GPT-5.6

Original Article ↗ Hacker News Discussion ↗

Model performance & benchmarks

  • Many see GPT‑5.6 Sol as a strong coding/model upgrade over GPT‑5.5, with better planning, fewer bugs, and more “get it done” behavior, sometimes rivaling or surpassing Fable in day‑to‑day work.
  • Others report Sol still trails Fable/Mythos on hard reasoning or code‑reading tasks; some say Fable remains uniquely good at deep code review and architecture.
  • Benchmarks spark debate:
    • OpenAI’s “Agents’ Last Exam” and other in‑house graphs are criticized as cherry‑picked, with misleading axes and non‑max Fable settings.
    • SWE‑Bench Pro is called “broken” or saturated by multiple commenters; some note OpenAI now deemphasizes it while promoting DeepSWE and FrontierCode where GPT models score better.
    • Sol’s 7.8% on ARC‑AGI‑3 is seen as genuine progress but still far from “AGI”.

Developer experiences & harnesses

  • Large subthread on Codex vs Claude Code:
    • Many praise Codex for lower drama, better stability, clearer permissions, and more generous usage; they dislike Claude Code’s bloat, bugs, outages, and opaque quotas.
    • Others feel Claude/Fable produce more elegant, intention‑aligned code and better design/UI work, with Codex described as more literal, verbose, or “austere”.
    • A common pattern: use both—one plans, the other implements or reviews; some chain multiple models (e.g., Fable + GPT + DeepSeek) via OpenCode, Pi, or other harnesses.

Pricing, quotas & efficiency

  • 5.6 pricing (Sol/Terra/Luna) is viewed as competitive versus Fable; Terra/Luna in particular look attractive on intelligence‑per‑dollar graphs.
  • Strong focus on token efficiency: several note Sol is far more efficient than Fable and Opus on cost‑per‑task benchmarks, but:
    • Sol Ultra and heavy subagent use can burn through 5‑hour windows extremely fast; some users see entire windows consumed in minutes.
  • OpenAI’s bankable quota resets are widely praised; Anthropic’s random resets and Fable’s tight subscription caps are widely disliked.

Safety, guardrails & policy

  • Fable’s safety filters (especially around biology, security, and even innocuous code with “DNA” or “security” terms) are a major pain point; many biologists and security researchers report abandoning Anthropic over this.
  • Some welcome OpenAI’s stated stance against over‑blocking cyber‑defense tasks, but others already see new “additional safety checks” and slowdowns in 5.6.
  • There is unease about all major labs’ ties to governments and defense (Palantir, Maven, etc.); some commenters say they now prefer open or non‑US models.

Prompting behavior & UX

  • OpenAI’s own guidance (shorter prompts, avoid generic “be concise”) is discussed heavily:
    • Some like the shift away from “verbal diarrhea”; others complain that needing long, precise brevity instructions defeats the purpose of “just be short”.
  • Users report 5.6 is slower than 5.5 in many harnesses, especially at higher reasoning levels, but often higher quality.
  • The new Sol/Terra/Luna naming is called both intuitive (sun/earth/moon by size) and confusing marketing compared to older “mini/nano” or simple “Pro/Max” labels.

Miscellaneous reactions

  • The long‑running “pelican on a bicycle” SVG test shows clear qualitative differences across Luna/Terra/Sol and reasoning levels; some argue the test is now in training data and partly saturated, others still find it a useful visual sanity check.
  • Overall sentiment: excitement about strong competition to Fable and better cost/efficiency, mixed with skepticism about benchmark marketing, safety trade‑offs, and rapidly escalating complexity in model/effort choices.