GPT-6 Sol: what it means for your marketing

GPT-6 Sol is OpenAI’s middle GPT-6 model, released on 22 September 2026. In the API it costs $2 per million input tokens and $10 per million output tokens, half GPT-5.6 Sol’s promotional rate. In ChatGPT it is in Work and Codex on paid plans, but not yet in Chat. Test it on source-led drafts, reports and supervised work across tools, and judge it on the cost of approved work. Prices and availability were checked on 24 September 2026.

GPT-6 family update (checked 2 October 2026)

On 29 September 2026 OpenAI released GPT-6.1 Sol (gpt-6.1-sol), an upgrade to GPT-6 Sol at the same $2 input and $10 output per million tokens, with cached input cut to $0.10. On the Artificial Analysis Intelligence Index v4.3.2 it scores 52 at max effort against 48 for GPT-6 Sol, and at medium effort it matches GPT-6 Sol’s best score for about a fifth of the cost per task. The prices and figures below are from GPT-6 Sol’s launch. For the newer model, see GPT-6.1 Sol: near-Astra results at a fifth of the price.

Correction (7 October 2026): an earlier version of this post said EU data residency for GPT-6 Sol was available only with Standard processing. OpenAI’s Sol model page and data controls guide say it is available with Standard, Flex and Batch processing. Fast mode is not available with EU data residency.

The answer in brief

• What it is: Middle GPT-6 model for complex coding and agentic workflows; Astra is the flagship, Luna is the high-volume tier. OpenAI released GPT-6.1 Sol as its upgrade on 29 September 2026, at the same token prices.
• API list price (short context): $2 input / $0.20 cached input / $2.50 cache writes / $10 output per million tokens.
• Where in ChatGPT: Work and Codex on Plus, Pro, Business, Enterprise and Edu (gradual rollout). Not ordinary Chat.
• Fast mode: API Fast is 2× Standard; Work/Codex Fast for GPT-6 Astra/Sol/Luna is 2.5× Standard. They are not the same.
• How to judge it: Cost per approved deliverable (model + review + rework), not token rate alone.

What is GPT-6 Sol, and where can you use it?

GPT-6 Sol is OpenAI’s middle model in the GPT-6 family: between flagship GPT-6 Astra and efficient GPT-6 Luna, aimed at complex coding and agentic workflows.

OpenAI’s Sol model page says it is “built to power complex coding and agentic workflows”. OpenAI calls Astra “our best model across the board”. Luna is, in OpenAI’s API documentation, its “most efficient model for focused, high-volume tasks”. That makes Sol a candidate for the middle of a marketing production process: enough capacity to reason through a brief or inspect a report, at about a fifth of Astra’s list price.

The name can mislead. In the GPT-5.6 family, OpenAI called Sol “our flagship model”, as we covered in our GPT-5.6 Sol guide. In the GPT-6 family, Astra is the flagship and Sol is the middle model. So the new Sol is cheaper than the old Sol partly because it now sits a tier lower, not only because list rates moved.

GPT-5.6 also had a middle model called Terra. OpenAI’s 22 September announcement does not mention a GPT-6 Terra. It names Sol and Luna as the new models, alongside Astra. GPT-5.6 Terra remains on the ChatGPT Work/Codex rate card; there is still no GPT-6 Terra listed there (checked 24 September 2026).

Access depends on the product you use. OpenAI’s launch announcement says Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with a gradual rollout. It also says the models “are not yet available in Chat”. Free and Go users can use Luna in the desktop app, but not Sol. Enterprise admins may need to enable the model.

These are separate routes. A ChatGPT subscription does not mean your own application or automation uses the API model. If the Sol option is missing from your usual chat window, that matches OpenAI’s launch statement. Look in Work and Codex instead. Our ChatGPT Work guide explains how Work differs from Chat.

How much does GPT-6 Sol cost?

GPT-6 Sol’s standard short-context API price is $2 input and $10 output per million tokens; Work and Codex use matching token rates but without the API’s separate cache-write line, and Fast mode is priced differently on each surface.

Standard API prices (short context)

OpenAI’s Sol model page lists these rates per million tokens:

OpenAI’s announcement compares this with $4 and $20 for GPT-5.6 Sol, and describes that older rate as promotional pricing. The halving is real, but it is measured against a promotional price, not every rate still quoted in an old proposal. The ChatGPT rate card still notes GPT-5.6 Sol’s promotional pricing as available at least through 21 November 2026.

