OpenAI Releases GPT-6 Sol and Luna at Half the Price

September 23, 2026Provided by Utku Ege Tuluk

OpenAI released GPT‑6 Sol and GPT‑6 Luna on September 22, 2026, three weeks after it launched the flagship GPT‑6 Astra. Sol is the mid-tier model and Luna the low-cost one. Both cost half as much in the API as the GPT‑5.6 models they replace: Sol is $2 per million input tokens and $10 per million output tokens, down from $4 and $20. OpenAI says it trained both models “with similar methods as GPT‑6 Astra.” Its headline claim is that Sol makes about half as many factual mistakes as GPT‑5.6 Sol. Astra is still OpenAI’s top model; Sol and Luna are meant to bring much of its capability to everyday work at a lower price.

Intermediate

GPT-6 Sol and Luna announcement artwork: a glowing sun and a crescent moon against a dark, star-filled sky — Image credit: OpenAI

Pricing and Specs

OpenAI credits the price cut to “improvements in caching and inference.” VentureBeat reports that the new prices are permanent, not a promotion. In the API, the models are called gpt-6-sol and gpt-6-luna.

Model	Input (per 1M tokens)	Output (per 1M tokens)
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

OpenAI’s developer documentation lists these specs for Sol:

Context window: 1,050,000 tokens. Input can be up to 922,000 tokens and output up to 128,000 tokens.
Knowledge cutoff: April 20, 2026.
Input and output: accepts text and images; returns text.
Reasoning effort: six levels: none, low, medium (the default), high, xhigh, and max. Higher levels spend more computation on each answer.
API endpoints: Chat Completions, Responses, and Batch. Fine-tuning and Realtime are not supported.

OpenAI also improved prompt caching, which lets the API reuse input it has already processed instead of processing it again. Reading tokens from the cache is 90% cheaper, and OpenAI says GPT‑6 gets higher cache hit rates by default. Two other changes: developers can now change the reasoning effort or turn tools on and off without losing the cached context, and they can set explicit breakpoints that control which part of a prompt gets cached. GitHub told OpenAI that over the past several months, these improvements more than halved the share of prompt tokens Copilot has to process fresh.

Benchmarks

The figures below come from OpenAI’s announcement. OpenAI notes that it took competitor scores from public reports. Where no Claude Fable 5.1 score was available, it used the Fable 5 score instead.

AutomationBench 1.0.6 (business workflows using 47 tools across sales, marketing, operations, support, finance, and HR): at the xhigh setting, Sol scores 33.2% at about $0.27 per task. That is ahead of GPT‑6 Astra at low effort (30.3%, at 3.9 times Sol’s cost) and Claude Opus 5 at max (26.9%, at 11.1 times the cost). OpenAI puts Claude Fable 5.1 at 31.4%. That run fell back to Opus 5 on about 40% of tasks, so OpenAI says Fable’s reported cost is too low.
Agents’ Last Exam V1 (long professional tasks across 55 sub-industries): Sol scores 56.4% at max. OpenAI says that beats Claude Opus 5’s highest score on this test at 60% lower cost per task.
DeepSWE v1.1 (long software engineering tasks): Sol scores 68.8% at max. Claude Fable 5’s best is 69.9%, at roughly five times Sol’s cost per task. Luna scores 66.6% at max.
OSWorld 2.0 offline (computer-use tasks): Sol scores 60.5% at xhigh. Claude Opus 5 scores 60.3% at medium, at about five times the cost per task.

OpenAI also says Sol improves on GPT‑5.6 Sol on FrontierCode 1.1. That benchmark grades code on whether it is ready to merge (test quality, scope, and style), not only on whether it works. For factuality, OpenAI tested on real ChatGPT conversations in which users had flagged a factual error. It cautions that these conversations are harder than typical use.

On safety, OpenAI says both models score better than their GPT‑5.6 versions on its alignment tests, including making fewer misleading claims about coding work they did. VentureBeat, citing the system card, reports that Sol’s rate on OpenAI’s coding-deception test fell from 10.4% to 1.3%.

Availability

Sol and Luna are rolling out gradually in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can use Luna in the ChatGPT desktop app. OpenAI says the models are “not yet available in Chat.” Both models also arrived in GitHub Copilot on launch day. Sol is available on the Pro+, Max, Business, and Enterprise plans. Luna is also available on Pro. In Copilot, both appear in the model picker across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, and github.com.

GitHub Copilot model picker with GPT-6 Sol selected and GPT-6 Luna listed below it — Image credit: GitHub Changelog

What This Means

The biggest change is the price. For teams running coding agents or other long-running agents, the cost of each task now matters as much as the top score, and GPT‑6 Sol aims to deliver near-Astra results at a fraction of Astra’s cost. VentureBeat points out that Sol now costs the same as Claude Sonnet 5 ($2/$10) and half as much as Claude Opus 5.5. It also notes that Luna costs $0.10/$0.50, against $0.75/$3.75 for Gemini 3.8 Flash at Google’s current introductory prices. TechCrunch reports that Anthropic released Opus 5.5 about 90 minutes before OpenAI’s announcement, so none of OpenAI’s published comparisons include Anthropic’s newest model.

These results should be read with some care. OpenAI ran most of the benchmarks itself or picked them for the announcement, and several comparisons pair Sol at its highest effort settings with competitors at lower ones. The cost-per-task figures are the most useful part, but they depend on the effort setting and on how well prompts are cached. Anyone choosing a model should test it on their own workloads before switching.

Related Coverage

OpenAI Launches GPT-6 Astra, Its First ‘Critical’ Cyber Model: the flagship GPT‑6 model that Sol and Luna build on
OpenAI Launches GPT-5.6 Family: Sol, Terra, and Luna Reach General Availability: the previous Sol and Luna generation
OpenAI Opens GPT-5.5-Cyber to Vetted Defenders via Trusted Access

This post was drafted with AI assistance and reviewed by RITS staff.