Claude Sonnet 5

Original Article ↗ Hacker News Discussion ↗

Price, performance, and effort levels

  • Many readers dissect Anthropic’s own cost/performance charts and system card.
  • Broad consensus: Sonnet 5 is clearly better than Sonnet 4.6, especially at low/medium reasoning.
  • But Opus 4.8 often beats Sonnet 5 on a Pareto basis for higher effort levels (agentic search/computer use): for the same cost per task, Opus tends to perform better.
  • Common takeaway: “Use Sonnet 5 on low/medium; if you need more reasoning, just use Opus.”
  • Several note that Sonnet 5’s new tokenizer yields ~1.0–1.35× more tokens for the same text. Combined with post‑promo pricing, some see this as a stealth price increase.

Positioning vs open and Chinese models

  • Multiple comparisons to GLM 5.2, Qwen, Kimi, DeepSeek, Gemma, etc.
  • Rough user consensus: Sonnet 5 is in the GLM 5.2 / “Haiku‑plus” band, but:
    • Often slower and/or more expensive than open‑weight competitors.
    • Still generally more reliable in long‑horizon, multi‑turn, and instruction‑following tasks.
  • Some argue open‑weight models are “benchmaxxed” and underperform in real workflows; others report the opposite.

Agentic vs assisted development

  • Anthropic markets Sonnet 5 as “more agentic.” Many are wary:
    • Complaints that increasingly agentic models ignore instructions, over‑act, or over‑complicate tasks.
    • A camp prefers “agent‑assisted” workflows: big model for planning/review, smaller model for mechanical edits.
  • Others report success with Sonnet/Opus subagent architectures but note that mis‑scoped tasks and weak small models can create technical debt and slop.

Cybersecurity nerfing and regulation

  • Anthropic highlights Sonnet 5’s weaker cybersecurity capabilities relative to Opus/Mythos.
  • Many see this as messaging to US regulators after the Mythos/Fable episode, not a user benefit.
  • Strong concern that “dumber at cyber” also means more insecure code and weaker defensive tooling.
  • Broader worry that top “defensive/offensive” models will be restricted to governments and large institutions.

Quotas, pricing, and routing

  • Users on Claude subscriptions note Sonnet often doesn’t feel much cheaper than Opus due to verbosity and hidden reasoning tokens.
  • Confusion over subscription usage bars (Sonnet vs Opus pools), automatic routing, and model downgrades.
  • Frustration with opaque costs, frequent model/effort changes, and perceived token‑burning behavior.

Sentiment about Anthropic and Fable

  • Noticeable shift toward skepticism: lobbying against open weights, export‑control drama, rapid deprecations, and perceived “nerfs.”
  • Many still like Claude’s UX and coding quality, but some are moving work to Codex, open models, or local setups.
  • Strong desire for return of Fable/Mythos; Sonnet 5 is widely seen as a workhorse upgrade, not an exciting frontier model.