---
title: 'In the News: September 22, 2026 (Extra 3)'
description: 'OpenAI halves API prices for GPT-6 Sol and Luna and says reasoning-effort and tool changes no longer break the prompt cache.'
canonical_url: 'https://darkfactory.dev/news/2026-09-22-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-09-22-extra-3.md'
collection: news
date_published: '2026-09-22T16:07:00-04:00'
date_modified: '2026-09-22T16:07:00-04:00'
---

# In the News: September 22, 2026 (Extra 3)


OpenAI released GPT-6 Sol and GPT-6 Luna today, the cheaper tiers below GPT-6 Astra, at half the API price of their GPT-5.6 predecessors. For anyone running agents, the caching changes in the same announcement may matter more than the price cut.

## 1. OpenAI halves Sol and Luna prices and lets agents change settings without losing the cache

**[Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/)** · OpenAI · September 22, 2026

GPT-6 Sol lists at $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's promotional $4 and $20. GPT-6 Luna lists at $0.10 and $0.50, down from $0.20 and $1.20. OpenAI says it improved prompt caching for GPT-6 to raise hit rates by default, with a 90 percent discount on cached input reads. Two controls now keep earlier context available for cache reuse: changing reasoning effort mid-conversation, and enabling or disabling tools "as your agent's needs change." Developers can also set explicit breakpoints that decide where a cached prefix ends. OpenAI cites GitHub as reporting that these caching improvements cut the share of prompt tokens needing fresh processing by more than 50 percent across billions of requests over the past several months.

On coding, OpenAI reports GPT-6 Sol at max effort scoring 68.8 percent on DeepSWE v1.1, 1.1 points behind Claude Fable 5's 69.9 percent at xhigh effort, at roughly 80 percent lower cost per task. It reports Luna at max effort at 66.6 percent, at 93 percent less per task than Claude Opus 5. These are OpenAI's own figures, and the competitor scores come from public reports rather than OpenAI's own runs. Both models are available today in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise and Edu users, and in the API as `gpt-6-sol` and `gpt-6-luna`. The Hacker News thread had 651 points and 372 comments about two hours after it was posted.

**Why it matters:** Per-step tuning inside a harness gets cheaper. On OpenAI's account, an agent can drop reasoning effort for an easy step or swap its tool set mid-run without giving up the cached context it has built. With list prices halved, and Anthropic's Claude Opus 5.5 listed at $4 and $20 earlier today, the cost comparison for long agent runs has moved twice in one day.