Work and Codex token rates (not full API parity)

On the ChatGPT Enterprise token-based rate card, GPT-6 Sol in Work and Codex is listed at $2 input / $0.20 cached input / $10 output per million tokens. Those figures mirror the API’s main token rates, but the Work/Codex card does not show a separate cache-write line like the API ($2.50). Do not assume every API pricing rule appears identically in ChatGPT billing.

Conditions that change the bill

The same OpenAI model page and rate card say:

Fast is not the same multiplier everywhere. On the API, Fast mode costs twice the applicable rate. On ChatGPT Work and Codex, the ChatGPT Enterprise rate card lists Fast for GPT-6 Sol (and Astra and Luna) at 2.5× the Standard rate. Do not budget Work Fast as if it were the API Fast multiplier.

If a team pastes a brand book, a year of exports and a style guide into every request, it can pass the 272,000-token line and lose much of the saving. Keep the stable part of the prompt short, and cache it. OpenAI says GPT-6 has improved prompt caching with “higher cache hit rates by default” and a 90% discount on cached input reads. The real benefit depends on how your application sends prompts.

For an agency, the better measure is cost per approved deliverable. Count model and tool charges, staff review, rework and failed runs. A landing page that passes factual, brand and technical checks at first review costs less than one that needs three rewrites. A lower token price alone does not justify a cheaper retainer or a promise of faster delivery.

What does a typical marketing job cost with GPT-6 Sol?

Very little in model charges for ordinary marketing jobs; the expensive part is usually human review. The table uses OpenAI’s standard Sol and Luna rates. The token counts are illustrative examples, not measurements. Check your own usage dashboard.

The last row shows the long-prompt rule at work. At 300,000 input tokens, the whole request is charged at 2× input and 1.5× output. Trim the same job to 250,000 input tokens and it costs about $0.55.

Caching changes the picture for repeated work. Suppose ten variant runs each reuse the same 30,000-token brand pack. Without caching, that pack costs $0.60 across the ten runs. With one cache write and nine cache reads, it costs about $0.13, if every read hits the cache.

Effort settings add output tokens. Artificial Analysis measured about 31,000 output tokens per task for Sol at maximum effort in its own tests, published on 22 September 2026. At the standard rate, that output alone costs about $0.31.

The lesson for budgets is simple. The model charge for most marketing jobs is pennies. The real cost is the time a person spends checking and fixing the result. That is why cost per approved deliverable matters more than the token rate.

Is GPT-6 Sol more accurate than earlier models?

Vendor and independent tests point to fewer wrong answers when Sol declines more often, not a blanket “half the errors in your client copy”.

OpenAI says Sol “makes about half as many mistakes as its predecessor” on an internal factuality test. Read the test conditions before you repeat the claim. OpenAI built the test from de-identified ChatGPT conversations in which users had flagged mistakes. The announcement says these conversations “are not representative of typical usage”, and that scores are not controlled for length.

Independent testing points the same way. Artificial Analysis reported on 22 September 2026 that Sol at maximum effort cut its hallucination rate on the AA-Omniscience test from 92% to 60%, compared with GPT-5.6 Sol. Read how it got there. Sol declined to answer more often: it attempted 83% of questions, against 99% for GPT-5.6 Sol. That cut wrong answers by about a quarter, but overall accuracy fell from 59% to 54%.

For marketing work, that trade-off is useful. A model that says “I don’t know” is easier to manage than one that invents a statistic. It still means Sol will leave gaps, and a person must fill them from real sources.

The knowledge cutoff also matters. OpenAI lists Sol’s cutoff as 20 April 2026. Prices, people, product names and results after that date must come from material you supply or a search tool. A person who knows the subject should then read the result.

OpenAI also reports that Sol and Luna make fewer misleading claims about their coding work than their GPT-5.6 versions. OpenAI says these alignment tests use challenging situations and “do not measure failure rates in typical use”. It publishes the detail in the GPT-6 Astra system card, which gained a Sol and Luna appendix on 22 September 2026.

How does GPT-6 Sol compare with Claude Opus 5.5 and Grok 4.7?

No launch chart is a direct head-to-head of GPT-6 Sol against Claude Opus 5.5 or Grok 4.7. Check which model versions each vendor actually compared.

