Claude 3.7 Sonnet and Claude Code

Original Article ↗ Hacker News Discussion ↗

Feature convergence & reasoning trend

  • Commenters note rapid copycatting: DeepSeek popularized visible “thinking,” xAI and now Anthropic follow with similar visual/reasoning modes.
  • Debate on whether reasoning is just a “meta-prompt bolt‑on” vs requiring RL and architectural changes; consensus in thread: serious reasoning needs RL and specific training, not just prompting.
  • Some see current releases as evolutionary (small steps since o1/R1), others argue going from GPT‑2‑level chat to IMO medals and agentic coding in <10 years is a massive shift.

Coding focus & Claude Code

  • Broad agreement that coding has been Claude’s comparative strength; many already preferred Sonnet 3.5 over GPT‑4o for real‑world codebases.
  • Claude Code (CLI agent) is seen as a smart way to be editor‑agnostic and “bring the model to the terminal,” though some would prefer IDE‑native plugins.
  • Early users report very strong capabilities (multi‑hour refactors, big speedups, complex scaffolding) but also rough edges: patch errors, bash commands hanging, incomplete long outputs, and no persistent history between accounts.
  • Anthropic staff say Claude Code intentionally exposes raw tool errors and model quirks; it currently relies on agentic search (grep‑style tools) rather than vector RAG for code.

Model behavior & UX preferences

  • Many like Claude’s code skills but dislike its eagerness to emit code when only high‑level discussion is desired; extensive use of custom instructions and “architect first” workflows to mitigate.
  • Some report better results with minimal context than with heavy project contexts; suspicion that long context can hurt answer quality.
  • 3.7 is perceived by some as “smarter but more aggressive,” occasionally ignoring instructions, looping, or overcomplicating solutions.

Costs, limits & billing concerns

  • Pricing is a major theme: Claude 3.7 and Claude Code can burn through dollars quickly; several users hit ~$1 after minutes or $5–10 per dev per day, with intensive sessions hitting “$100/hour” as Anthropic’s own blog notes.
  • Cache reads help a lot in Claude Code, but people still worry about unpredictable bills and want per‑key spend caps, flat‑rate “Ultimate” tiers, or more generous Pro limits.
  • Persistent frustration with tight web‑UI rate limits; heavy users routinely hit caps mid‑debug and fall back to other models.

Comparisons with other models & benchmarks

  • Reports are mixed:
    • Some claim Grok 3 and o1/o3‑mini beat earlier Claude models on complex algorithms; others say they’ve never seen o1 solve something Claude 3.5 couldn’t.
    • New Aider benchmarks put 3.7 Sonnet (no thinking) at the top among non‑reasoning coders, and 3.7‑thinking at SOTA with a large thinking budget—though DeepSeek‑R1+Claude mixtures are very competitive on cost.
  • Several note benchmarks rarely reflect their “vibes”: Claude often “feels right” in large codebases even when charts put it behind.

Open vs closed, privacy & hosting

  • Skepticism toward closed APIs: no way to prove inputs aren’t used for training; some insist only open‑weights or self‑hosted setups are truly trustworthy.
  • Others point to contractual guarantees, use via Bedrock/Vertex, and argue they’re sufficient for most businesses.
  • Discussion on Meta and open‑weights models undercutting economics; expectation that general‑purpose LLMs will commoditize and inference prices trend toward raw compute.

Capabilities, creativity & humor

  • Multiple users are impressed by 3.7’s SVG generation and UI design quality, and by complex math/physics/engineering derivations on first try.
  • A side project (“HN Wrapped”) that uses Claude to roast Hacker News profiles is widely praised as genuinely funny—some see this as evidence of a step‑change in LLM humor and “feel” compared to prior models.

Economic & career anxieties

  • Long subthread on whether AI will erode software jobs: some foresee massive disruption and advise becoming “T‑shaped” (broad stack + deep niche) and using AI as a force multiplier; others think edge‑case complexity, legacy systems, and real‑world ambiguity will keep good engineers in demand.
  • Students express pessimism about picking CS just as AI coding tools accelerate; responses range from “learn to code anyway, you must be able to evaluate AI output” to suggestions to pivot toward products, domain expertise, or starting niche businesses.