# Astra Is the Ceiling. Sol and Luna Are What Agent Loops Will Pay.

Times of AI Desk · 2026-09-22 · Models

[https://timesof.ai/2026/09/openai-gpt-6-sol-and-luna](https://timesof.ai/2026/09/openai-gpt-6-sol-and-luna)

> OpenAI halved Sol and Luna API prices versus GPT-5.6 promo rates and put them in Work and Codex the same day. Lab cost-per-task gaps look large; Artificial Analysis still has Opus 5.5 clearly ahead of Sol on intelligence — the trade this launch asks buyers to accept.

Astra stays the ceiling. Sol and Luna are the prices ordinary agent loops will actually pay.

OpenAI released **GPT-6 Sol** and **GPT-6 Luna** — the cheaper tiers under **GPT-6 Astra** (September 3) — and says caching and inference gains let it **cut API prices 50%** versus **GPT-5.6 promotional** rates. The lab’s cost-per-task gaps are large. The independent intelligence index still has **Claude Opus 5.5** clearly ahead of Sol. That is the trade this launch is asking buyers to accept.

## What shipped

| | Input / MTok | Output / MTok |
| --- | --- | --- |
| GPT-5.6 Sol → **GPT-6 Sol** | $4 → **$2** | $20 → **$10** |
| GPT-5.6 Luna → **GPT-6 Luna** | $0.20 → **$0.10** | $1.20 → **$0.50** |

Same-day availability: ChatGPT **Work** and **Codex** for Plus, Pro, Business, Enterprise, and Edu. **Free and Go** get **GPT-6 Luna in the desktop app**. **Not yet in Chat.** API slugs: **`gpt-6-sol`** and **`gpt-6-luna`**. Rollout through the day for stability. OpenAI says they were trained with methods similar to Astra.

**Caching** (claims in the Sol post; a separate “Better prompt caching for GPT-6” item ran the same day): higher default cache hit rates, **90%** discount on cached input-token reads, a prompt-caching dashboard, and the ability to change reasoning effort or tool availability without breaking the cache. OpenAI says GitHub, over several months, cut the share of prompt tokens needing fresh processing by **more than 50%** across billions of requests.

## Lab claims

Effort levels called out; competitor scores taken from public reports; Fable 5 used where Fable 5.1 was unavailable.

- **AutomationBench 1.0.6** (Zapier, 47 tools): Sol at xhigh **33.2%** and **$0.27 per task**. OpenAI says that beats Claude Opus 5 at max (**26.9%**) at **9%** of Opus 5’s cost, beats low-effort Astra (**30.3%**, **3.9×** the cost), and beats Fable 5.1 with an Opus 5 fallback (**31.4%**) whose fallback cost OpenAI says is **omitted** — the Fable point “understates its actual cost” because fallbacks hit on about **40%** of tasks. Luna at high effort is **5.4 points** above its predecessor at **58%** lower cost per task.
- **Agents’ Last Exam V1**: Sol at max **56.4%**, which OpenAI says is above Opus 5’s highest score on that eval at **60%** lower cost per task. Opus 5’s score is not printed in the post.
- **DeepSWE 1.1**: Sol at max **68.8%**, within **1.1 points** of Fable 5 at xhigh (**69.9%**), at about **80%** lower cost per task. Luna at max **66.6%**, described as comparable to Opus 5 and Fable 5 at medium effort, at **93%** and **96%** lower cost.
- **OSWorld 2.0 offline**, partial reward, v2026.08.08: Sol at xhigh **60.5%** versus Opus 5 at medium **60.3%**, about **80%** lower cost. Astra remains OpenAI’s computer-use leader.
- Internal factuality set (de-identified chats where users had flagged an error — not typical traffic): Sol makes **about half as many mistakes** as its predecessor. Luna at higher effort matches GPT-5.6 Sol at about a **hundredth** of the cost.

## Independent checks

- **Artificial Analysis** comparison page: Sol at max effort intelligence **48**, against Opus 5.5 max-effort **58**. Sol medium is the faster of the two on that page at **114 tokens/s**. Sol low is the cheaper task at **$0.13**, against Opus 5.5 low at **$0.55**.
- **Andon Labs**, September 24, Vending-Bench 2: Sol averaged **$14,428**, second only to Astra, and won **three of four** arena games against Opus 5.5 and Grok 4.7. API cost for a simulated year: **$104**, against **$810** for Astra. Andon says that is about **93%** of Astra’s score at an eighth of the price, and about **half again** what GPT-5.6 Sol made.
- Arena Code Elo: a September 25 public snapshot still showed **GPT-5.6 Sol**, not GPT-6 Sol. No new Elo is claimed.

## Limits

- Chat availability is still missing; Work and Codex are not a full ChatGPT substitute.
- Several lab benches lean on effort-level and fallback-skewed comparisons; OpenAI itself flags the Fable AutomationBench cost omission.
- Arena Code Elo has not yet priced GPT-6 Sol.

## Sources

- [OpenAI: Introducing GPT-6 Sol and Luna (September 22, 2026)](https://openai.com/index/introducing-gpt-6-sol-and-luna/)
- [Artificial Analysis: Claude Opus 5.5 vs GPT-6 Sol](https://artificialanalysis.ai/models/releases/comparisons/claude-opus-5-5-vs-gpt-6-sol)
- [Andon Labs: Opus 5.5, GPT-6 Sol and Grok 4.7 on Vending-Bench (September 24, 2026)](https://andonlabs.com/blog/opus-5-5-gpt-6-sol-grok-4-7-vending-bench)