GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday

Original Article ↗ Hacker News Discussion ↗

Model quality vs. Fable/Claude/Gemini

  • Many see GPT‑5.5 as strong for data analysis and coding, but notably weaker than Fable for deep math/linear algebra, orchestration, and long-horizon work.
  • Early preview feedback on GPT‑5.6 Sol: “very capable,” better instruction following and “tenacity,” fixes many 5.5 issues, but still generally perceived as below Fable in raw “smartness.”
  • Some claim 5.6 Sol feels Fable‑level for coding and computer use; others are skeptical this can match a much larger “Mythos‑class” model.
  • Several users still prefer Claude/Fable for backend, architecture, and interface design; GPT models are praised for speed, crispness, and pedantic code review.

Reasoning tokens, context, and agents

  • An OpenAI employee confirms: in the Responses API, reasoning tokens are discarded after each user turn; only input/output tokens are carried forward.
  • Rationale: dropping reasoning tokens allows more work within context before needing lossy compaction, especially with older, shorter‑context reasoning models.
  • Some users feel this design explains why GPT‑5.5 loses “session understanding” over turns; workarounds include saving running context into markdown and re‑feeding it.
  • Others emphasize logging and explicit orchestration (decision logs, task queues, project maps) to make agent work auditable and transferable across models.

Performance, pricing, and size debates

  • Official pricing for Sol reportedly matches GPT‑5.5; some argue Sol/Terra/Luna are mainly rebranded 5.6/mini/nano with extra post‑training, not larger models.
  • One camp predicts similar tokens/second to 5.5 (implying similar size); others note tok/s isn’t a clean proxy for parameters and “bigger isn’t always better.”
  • There’s ongoing argument over whether Mythos‑class models are fundamentally larger than anything OpenAI currently exposes.

Developer UX: Codex vs Claude Code

  • Many prefer Codex for responsiveness, steering, shorter responses, and better image handling; others dislike missing features like /revise and /undo.
  • Claude Code is seen as more powerful for complex engineering but slower, more verbose, and heavily rate‑limited.
  • Some users build tools to share context across Codex/Claude and normalize transcripts.

Naming, access, and product strategy

  • New Sol/Terra/Luna names draw criticism as confusing and marketing‑driven; defenders say they’re better than “mini/nano” which sounded “inferior.”
  • Preview access is being expanded before a broader Thursday launch; exact subscription/TPS details for $20 Codex‑tier users remain unclear.
  • Anthropic’s shifting Fable availability and capacity limits are seen as a weakness versus OpenAI’s larger infrastructure.