OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, 2026, nineteen days after Astra. The headline was price. Sol costs a fifth of what Astra does. Luna costs a hundredth.
The more useful line was in OpenAI’s own chart. On AutomationBench, a test of business workflows across apps, Sol at its extra high effort setting scored 33.2%. Astra at low effort scored 30.3%, at 3.9 times the cost per task.
For anyone doing SEO with an agent, that changes the question. It is no longer which model is smartest. It is which model, at which effort, for which job. Most SEO work now belongs on Sol. The bulk work belongs on Luna. Astra keeps the calls you cannot afford to get wrong.
None of the three can see a search volume, though. ContextBolt SEO is the live data they are missing. It pulls keyword volumes, the current top ten, backlinks, audits and your own Search Console numbers into Codex over MCP. The model does the thinking. The server brings today’s facts.
- Sol is the new default for SEO work. OpenAI built it for complex, agentic work. It costs $2 per million input tokens and $10 per million output, a fifth of Astra’s price.
- Luna is for volume. At $0.10 and $0.50 per million tokens, it is the model for clustering, tagging and metadata runs across thousands of rows.
- Astra keeps the hard calls. A traffic drop investigation, a migration plan, a strategy that sets the next quarter.
- Effort matters as much as the model. Sol on extra high beat Astra on low in OpenAI’s workflow test. Luna on max beat GPT-5.6 Sol on medium in computer use, at a tenth of the cost.
- None of them has live SEO data. Sol’s knowledge stops at April 20, 2026 and Luna’s at May 18. ContextBolt SEO connects keyword, SERP and Search Console data over MCP, free for 7 days and then $35 a month.
What are GPT-6 Sol and Luna?
GPT-6 Sol and GPT-6 Luna are the two cheaper models in OpenAI’s GPT-6 family. OpenAI trained them “with similar methods as GPT-6 Astra” and released both on September 22, 2026. They take over from GPT-5.6 Sol and GPT-5.6 Luna at half those models’ promotional prices, per OpenAI’s announcement.
The two have different jobs. OpenAI describes Sol as “built to power complex coding and agentic workflows.” It calls Luna its “most efficient model for focused, high-volume tasks.”
On paper they share a lot. Both have a 1,050,000-token context window, take up to 922,000 tokens of input and return up to 128,000. Both accept text and images. Both support the tools an SEO workflow leans on, including web search, file search, code interpreter, computer use and MCP. Reasoning effort runs from none to max on each.
The real differences are price and knowledge cutoff. Sol’s training data stops at April 20, 2026. Luna’s stops at May 18.
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
Prices are per million tokens, from OpenAI’s model pages as of September 23, 2026. Cached input costs a tenth of the input price on all three, so $1, $0.20 and $0.01. Prompts over 272,000 input tokens cost twice the input rate and 1.5 times the output rate. Batch and Flex cost half.
Where you can use them
In ChatGPT, Sol and Luna are in Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users get Luna in the desktop app. OpenAI’s announcement says plainly that “these models are not yet available in Chat.” In the API they are gpt-6-sol and gpt-6-luna.
That last detail decides the setup. If you want Sol working on live SEO data today, Codex is the simplest place to do it.
What do OpenAI’s benchmarks say?
There is still no benchmark called SEO. What OpenAI published is a set of workflow and agent tests, and a few of them sit close to how SEO work actually runs.
AutomationBench is the closest. It tests agents on end-to-end business workflows using 47 tools across sales, marketing, operations, support, finance and HR. OpenAI put the cost per task next to each score, which is the column worth reading.
| Model and effort | Score | Cost per task |
|---|---|---|
| GPT-6 Sol, xhigh | 33.2% | $0.27 |
| Claude Fable 5.1, max | 31.4% | Over 8.9x Sol |
| GPT-6 Astra, low | 30.3% | 3.9x Sol |
| Claude Opus 5, max | 26.9% | 11.1x Sol |
OpenAI’s Fable row counts Opus 5 as a fallback model, and OpenAI notes that its cost understates the real figure because it leaves out those fallbacks, which ran on about 40% of tasks.
