GPT 6 Sol and Luna
Photo by Jongsun Lee on Unsplash
GPT 6 Sol and Luna are OpenAI's cheaper additions to the GPT 6 family, released on 22 September 2026.
GPT 5.6 Luna was already wild - absolutely wild - value for money. And now it's half the price again. Incredible.
The launch came on the same day as Claude Opus 5.5. OpenAI releasing right on top of an Anthropic launch has become a pretty familiar story.
Context
Context window is a total of 1M, for all models, and they're multi-modal.
| Specification | GPT 6 Sol | GPT 6 Luna |
|---|---|---|
| API identifier | gpt-6-sol |
gpt-6-luna |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Maximum input | 922,000 tokens | 922,000 tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Input / output | Text and images / text | Text and images / text |
Both support tools and structured outputs. Reasoning effort ranges from none through low, medium, high, xhigh and max, with medium as the API default. Unlike Opus 5.5's always-on adaptive thinking, these models still offer none. Use the Responses API for tools with reasoning enabled. Sol documentation, Luna documentation.
Pricing and Cache Settings
Standard direct API prices, checked on 23 September 2026, in USD per million tokens. The GPT columns apply to requests with at most 272K input tokens.
| Usage | GPT 6 Sol | GPT 6 Luna | Claude Opus 5.5 |
|---|---|---|---|
| Uncached input | $2.00 | $0.10 | $4.00 |
| Cached input read | $0.20 | $0.01 | $0.20 |
| Cache write | $2.50 | $0.125 | $5.00 (5 minutes) / $8.00 (1 hour) |
| Output, including reasoning | $10.00 | $0.50 | $20.00 |
Sol is half Opus 5.5's price for uncached input and output, but their cache reads cost the same. Different tokenizers and reasoning lengths also mean a token-price ratio is not a task-cost ratio. OpenAI pricing, Claude pricing.
Sol also now matches GPT 5.6 Terra's input and caching prices, with cheaper output: $10 versus $12 per million tokens, a 16.7% reduction. OpenAI pricing.
For Sol and Luna, prompts above 272K input tokens cost 2x the input and cache rates and 1.5x the output rate for the whole request. Batch and Flex halve standard rates; Fast mode doubles the applicable rates. Model pricing notes.
GPT 6 caching is enabled by default. A cacheable prefix needs at least 1,024 visible input tokens, and entries last at least 30 minutes after the latest write or reuse. Cache writes are billed at 1.25x the base input rate, replacing rather than adding to the ordinary input charge. Reads cost 10% of the base rate. Explicit breakpoints give more control over which prefixes can be reused; a repeated passage alone does not guarantee a hit. Prompt caching documentation.
Benchmarks
OpenAI's launch comparisons mostly use Opus 5, which Opus 5.5 has just replaced. The table below adds the new model using Anthropic's system card, sections 8.3 and 8.14.6 and Zapier's AutomationBench 1.0.6 leaderboard, checked on 23 September 2026.
| Benchmark | GPT 6 Sol | GPT 6 Luna | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE v1.1 | 68.8% (max) |
66.6% (max) |
74.2% (max) |
| AutomationBench 1.0.6 | 33.2% (xhigh) |
No exact score in launch text | 40.0% (max) |
DeepSWE tests extended software engineering tasks. The GPT scores come from OpenAI's announcement; Anthropic reports Opus 5.5's score averaged over five trials. Opus leads these reported results by 5.4 percentage points over Sol and 7.6 over Luna. These are developer reports, not my own tests with an identical Coding Harness.
AutomationBench tests business workflows across apps. Zapier reports $0.27 per task for Sol at xhigh, versus $1.28 for Opus 5.5 at max: Sol costs about 79% less, but succeeds on fewer tasks. Opus 5.5 at high scores 32.0% at $0.65 per task. Effort labels are not equal compute budgets across providers, and cost per attempted task is not cost per successful task. Leaderboard.
Computer-use results need extra care. OpenAI reports 60.5% for Sol at xhigh on OSWorld 2.0's offline set, using partial credit and the August 8 release. Anthropic reports 81.8% partial credit and 48.7% strict success for Opus 5.5, but explicitly says its updated task files and context-management setup prevent direct comparison with earlier releases or other harnesses. Comparing Sol's 60.5% with Opus's 48.7% would mix different metrics as well as setups. OpenAI results, Anthropic system card, section 8.13.3.
Real example
This example uses GPT 6 Sol at high effort to turn the article into a 3D gameshow with an N64 look, using the same prompt as the Opus 5.5 example.
It uses the OpenAI Python SDK and my Obsidian Markdown Notebook plugin. The model call stays visible; the Notebook HTML helper saves the game and token usage, then renders it.
Download the Notebook HTML helper, then install it in your notebook's Python environment:
pip install openai ./notesbylex_notebook_html-0.1.2-py3-none-any.whl
Set OPENAI_API_KEY in the environment used to launch Obsidian. Running the cell makes a paid request; the saved game plays without an API key or another model call.
import os
from openai import OpenAI
from notebook_html import read_article, render_openai_html
api_key = os.environ["OPENAI_API_KEY"]
article = read_article("gpt-6-sol-and-luna.md", before="Real example")
with OpenAI(api_key=api_key).responses.stream(
model="gpt-6-sol",
max_output_tokens=128_000,
reasoning={"effort": "high"},
service_tier="default",
store=False,
instructions="Return a complete, self-contained HTML document.",
input=f"""
Turn this article into a 3D gameshow that looks like an N64 game.
Use around 8 questions that test understanding of the article, with a short
explanation after each answer.
Support phones and keyboards. Use inline CSS and JavaScript, no external assets,
network requests or browser storage, and no em dashes. Keep the code compact,
with one shared question renderer.
Article:
{article}
""",
) as stream:
response = stream.get_final_response()
render_openai_html(response, "../_media/gpt-6-sol-and-luna-quiz-high.html")
This game was generated on 23 September 2026 using gpt-6-sol at high effort, with the standard service tier. It cost about US$0.0868, calculated from the saved token usage and standard API prices:
| Usage | Tokens | Estimated cost (USD) |
|---|---|---|
| Input, ordinary uncached | 3 | $0.0000060 |
| Input, cache write | 1,647 | $0.0041175 |
| Output, including reasoning | 8,270 | $0.0827000 |
| Total | 9,920 | $0.0868235 |
There were no cache reads. Of the output tokens, 3,485 were reasoning tokens and 4,785 were the HTML response. Reasoning is included in the output charge. This request incurred cache writes because caching was left at its default setting.
For comparison, the Opus 5.5 quiz cost about US$0.96, making this Sol run roughly 91% cheaper. Both used the same task prompt at high effort, with different articles as input.
Play the quiz in its own page.
The generated HTML is unchanged. Browser checks covered correct and incorrect answers, scoring, replay, keyboard controls, a 390px phone viewport and the sandboxed notebook embed. This is one small coding example, not a benchmark of either model's overall ability.
To render the saved game again without calling the model, use load_html("../_media/gpt-6-sol-and-luna-quiz-high.html") from notebook_html.