GPT-6 Sol and Luna for marketers: what they change

What OpenAI released

In September 2026 OpenAI added two models to its GPT-6 family. GPT-6 Astra, launched earlier the same month, stays the flagship. Sol and Luna sit below it on price: Sol is the capable middle, Luna the fast and cheap one. OpenAI says both were trained with methods similar to Astra's and carry its gains in professional work, factual accuracy, coding and computer use into cheaper models.

In the API they are gpt-6-sol and gpt-6-luna. In ChatGPT they are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu plans, and Free and Go users get Luna in the desktop app. The full details are in OpenAI's announcement.

Per million tokens, from OpenAI's announcement. Cached input is read at a 90% discount. Every benchmark figure below is also OpenAI's own, run on its own setup, so read them as a direction rather than a promise for your account.

Why the price matters more than the benchmarks

Most of the marketing work AI helps with is not one brilliant answer. It is the same small job done hundreds of times: ten headline variants per ad set, a caption rewritten for five channels, a thousand comments sorted by intent, a weekly summary for every client. At that volume the price per task decides whether you use AI for the job at all.

A rough illustration. Say one caption variant takes about 300 tokens of instructions and 150 tokens of output. Five hundred variants cost about five cents on Luna at list price, and about a dollar on Sol. Reasoning tokens bill as output and grow with the effort setting, so real bills run higher, but the point holds: producing options is no longer the expensive part. Choosing and checking them is.

That moves the bottleneck from writing to approving, which is why every workflow in overads starts out asking a person before anything goes out. When drafts are nearly free, the approval step is the part worth designing well.

Which model for which marketing job

Start high-volume, low-stakes jobs on Luna, give multi-step jobs to Sol, and keep Astra for work that is expensive to get wrong.

Start each repeatable job on Luna and move it up only when you can see it failing. The common mistake is the reverse: paying mid-tier prices to sort comments.

Agents that work across your tools

The benchmark closest to a marketing team's week is AutomationBench, which tests agents on end-to-end business workflows using 47 tools across sales, marketing, operations, support, finance and HR. OpenAI reports Sol at its xhigh effort setting scoring 33.2% at $0.27 per task. That puts it ahead of Claude Opus 5 at maximum effort, at 26.9%, for roughly 9% of the cost per task. At high effort, Luna improves on its predecessor by 5.4 points at 58% lower cost per task.

Read the absolute number as well as the ranking: a score of about a third leaves plenty of room for a long, multi-app job to go wrong when a model runs it alone. What works is the shape that already works for people: small steps, each with a clear input and output, and a person signing off on anything that leaves the building.

That is how Workflows are built in overads. A schedule or a number crossing your line starts the run, a few steps do the work, and an approval sits before anything publishes until you decide a workflow has earned the right to post its own drafts. The templates are a quick way to see the pattern, and if you are weighing a general-purpose automation tool instead, we compared the two in overads vs n8n.

Fewer made-up facts, same rule about numbers

OpenAI says Sol makes about half as many factual mistakes as its predecessor on an internal test built from real conversations where users had flagged an error, and that Luna at higher effort matches GPT-5.6 Sol at about a hundredth of the cost. For marketers that should mean fewer invented statistics in drafts, fewer wrong feature names and fewer confident claims about a competitor's pricing.

It does not mean the model knows your account. Spend, return on ad spend, reach and lead counts have to come from the platform that recorded them. Let the model write the sentence and let the source supply the number. That is the rule behind the ads manager, which reads spend and results straight from Meta, Google and Snapchat, and behind competitor tracking, where a figure we could not observe shows as missing rather than as a guess.

Your buyers read these answers too

Accuracy gains also reach the answers your buyers get when they ask an assistant which product to choose. Our read is that a model less willing to invent leans more on what it can look up, which favours brands whose facts are stated plainly and the same way everywhere. We covered what gets a brand cited in how to rank in ChatGPT and Perplexity SEO. Signals is built to ask Gemini and Perplexity how they describe your brand, but its AI-answer scan is switched off on the live product while we finish it, and ChatGPT is not covered.

Use Sol or Luna with your overads channels

To be clear about where we stand: overads does not run on GPT-6. What you can do today is connect the two. Both models are in Codex, and Codex reads MCP servers from its config file. Add the overads MCP server with an API key and the model can list your connected channels, check your best times to post, pull a metrics summary, draft and schedule posts, and run one of your saved workflows. Setup is one block of config, and every tool is listed on the MCP tools page. The MCP server and API keys are included on paid overads plans.

A loop that suits Luna well: ask for last week's summary, have it draft three posts that build on what performed, and save them for approval at your best times. Posts can be saved as drafts, or the model can set requireApproval on each post, the same option the composer shows as "Needs approval first", so it waits for a teammate to approve it. Holding every scheduled post in a workspace for approval is a setting overads turns on for you. Our scheduling and approvals guide covers the rules. Posts go out through Publish to X, LinkedIn, Bluesky, Mastodon and Pinterest.

If you build your own agent on the API, the caching changes are worth using. Put the parts that never change first, such as the brand voice guide, product facts and channel rules, then the day's request. OpenAI says cached reads are 90% cheaper and that changing the effort setting or the available tools no longer throws the cache away.

What to try this week

• Move one high-volume job, such as comment tagging or hook variants, to Luna and compare cost and quality for a week.
• Give one multi-step job, such as the weekly client summary, to Sol with an approval step before anything is sent.
• Write your brand facts down once and reuse them as the fixed start of every prompt.
• Ask assistants how they describe you, before and after you tidy those facts. Signals is built to put those questions to Gemini and Perplexity for you, but its AI-answer scan is switched off on the live product while we finish it, and ChatGPT is not covered.