GPT-4.1 in the API

Original Article ↗ Hacker News Discussion ↗

OpenAI’s release of GPT‑4.1 through its API, along with cheaper “mini” and “nano” variants and a 1M‑token context window, is seen as a major step aimed at coding and agentic workloads, but not clearly state‑of‑the‑art compared with rivals like Gemini 2.5 Pro, Claude 3.7 and DeepSeek on several public benchmarks. Commenters welcome lower prices, faster responses and stronger instruction following, yet highlight that long‑context quality, real‑world coding performance and model stability still lag behind marketing claims. Many are frustrated by opaque benchmarking and confusing model names—especially the deprecation of the more capable but costly GPT‑4.5—and argue that end users shouldn’t have to understand the entire model matrix just to pick the “right” one.

Model naming, versioning, and GPT‑4.5 deprecation

  • Many find the 4.x naming “wild”: 4.1 arriving after 4.5, and 4.5 being deprecated in three months, is seen as confusing and “retconning” the line.
  • Some argue the scheme roughly reflects capability families (4/4o vs o‑series reasoning vs 4.1‑mini/nano), but others say it’s impossible to rank models without documentation.
  • The 4.5 deprecation is attributed by commenters to GPU cost, low usage, and poor cost/latency vs 4.1, despite 4.5 often feeling stronger in creativity and world knowledge.

Benchmarks and SOTA competitiveness

  • OpenAI only compares 4.1 to its own models, which several posters read as a sign they’re no longer clearly ahead.
  • Community benchmarks cited show 4.1 strong but not SOTA in coding: Claude 3.7 and Gemini 2.5 Pro generally score higher on SWE‑bench and Aider Polyglot, often at competitive or lower cost. DeepSeek R1/V3 also feature prominently.
  • Some think 4.1 is likely a distilled 4.5 optimized for efficiency and coding benchmarks.

Coding focus and agentic behavior

  • The release is widely read as a response to Claude 3.7 and Gemini 2.5’s success in coding and agents.
  • GPT‑4.1 mini being roughly 2x faster than 4o at similar reasoning is seen as important for interactive coding tools.
  • Early reports: 4.1 is more “agentic” than 4o but still weaker than Claude/Gemini on large, cross‑cutting refactors; better for small, targeted tasks than complex multi‑scope changes.

Pricing, mini/nano tiers, and context

  • 4.1 is cheaper than 4.5 and 4, with 4.1‑mini and 4.1‑nano targeting Gemini Flash–like price points.
  • Some complain mini got ~2–3× more expensive vs 4o‑mini; others see nano as the real 4o‑mini successor.
  • 1M‑token context across 4.1 models is praised, but several note that beyond ~100–200k tokens most models degrade sharply; announced limits may outstrip practical usefulness.

ChatGPT vs API and routing

  • GPT‑4.1 is API‑only; ChatGPT is said to include “many” of its improvements within 4o‑latest, which some consider vague marketing.
  • Developers value 4.1 as a pinned, stable snapshot, while end‑users express confusion over the growing list of models in the ChatGPT UI and want better automatic routing.

Developer impact and automation debate

  • Some argue front‑end/TypeScript work is “cooked” given tools like v0 and modern models; others report LLMs still fail on non‑trivial refactors and require heavy supervision.
  • There’s concern that labs are explicitly targeting software automation as their key business case, using developer fear as a powerful engagement and marketing driver.

Prompting guidance and eval skepticism

  • OpenAI’s new 4.1 prompting guide draws attention: “persistent” instructions, explicit planning, XML/GDM over JSON for structure, and duplicating instructions at top and bottom. This clashes with prompt‑caching patterns and is seen as more trial‑and‑error empiricism.
  • Benchmarks based on specific tools (e.g., Aider, Qodo) are viewed as useful but also vulnerable to tuning and marketing spin; many insist real‑world testing per use case remains essential.

Overall sentiment

  • Mixed to skeptical: 4.1 is welcomed as cheaper, faster, and better for coding than 4o, but not seen as a clear frontier leap.
  • Several users say they now prefer Gemini 2.5, Claude 3.7, or DeepSeek for many serious tasks, with 4.1 viewed as a strong but no longer dominant option.