GPT‑5.4 Mini and Nano

Original Article ↗ Hacker News Discussion ↗

Pricing, Performance, and Positioning

  • Mini/Nano are seen as attractive for “simple” or high-volume tasks due to lower cost and latency, though prices are notably higher than prior GPT‑5 mini/nano generations.
  • Some argue models are “more expensive but cheaper per unit of capability,” others say the low-end pricing has been “thoroughly hiked” and hurts volume use cases.
  • Reported speeds over API: GPT‑5.4 mini ~180–190 tokens/s, nano ~200 tokens/s, substantially faster than older GPT‑5 mini and competitive with Gemini Flash; however, prompt-processing latency and TTFT remain unclear and are a pain point for some.
  • Benchmarks: GPT‑5.4 mini scores well on many tests (including “how many Rs in strawberry” type sanity checks and OSWorld computer-use), sometimes approaching or matching more expensive models, but long‑context performance is criticized.

Mini vs Nano and Reliability

  • Several commenters find GPT‑5.4 mini strong and a good default when precision matters.
  • GPT‑5.4 nano is praised for speed and cost, but often seen as less reliable for precise tasks; some benchmarks oddly show nano > mini, and others report mini behaving inconsistently even at temperature 0.
  • For multi-agent pipelines, there’s concern that naïve orchestrators send huge contexts to “cheap” nano calls, negating cost/latency advantages.

Comparisons with Competitors and Open Models

  • Claude’s Haiku/Sonnet and Gemini Flash/Flash Lite are frequent reference points; many find Claude better for tool use, instructions, and agentic work, with GPT models described as slower, more “robotic,” and more prone to guardrail refusals.
  • Others strongly prefer Codex/GPT for coding quality, using mini models as cheaper subagents in workflows.
  • Some report open models (Qwen, GLM, K2.5, etc.) as competitive at lower cost, though opinions vary on whether they match GPT‑5.4 mini/nano.

Use Cases and Practical Experiences

  • Common use cases: code generation and refactoring, automated PRs, computer-use agents (OpenClaw/OSWorld), PDF/invoice parsing, log analysis, content labeling at scale, and voice agents where latency is critical.
  • Mini models are viewed as especially important for making these “real-world” applications economical.

Transparency, Strategy, and Fatigue

  • Frustration that OpenAI doesn’t disclose model sizes or open-source weights; some say without open weights these releases are less interesting.
  • Concerns about rising safety friction (overactive guardrails, anti‑sycophancy) and “version fatigue” from frequent incremental releases and confusing naming.
  • Some threads criticize OpenAI’s business trajectory versus Anthropic and express general numbness to yet another model announcement.