Three more results are worth knowing.
- Agents’ Last Exam: Sol at max effort scored 56.4%, above Claude Opus 5’s best score in that evaluation, at 60% lower cost per task.
- Factuality: on OpenAI’s internal test, built from conversations where users flagged a mistake, Sol makes about half as many mistakes as GPT-5.6 Sol. Luna at higher effort matches GPT-5.6 Sol at about a hundredth of the cost.
- Computer use: on OSWorld 2.0, Sol at extra high scored 60.5% against 60.3% for Claude Opus 5 at medium, for about 80% less per task. Luna on max beat GPT-5.6 Sol on medium at a tenth of its cost.
Take all of it with a grain of salt. These are OpenAI’s own evaluations. The rival scores come from public reports, and OpenAI used Claude Fable 5 scores where Fable 5.1 scores were missing. None of it measures whether a model picks the right keyword. The Claude side of the same question, with Anthropic’s own numbers, is in Claude Fable 5.1 and Opus 5.5 for SEO.
What the chart does show is the thing to take from this launch. An effort setting moved a cheaper model past the flagship. Which model you pick matters less than it did last month. How much thinking you pay for matters more.
Which SEO jobs go to which model
Here is how I would split the work. The models follow OpenAI’s description of each one. OpenAI’s own starting efforts are medium for Sol, high for Luna and light for Astra, and the table goes lower on jobs so well-scoped that a sample tells you in a minute whether the output holds.
| SEO job | Model | Effort |
|---|---|---|
| Keyword research and picking targets | Sol | Medium to high |
| SERP and competitor analysis | Sol | High |
| Content briefs and page rewrites | Sol | Medium |
| Audit triage and fixes in the repo | Sol | High to extra high |
| A traffic drop that spans data and code | Astra | High |
| Site strategy or a migration plan | Astra | High to max |
| Clustering thousands of keywords | Luna | Low |
| Title and meta variants at scale | Luna | None to low |
| Tagging a crawl export by page type | Luna | None to low |
Sol does the daily work: Keyword research, SERP reads, briefs and audit fixes are judgment jobs, but they repeat. You run them every week. Sol is cheap enough to iterate on and strong enough that the answer is worth reading. Start at medium, the default, and raise it when the job needs more planning.
Luna does the volume: Clustering 5,000 keywords is not a hard problem. It is a long one. So is writing title variants for every product page. Luna is built for exactly this, and at its price the model stops being the expensive part of the job.
Astra keeps the expensive mistakes: A traffic drop that crosses Search Console, a competitor’s new page and last Tuesday’s deploy is where a model that holds the whole thread pays for itself. So is anything that sets direction for the next quarter. GPT-6 Astra for SEO covers what the flagship can do and when its price is worth paying.
What a bulk job costs on each model
Here is the arithmetic on OpenAI’s list prices. Say you want new title tags for 500 product pages. Each page sends about 1,500 tokens of text and instructions, and each title comes back in about 30. That is 750,000 input tokens and 15,000 output tokens.
- Luna, no reasoning: about 8 cents.
- Sol, no reasoning: about $1.65.
- Astra: about $8.25 before reasoning tokens, which Astra always spends because its lowest effort setting is low.
At the same token counts, Sol always costs 20 times Luna and Astra always costs 100 times Luna. Only reasoning changes the ratio, because reasoning tokens are billed as output.
One more lever if you build on the API. OpenAI says GPT-6 caching now hits more often by default, and cached input costs 90% less. Put the parts that never change first, like your site context, rules and examples, and the changing page last. Calls that share that opening then read it from the cache at a tenth of the price.
The model is still not the SEO data
A cheaper model does not change this part.
Sol’s knowledge stops at April 20, 2026. Luna’s stops at May 18. Search results change every day. Search volume, difficulty and backlinks come from specialist indexes, and your clicks and impressions live in your own Search Console.
Web search helps, and both models have it. It reads today’s pages. It does not turn Google’s results into a keyword database, because Google does not print a monthly volume next to each result.
