OpenAI O3-Mini

Original Article ↗ Hacker News Discussion ↗

Model role, hierarchy, and benchmarks

  • Many try to place o3-mini in a rough hierarchy (e.g. somewhere between GPT‑4o and o1, above o1‑mini/4o‑mini), but there’s no consensus; performance is clearly task‑dependent.
  • On coding, several report o3‑mini‑high tying or beating o1/o1‑mini on their own tasks, especially with high reasoning effort, while others find o1 still clearly better on tricky math or geometry.
  • SWE‑Bench numbers are scrutinized: the headline 61% involves an internal tools scaffold; the “agentless” setting is closer to high‑40s and only slightly above o1‑mini, which some see as benchmark “benchslop.”
  • Codeforces/ARC‑style reasoning scores look strong, but multiple commenters argue competitive programming is a poor proxy for real software engineering.
  • Some see evidence of diminishing returns: newer models often feel “incremental” and it’s getting hard for non‑experts to tell which is better.

Speed, cost, tiers, and rollout

  • o3‑mini is praised for speed and cost: reasoning comparable to o1 on many coding tasks at a fraction of o1’s token price, with 200k/100k context windows.
  • The three “reasoning_effort” levels (low/medium/high) are liked conceptually; people want similar control on o1 via API.
  • ChatGPT Plus limits: ~150 o3‑mini messages/day and ~50 o3‑mini‑high/week, separate from o1 limits.
  • Rollout was staggered by API tier; several complain the blog said “available today” but it appeared hours later.
  • Some o1‑pro subscribers say the model was silently changed (shorter thinking time, lower quality), reinforcing a long‑standing complaint that OpenAI swaps models behind the same name without notice.

Comparisons: DeepSeek, Claude, Gemini, Mistral

  • DeepSeek R1 is widely praised for visible chain‑of‑thought, local runnability, and low price, but also criticized as buggy, often down, and less reliable in real apps; some find OpenAI/Gemini more robust.
  • Claude 3.5/3.6 Sonnet is frequently described as the best day‑to‑day coding assistant (especially in Cursor/Aider), with o‑series or R1 used as “architect” models for harder reasoning.
  • Gemini 1.5 Pro/Flash and Flash‑Thinking get good marks for reasoning and huge context, but pricing and production robustness are questioned.
  • Mistral Small 3 is seen as roughly 4o‑mini‑class; o3‑mini is considered a different (stronger) tier, pending more external benchmarks.

Chain‑of‑thought visibility and alignment

  • Many want OpenAI to expose reasoning traces like DeepSeek and Gemini, both for debugging prompts and for trust; paying for hidden “thinking tokens” is resented.
  • Others note OpenAI staff have hinted at a more detailed but not fully raw CoT view “coming soon.”
  • The rename of system → developer messages is widely interpreted as a jailbreak‑hardening move; some suspect it’s also meant to break drop‑in “OpenAI‑compatible” stacks.
  • Alignment vs capability trade‑offs (e.g. safety filters lowering benchmark scores) are debated; some call safety “lobotomization,” others argue guardrails are necessary and analogous to bug‑fixing.

Naming, branding, and product strategy

  • Model naming (4o vs o1/o3, minis, previews, high/medium/low) is a dominant complaint: people find it opaque, hard to teach to non‑technical users, and reminiscent of Xbox/USB/Azure SKU chaos.
  • Several argue OpenAI should surface just “ChatGPT” and “ChatGPT‑mini” for normal users and hide model SKUs behind an advanced menu, or adopt a simple versioned scheme (e.g. ChatGPT 5, 5‑mini).
  • Some think the confusion is intentional marketing: every “o‑something” sounds like a breakthrough even when improvements are modest.

Competition, moats, and trust

  • There’s sharp disagreement on OpenAI’s “moat”: some see brand and infra scale as enduring advantages; others say DeepSeek’s efficiency and open weights show big‑spend, closed models are vulnerable.
  • Closed vs open is a major axis: DeepSeek’s open weights and local deployability inspire trust for some; others worry about PRC alignment, censorship, or subtle steering even in local copies.
  • OpenAI’s data‑use policies are debated: API/enterprise are opt‑out‑by‑default from training, but some remain skeptical given earlier “open” rhetoric and evolving terms.

User experience, workflows, and fatigue

  • Many now mix models: e.g. Sonnet for implementation, R1 or o‑series for reasoning, Gemini for long‑context analysis, DeepSeek locally for experiments.
  • Several say the real frontier is moving from raw model IQ to UX, agents, and stability; frequent model churn and subtle regressions make people hesitant to “bet the company” on any single provider.
  • A noticeable group is underwhelmed, comparing the current pace to smartphones: faster, cheaper, more SKUs, but not a qualitative leap beyond GPT‑4‑class for everyday use.