Claude Opus 4.1

Original Article ↗ Hacker News Discussion ↗

Initial impressions & performance

  • Early testers report Opus 4.1 feels similar to Opus 4 in casual use: sometimes slightly better at coding and planning, but often slower and not obviously improved.
  • Some users see noticeably better adherence to instructions and multi-step plans, especially in Claude Code and long troubleshooting sessions.
  • Others say it performs worse than Opus 4.0 in Claude Code, with more mistakes and a “Sonnet-like” feel.

Benchmarks, versioning, and expectations

  • Many note Anthropic’s own charts show only modest gains; some argue improvements look small enough to be noise or “one more training run.”
  • Others point to specific coding benchmarks (e.g., “agentic coding,” junior dev evals) where 4.1’s jump is described as a full standard deviation and “a big improvement.”
  • The minor version bump (4 → 4.1) is seen as signaling incremental, not transformative, progress; some lament a perceived slowdown in frontier-model leaps.

Opus vs Sonnet for coding

  • Strong disagreement:
    • One camp: Opus is clearly superior for complex reasoning, debugging, architecture, long unsupervised tasks, and big-picture analysis.
    • Another camp: Sonnet is faster, cheaper, more predictable, and often “good enough” for interactive coding; some even call Sonnet “much better overall.”
  • Common hybrid strategies:
    • Opus for design, analysis, planning, or “plan mode”; Sonnet for implementation and routine edits.
    • Use Sonnet by default, switch to Opus when Sonnet gets “stuck” or hallucinates.
  • Several note Opus is “ridiculously overpriced” via API and only attractive under subscription plans.

Pricing, economics, and limits

  • Heavy users complain Opus API costs and Claude Max usage caps make serious work difficult; some hit Opus limits within minutes.
  • Others report excellent economics on Max plans when combined with caching and disciplined model selection; tools like ccusage are used to estimate “real” API-equivalent spend.
  • Debate over whether Opus’s marginal quality gain justifies ~10x Sonnet’s price, especially when differences feel small in practice.

Release timing & competition

  • Many notice multiple labs (Anthropic, OpenAI, others) releasing models within hours and interpret it as PR “counterprogramming,” not pure coincidence.
  • Some speculate Anthropic’s teaser about “substantially larger improvements in the coming weeks” is partly defensive against an anticipated GPT-5 launch.
  • Others with industry experience argue coordinated release-by-vibes is overstated: real launches take weeks of prep and are often queued, then timed for attention.

Claude Code, tools, and onboarding confusion

  • There is extensive discussion that the ecosystem (Claude web, Claude Code CLI, API, third-party IDEs like Cursor/Cline/Copilot, multiple models/tiers) feels overwhelming to newcomers.
  • Suggested “simple starts”:
    • Pay for Claude Pro or Max and use Claude Code in a terminal, with your usual editor.
    • Or install Cursor (VS Code-based) and switch between Sonnet/Opus there.
  • Clarifications:
    • Claude Code can be used via subscription or per-token billing and essentially wraps the API with an agentic, project-wide editing loop.
    • Sub-agents in Claude Code are highlighted as powerful for isolating context, delegating sub-tasks, and combining models.

Quality regressions, slowness, and behavior

  • Several users complain that:
    • Opus 4.1 and Sonnet 4 feel slower at times.
    • Sonnet’s style has drifted toward more filler, lists, and “sycophancy,” undermining earlier appeal.
    • On some days, overall output quality feels degraded, with more flailing and less crisp reasoning.
  • Others counter that expectations rise quickly, projects grow in size, and “context rot” or long sessions might explain perceived decline.

Benchmarks, reliability, and skepticism

  • Some external benchmarks (e.g., LLM-to-SQL) reportedly do not show Opus 4.1 topping Opus 4.0, raising questions about Anthropic’s highlighted metrics.
  • Users call for more rigorous, repeated, statistically sound benchmarking instead of single-run numbers and glossy charts.
  • There is skepticism that frontier models may be overfitted to benchmark suites, reducing their value as indicators of real-world performance.

Openness and model strategy

  • One thread criticizes Anthropic for never open-sourcing models, branding them “less open” than some competitors.
  • Others note positives on “openness of behavior”: visible chain-of-thought in some settings, explicit “thinking budget,” and relatively low-friction API access compared to KYC-heavy rivals.
  • No consensus emerges on whether this constitutes meaningful openness versus just better product ergonomics.

Productivity claims and limits

  • Some report dramatic productivity boosts (2–10x) using Claude Code for refactors, test coverage, CI pipelines, and tech-debt cleanup; others argue such gains are overstated.
  • A recurring theme: the new bottleneck is code review and trust. Reviewing AI-generated code (which you didn’t author) can be slower and cognitively heavier, capping real-world speedups.
  • A few emphasize that large wins often come from using LLMs to tackle tasks previously too tedious to attempt at all, not just speeding up existing workflows.