Research via TinyFish, 2026-09-11
Side-by-side: Luna effort variants
GPT-5.6 Luna supports six reasoning.effort settings: none, low, medium (default), high, xhigh, max (source: OpenAI API docs). All share the same per-token price; cost rises with longer reasoning traces.
| Effort | AA Intelligence Index | Speed (tps) | TTFT | Cost to run AA Index | Notes |
|---|---|---|---|---|---|
| none | ~14 | - | - | ~5¢/task | Same cost as Medium, but 11 index points lower |
| low | - | - | - | - | Barely above none; rarely worth it |
| medium (default) | 25 | 104.4 | 3.55s | $44 total / 18M tokens | Volume tier; beats GPT-5.5 medium (34) at a fraction of the cost |
| high | 32 | 108.1 | 18.57s | $108 total / 50M tokens | Most compelling upgrade in the family |
| xhigh | 35 | - | - | - | Credible intermediate step |
| max | 38 | 111.8 | 138.61s | $319.93 total / 150M tokens | Highest Luna tier; ~$0.18/task, verbose |
Sources: Artificial Analysis (medium/high/xhigh/max pages), christiant.io value analysis, OpenAI launch post.
Pricing across variants
All effort levels share the same token price — the cost difference comes from longer reasoning traces at higher efforts:
- $0.20/M input, $1.20/M output, $0.02/M cache read (post July 30, 2026 "80% cheaper" cut; was $1/$6)
- Codex credits: Luna ~1x baseline (25 / 2.5 / 150 credits per 1M at low/medium/high effort)
Recommended routing (christiant.io)
- General workloads: Luna Medium (volume) → Luna High (cost-aware default) → Sol Medium
- Coding workloads: Luna Medium → Luna High → Luna Max → Sol High
- Luna Max = premium-value intermediate; Sol xhigh/Max for high-stakes review
Luna vs DeepSeek V4.1 Flash
| Luna (all efforts) | DeepSeek V4.1 Flash | |
|---|---|---|
| Max AA Intelligence Index | 38 (max) | 40 |
| Input price | $0.20/M | $0.15/M |
| Output price | $1.20/M | $0.60/M |
| Cache read | $0.02/M | $0.003/M |
| Speed | 104-112 tps | ~190 tps |
| Context | 1M | 1M (384K max out) |
| Open weights | No | Yes (MIT) |
| Self-host | No | Yes (166.9 GB, 2x H200) |
Key insight
Luna High (32) already beats Claude Fable 5 at a fraction of the cost, and Luna Max (38) is within 2 points of V4.1 Flash's 40 — but V4.1 Flash is still ~25% cheaper on input, half the output price, ~1.8x faster, and self-hostable. Luna's edge is that you can scale effort down (none to medium) for near-free cheap tasks, which V4.1 Flash cannot do.
Bottom line
- Luna for cost-sensitive, high-volume work where you want to dial quality up or down per task.
- V4.1 Flash for the cheapest per-token, fastest, and most open option at the frontier.
- On the same AA v4.3 index they are essentially tied (38 vs 40), but V4.1 Flash wins on price, speed, and openness.
Sources
- OpenAI: GPT-5.6: Frontier intelligence that scales with your ambition (Jul 9, 2026) + July 30 / Aug 21 price updates
- OpenAI API docs (developers.openai.com) — reasoning.effort levels
- Artificial Analysis: GPT-5.6 Luna (medium/high/xhigh/max) model pages + comparisons
- christiant.io — GPT-5.6 Value Analysis
- OpenRouter: deepseek-v4.1-flash, gpt-5.6-luna/sol/terra model pages
- Wikipedia: GPT-5.6