GPT-5.6 Sol, Terra, Luna: OpenAI's cost-efficiency bet vs Fable 5

My notes

Summary

OpenAI launched the GPT-5.6 family - Sol (flagship), Terra (mid), Luna (cheapest) - pitched almost entirely on performance-per-dollar rather than peak score, plus a new “ultra” mode that runs four agents in parallel. The announcement leans hard on beating “Claude Fable 5” on cost-efficiency benchmarks but quietly trails it on SWE-Bench Pro. The Hacker News thread is dominated by a Codex-vs-Claude-Code debate whose rough consensus is that the harness and usage economics now matter more than the underlying model.

Key Insight

Tiers, pricing, effort levels

• Three tiers, priced per 1M tokens: Sol $5 in / $30 out, Terra $2.50 / $15, Luna $1 / $6. The number = generation; Sol/Terra/Luna are durable capability tiers (sun > earth > moon size metaphor, which several people found unintuitive).
• Six reasoning-effort levels (none -> max) plus ultra = four parallel agents by default (more tokens, faster wall-clock). In practice effort levels act as spend dials, not free quality.
• Prompt-caching change: explicit cache breakpoints + 30-minute minimum cache life. Cache writes now billed at 1.25x the uncached input rate; cache reads keep the 90% discount.
• API adds Programmatic Tool Calling (model writes in-memory programs to coordinate tools, filter intermediate data; ZDR-compatible) and a multi-agent beta.

Positioning is cost-efficiency, not peak

• Claims: Sol beats Fable 5 on Agents’ Last Exam by 13.1 pts at ~1/4 the estimated cost; Terra and Luna beat Fable at ~1/16 the cost. On coding, Sol matches Fable’s index (80) using less than half the output tokens, less than half the time, ~1/3 the cost.
• But on SWE-Bench Pro, Sol scores 64.6% vs Fable 80% and Opus 4.8 69.2% - it trails. OpenAI conveniently published a piece days earlier arguing SWE-Bench Pro is unreliable/contaminated (self-serving, but the contamination critique is broadly shared).
• “Fable” is mentioned ~15x in the release notes; commenters read that as competitive anxiety.
• Cyber safeguards claim to block ~10x more harmful activity via a reasoning monitor (not just classifiers); highest-risk capability gated behind Trusted Access with hardware passkeys by Sept 1.

Prompting guidance that generalizes (from the dev guide, surfaced by minimaxir)

• Shorter prompts win: replacing long system prompts with minimal ones raised scores ~10-15% while cutting tokens 41-66% and cost 33-67%. Information density beats length.
• Avoid generic brevity instructions (“be concise”, “keep it short”, “no yapping”) - 5.6 over-corrects and drops required info. Use BLUF prioritization instead: “lead with the conclusion, keep required facts/caveats/next steps, trim intros and repetition.”
• “Be friendly/warm” does nothing measurable; give concrete guidance (“be direct and tactful, avoid canned reassurance and unnecessary sign-offs”).
• Caveat raised in-thread: a non-thinking/auto turn can’t actually “lead with the conclusion” or “trim” before generating - put the polished summary last, or force the thinking model.

Workflow patterns from practitioners

• Cross-model review is the single highest-signal tactic: run one model to write and a different one to review - each reliably catches bugs and over-engineering the other missed. Common split: Claude/Fable plans + reviews, Codex/GPT implements (reversed for math/3D-heavy work).
• Harness > model for most users: Codex and Claude Code are seen as close; the real differentiators are usage economics, reliability/uptime, and remote-control/mobile features - not raw intelligence.
• Codex’s “banked usage resets” (save a reset, spend it when you want) is repeatedly cited as reducing quota anxiety vs Anthropic’s random weekly resets - a loss-aversion retention lever.
• Over-restrictive safety is a real churn driver: Fable reportedly refuses code containing “DNA”/“security”/“capability” tokens, plus plant-care and nutrition questions - pushing biologists and devs to cancel.
• Sandboxing reality: most run --yolo inside a git repo or a disposable VM; auto-reviewers still let through rm and git push --force.
• Ecosystem smoke test: the pelican-riding-a-bicycle SVG (Simon Willison) is the informal “did you actually try the model” check; ongoing debate about whether it’s now saturated/benchmaxxed.