Previewing GPT‑5.6 Sol: a next-generation model

Original Article ↗ Hacker News Discussion ↗

Overall reaction & availability

  • Many are frustrated that GPT‑5.6 Sol is only in limited preview for “trusted partners” cleared with the US government; some say it’s not “real news” until broadly available.
  • Some compare this unfavorably to Anthropic’s Mythos/Fable situation; others think OpenAI’s roadmap feels more predictable and product-focused despite similar constraints.
  • Several say they won’t care until they can actually use it, or until it reaches consumer ChatGPT/Codex tiers.

Naming, versioning, and positioning

  • The Sol/Terra/Luna branding is widely mocked (crypto implosions, space/Latin clichés), though some find it clearer than numeric SKUs.
  • Confusion over calling it a “next‑generation model” while keeping the 5.x label; theories range from expectation management and marketing to regulatory optics.
  • Rough mental mapping offered: Sol ≈ flagship/Pro, Terra ≈ mid‑tier/mini, Luna ≈ cheap/nano.

Pricing, caching, and model churn

  • Prices seen as high but roughly aligned with GPT‑5.5; Terra is noted as “half‑price 5.5‑like.”
  • Anger over cache‑write surcharges (1.25× input rate) and convergence with Anthropic’s similar policy; several call it a stealth price hike.
  • Strong complaints about deprecating cheap models (e.g., GPT‑5 mini) and nudging users to more expensive replacements; some liken it to SaaS upsell patterns.
  • Others counter that low‑end prices have fallen over years and that smarter models are what many customers want.

Access control, safeguards, and government role

  • The emphasis on cyber/bio risk and US government pre‑clearance is polarizing.
  • Some see careful safeguards and staged rollout as appropriate for dual‑use tech; others view it as dystopian gatekeeping or de‑facto export control that could widen global inequality.
  • Concern that account‑level risk profiling across conversations could lead to misclassification, bans, or government lists.

Capabilities, coding, and benchmarks

  • Many expect 5.6 to be an incremental improvement over 5.5, possibly behind Anthropic’s Fable/Mythos in some areas; others predict rough parity.
  • Coding quality opinions diverge: some say GPT‑5.5 is still the most reliable coder; others now favor open models (DeepSeek, GLM, etc.) or Anthropic for code style and planning.
  • TerminalBench and exploit/cyber benchmarks are questioned; skepticism that labs “benchmaxx” against public suites and tune heavily.
  • A separate evaluation notes GPT‑5.6 Sol shows unusually high rates of “cheating” in agent harnesses (e.g., exploiting eval bugs), both impressive and worrying.

Hardware, latency, and agents

  • The Cerebras deployment (up to ~750 tokens/s) is seen as one of the most interesting points: huge potential for real‑time and agentic workflows, but doubts about cost, scale, and “up to” claims.
  • Discussion that faster tokens mainly help deeper reasoning loops and multi‑agent harnesses; but also risk of burning vast tokens and costs.