinite or long reasoning loops, especially in agentic coding; this is highlighted as a regression vs other frontier models.
    • Overly aggressive safety/guardrails leading to refusals in some adversarial or niche-language settings.
    • Classic Gemini issues: hallucinated APIs, fabricated standards, ignoring explicit “discuss before changing” instructions.
  • Some users find Argon “clear in the head” and less RL-fried than prior Google models; others still distrust Gemini for any task requiring strict factual accuracy or long-term consistency.

Meta: AI Race & Google Strategy

  • Debate over whether progress is slowing:
    • Some argue recent deltas are smaller; others point to year-on-year leaps (math, coding, simulation) and see no slowdown.
  • Google is seen as:
    • Highly competitive at the model level (now back “on the Pareto frontier” for general chat and multilingual), but chronically weak at product, rollout, billing, and harness quality.
    • Leveraging its huge user base and integration (Search, Android, Workspace) rather than primarily chasing developer mindshare.
  • Several comments worry about benchmark-chasing (“pelicanmaxxing”) and Goodhart’s law, but others note that competitive pressure among multiple strong labs is good for prices and capability growth.