GPT-6 Sol and Luna

Original Article ↗ Hacker News Discussion ↗

OpenAI’s launch of GPT‑6 Sol and Luna is seen as a major price and capability move, especially with Luna’s token costs cut roughly in half while remaining “good enough” for many coding and automation tasks. Commenters contrast the new models with Anthropic’s Opus/Fable line and Chinese or open‑weight alternatives, debating real cost‑per‑task, subscription limits, and whether OpenAI is temporarily underpricing to win back market share. There is enthusiasm for cheaper long‑running agents and everyday automation, tempered by worries about model regressions, confusing model/effort choices, and the fragility of building workflows on rapidly changing, opaque pricing and quota policies.

Pricing, economics, and sustainability

  • GPT‑6 Sol and especially Luna are praised as “insanely” cheap: ~50% lower API prices than 5.6, with batch another 50% off. Many say Luna is now effectively “too cheap to meter” for many workloads.
  • Some suspect this is strategic undercutting of Anthropic, DeepSeek, and Chinese/open providers; others argue current prices may be below sustainable cost and enabled by large inference margins or investor subsidy.
  • Several note that cost per task (including cache use and “thinking” tokens), not per-token list price, is what actually matters. On some workflows Luna or Sol can be cheaper than DeepSeek even if sticker prices differ.

Subscriptions, limits, and UX

  • Price cuts reportedly apply to subscriptions (Plus/Pro/20x) via higher effective usage, but details remain opaque; users only see percentage meters, not real token counts.
  • Strong disagreement on which provider is more generous: some find Codex plans burn out far faster than Claude Code; others report the opposite.
  • The Codex 20x plan has been intermittently unavailable to new users, frustrating people who downgraded. Resets are called both a lifesaver and “shenanigans” that distort perceived value.

Model quality vs 5.6 and Anthropic

  • Many describe GPT‑6 Sol as only a small upgrade over 5.6 Sol, with some reporting regressions (e.g., DeepSWE, more “sloppy” code, heavier usage). A number of developers are sticking with or reverting to 5.6 Sol.
  • Luna 6 often tests as slightly worse than Luna 5.6 on coding benchmarks but ~50–60% cheaper, leading some to view it primarily as a price cut rather than a capability jump.
  • Several say Anthropic’s new Opus 5.5 feels stronger for frontier coding, while GPT‑6 shines more on cost/performance.

Model roles: Luna, Sol, Astra, Terra

  • Luna is widely seen as the best value workhorse for data tasks, classification, simple coding, and long-running automation, especially at high/max effort.
  • Sol is treated as a mid‑tier “engineer” or orchestrator; Astra as a powerful but slow, token-hungry planner/designer that can over‑engineer and wander.
  • Many mourn the removal of Terra, which some found to be the ideal “medium‑intelligence” workhorse; others argue Sol High + cheap Luna now dominate its niche.

Agents, context, and tooling

  • Heavy use of agentic workflows: Astra or Sol as planner, Luna or Sol as implementer/tester, with subagents in Codex or third‑party harnesses.
  • Mixed views on aggressive context compaction: some avoid it as lossy and destabilizing; others report hundreds of compactions working fine in long sessions.
  • The proliferation of models × effort levels is seen as confusing; several request an “auto-router” that chooses model and effort based on task instead of manual tuning.