Sonnet 5.5

Original Article ↗ Hacker News Discussion ↗

Anthropic’s release of Claude Sonnet 5.5 is seen as a major step up in coding and “agentic” capabilities, often matching or nearly matching the flagship Opus 5.5 in benchmarks while being cheaper at lower reasoning levels. Commenters are split on its practical value, noting that at higher effort levels Sonnet’s cost approaches or exceeds Opus and that ultra-cheap Chinese models (GLM, DeepSeek, MiMo) and OpenAI’s Luna undercut both on price for many workloads. Much of the debate centers on where Sonnet fits in a rapidly crowding model landscape, ongoing frustration with Anthropic’s aggressive cyber-safety guardrails, and broader worries about how increasingly capable coding agents affect software quality, developer roles, and long‑term economic dynamics.

Model performance & benchmarks

  • Many see Sonnet 5.5 as roughly on par with Opus 5.5 for agentic coding: strong Terminal-Bench, FrontierCode, CursorBench scores; sometimes slightly ahead, sometimes behind.
  • Some argue benchmarks are “benchmaxxed” and overemphasize specific tasks; others point to external community benchmarks (Pac-Man game, ArtificialAnalysis) showing Sonnet 5.5 close behind Opus 5.5 at max effort.
  • Terminal-Bench difference is partly attributed to fewer safety-triggered fallbacks for Sonnet vs Opus, raising questions about how guardrails skew scores.

Cost, pricing, and positioning

  • Repeated complaint: at high/xhigh/max effort, Sonnet 5.5 often costs similar to or more than Opus low/medium for similar or better results; several ask why Sonnet exists on that part of the curve.
  • At low/medium effort, Sonnet 5.5 appears ~30% cheaper and faster than Sonnet 5 and cheaper than Opus, which some see as its real niche.
  • Many compare it unfavorably to cheaper models (OpenAI Luna, Chinese models like DeepSeek/GLM/MiMo), arguing Anthropic is “10–100x” behind on $/speed for “good enough” work.
  • Upcoming Haiku 5.5 is expected to target the low-cost, high-volume slot, but some think it’s already “DOA” unless priced near Luna.

Use cases and harness patterns

  • Common pattern proposed: use Opus/Fable as orchestrator and Sonnet as subagents for grunt work (summaries, refactors, web scraping) to save tokens and increase parallelism.
  • Others say overhead of subagents and cache misses may negate savings; some just run Opus for everything under subscriptions.
  • Sonnet 5.5 is suggested for product-side API calls (parsing user text, simple tasks) rather than heavy dev harnesses where Opus or open models dominate.

Safety, cyber guardrails, and CVP

  • Strong frustration with “cyber” safeguards: authorized security work, old C/C++ code, Bluetooth, WALs, compilers, etc. routinely trigger downgrades to older models.
  • Several say the Cyber Verification Program doesn’t meaningfully relax guardrails and mainly affects post-hoc enforcement, not live blocking.
  • Some argue Anthropic’s safety posture undermines writing secure code and pushes users to uncensored open/Chinese models.

Model behavior & UX

  • Sonnet 5 is widely described as bad value (verbose, pedantic, weaker than open models); 5.5 is perceived as a substantial improvement, faster and more balanced but still below Fable in “feel” for some.
  • Complaints about overlong reasoning at max effort and giant token burn; praise for speed at lower efforts.
  • Anthropic’s marketing charts are criticized as confusing or unfavorable to Sonnet vs Opus; omission of Fable benchmarks is called out.

Chinese and small models vs frontier labs

  • Many report excellent cost/performance from GLM 5.3 Flash, DeepSeek v4.1 Flash, Qwen, MiMo, especially for cyber tasks and “grunt work”.
  • Debate over whether Chinese models are truly close to frontier quality or just “decent value”; some see them as 6–12 months behind but economically decisive, others say frontier models are still clearly better.

AI, work, and existential concerns

  • Extended subthread on “vibe coding”: letting agents generate most code vs humans retaining deep understanding.
  • Some embrace parallel agent swarms and accept less understanding; others fear unmaintainable systems and loss of human agency and meaning in work.
  • Broader worries about wage suppression, K‑shaped outcomes, and coping with AI-driven uncertainty; countered by arguments that understanding problem domains and data structures still matters and limits pure “vibe coding”.

Bugs, pelicans, and benchmarks culture

  • The recurring “pelican on a bicycle” SVG test shows Sonnet 5.5 on max effort hits a 128k output-token cap and fails to respond, highlighting a practical bug; lower efforts work fine and are cheap.
  • Some find this recurring test overexposed; others see it as a useful, visual, and cost-revealing benchmark tradition.