So the failure is the old one. Ask any model for search volume without a data source and you get a number shaped like 2,400 that came from nowhere. Why ChatGPT invents keyword numbers walks through how to spot one. A smarter, cheaper model makes a better-sounding guess. It is still a guess.
ContextBolt SEO is ours, and it exists for this gap. It is a hosted MCP server. You start the trial, add one URL to Codex, sign in once with the same email, and ask in plain English for keyword research, difficulty, the live top ten, competitor rankings, backlinks, site audits, AI visibility checks or your own Search Console data. The answer comes back inside the session, where Sol can act on it.
The volumes are estimates from a large keyword index, close enough to decide on, and the Search Console numbers are your own. Every lookup saves automatically to your SEO Dashboard, so the evidence outlives the session. The trial gives you 100 credits for 7 days. After that it is $35 a month for 1,000 credits. Research calls cost one credit, and backlink and AI visibility checks cost more.
How to run SEO research with Sol in Codex
Codex is the simplest place to pair Sol with live data today. Start the 7-day free trial first, because sign-in only works for an email with an active trial or subscription. Then it is two commands.
codex mcp add contextbolt-seo --url https://seo.contextbolt.app/mcp
codex mcp login contextbolt-seoThe first saves the server to Codex’s configuration. The second opens a browser, where you sign in with the email you used for the trial. SEO inside Codex covers the config file and what to do if the tools do not show up.
Then start Codex on Sol.
codex -m gpt-6-solInside a session, /model switches the model or its reasoning effort without a restart. OpenAI’s Codex model guide lists all three GPT-6 models and gives a starting effort for each. It starts Sol at medium, Luna at high, and Astra at light, which is low in the CLI. Raise it when a task needs more planning or checking. If Sol is missing from your list, OpenAI says availability depends on the rollout, your sign-in method and your client.
Copy this prompt
Use ContextBolt SEO to find one content opportunity for my site.
Market: United States
Goal: reach people who do their own SEO inside an AI agent
Seed: ai seo agent
Run one keyword expansion. Shortlist the 5 terms that best fit
the goal. Pull volume, difficulty and intent for those 5, then
the live top 10 for the best 2. Tell me which one my site can
realistically win and why. Stop before drafting.The stop is deliberate. Sol is cheap, but a 2,000-word draft on a keyword you cannot win still costs you an afternoon.
When the list is settled, hand the long part to Luna. Type /model and pick GPT-6 Luna. OpenAI’s starting point for Luna is high. A clustering pass is a familiar, well-scoped job, so the guide’s own advice applies. Try a lower setting on a sample, and step back up if the groups come out wrong.
Then hand it to Luna
Take the full keyword list from the research above. Group it by
search intent into clusters of closely related terms. For each
cluster, name the page that should target it, and flag any term
that could fit two clusters. Return one table.The keyword clustering walkthrough shows the same job step by step, and the method does not depend on the model.
Where Sol and Luna fall short for SEO
Chat access is not there yet: Sol and Luna run in ChatGPT Work and Codex, not in Chat. If you do your SEO in a chat window, SEO inside ChatGPT covers the connector for the models Chat has today.
The benchmarks are not SEO tests: A model can top AutomationBench and still chase a keyword your site cannot win. Run a small test of your own. Give two models the same real data and score whether they rejected terms you cannot win and stayed in your market. Cost per usable answer is the number to compare.
Luna is for work you can check at a glance: A clumsy title or a wrong cluster is quick to spot and cheap to fix. A wrong strategy is neither. Keep Luna on jobs where a mistake shows up on the page, and keep the judgment calls on Sol or Astra.
Cheap output makes the old trap cheaper: This is the uncomfortable one. Search feedback takes months. A model that drafts a page for a fraction of a cent makes it easy to publish 500 pages before the first one proves the plan was wrong. Sol costs a fifth of Astra. A wrong page costs what it always did.
So use the cheap models for the work that repeats, and put the savings into data you can trust. ContextBolt SEO connects to Codex in two commands, starts with a 7-day free trial, and saves every lookup to your SEO Dashboard. The research will still be there when the next model ships, which at this pace is about three weeks away.