• OpenAI’s charts compare Sol mostly with Claude Opus 5 and Claude Fable 5 or 5.1, not with Claude Opus 5.5. For example, OpenAI reports Sol at 60.5% on OSWorld 2.0 offline at xhigh effort, against 60.3% for Claude Opus 5 at medium effort (vendor chart: treat it as OpenAI’s own comparison).
• Anthropic’s Claude Opus 5.5 announcement, also on 22 September 2026, compares Opus 5.5 with GPT-6 Astra and GPT-5.6 Sol, not with GPT-6 Sol. Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens.
• SpaceXAI’s Grok 4.7 announcement on 21 September 2026 compares Grok 4.7 with older models, including GPT-5.6 Sol and Opus 5. SpaceXAI lists Grok 4.7 at $2 input and $6 output per million tokens.

Independent tests are more useful, because they complicate the sales pitch. At launch on 22 September 2026, Artificial Analysis found that Sol at maximum effort cost $1.06 per task to run its Intelligence Index, about 50% less than GPT-5.6 Sol at $1.99. Sol scored 57 on its Coding Agent Index, up two points. But on its GDPval-AA v2.1 knowledge-work test, Sol dropped about 100 Elo points. Artificial Analysis linked the regressions to weaker presentation and “deliverables that omit rubric elements”.

That last finding matters to a marketing manager. A fluent campaign plan can still leave out the one requirement the client cares about.

Zapier’s AutomationBench (version 1.0.6) tests agents across 47 tools in simulated business systems. It grades the final state of those systems, not the quality of the write-up. Zapier’s live leaderboard showed these rows when we checked it on 24 September 2026:

Zapier’s Opus 5.5 rows use default fallbacks, which route a refused step to the provider’s own fallback model. Anthropic’s launch page reported 40.0% for Opus 5.5 at max effort on runs without fallback models, where safeguard interventions counted as failures. They are two different set-ups, so quote the one that matches how you would run the model.

These are Zapier’s results for its own test setup. They are not a forecast of agency productivity. Even the leading model failed most tasks. So connected agents are worth testing, but they still need limited permissions, checks and human approval before anything goes live.

Note the effort settings too. Sol’s default reasoning effort is medium, according to OpenAI. Most of the published scores use xhigh or max, which a busy team may never select.

Artificial Analysis also tested Grok 4.7 and found gains over Grok 4.6. It also found that Grok 4.7 at xhigh used about 81,000 output tokens per Intelligence Index task, against 36,000 for Grok 4.6 (Artificial Analysis figures, last checked 23 September 2026). A low token rate can still produce an expensive task.

Where does Sol fit in AI implementation?

Sol belongs after you know which process repeats, costs a lot to review, or causes avoidable errors, not before you have named that process.

Take a monthly marketing report. A useful workflow collects approved data, finds the changes worth explaining, drafts a plain-English commentary and flags figures that do not reconcile. Sol may suit the analysis and first draft. A person still checks the data source, attribution, commercial interpretation and wording before the client sees it.

Campaign production follows the same pattern. Give the model an approved offer, audience, exclusions, landing page and brand examples. Ask for a small set of ad and email variants, and tie each claim to the supplied material. Review the variants against platform, legal and client requirements. Then move them into the publishing system. The model prepares the work. The team decides what to send or spend.

OpenAI’s Sol model page lists support for web search, file search, image generation, code interpreter, computer use and MCP, and tells developers to use the Responses API for built-in tools. These features connect a model to information or actions. They do not make a given integration secure, reliable or suitable for a particular client. Define what the model can read, what it can change, and who approves the change.

OpenAI lists fine-tuning as not supported for Sol. If a project needs a tuned model, choose a model that supports tuning, or design the workflow around prompts, retrieval and examples instead.

AIWIZ’s AI adoption services cover AI strategy, workflow automation, predictive analytics, chatbot development, custom models and fine-tuning, and platform and API integration. A good first brief is specific. Name one process, its inputs and outputs, who signs it off, and what a good result looks like. That gives a project something to test, which a general wish to “use AI” does not.

Can GPT-6 Sol help with SEO and AI search visibility?

It can help with the research and drafting work. It cannot guarantee rankings or AI citations on its own.

There are two separate jobs here. One is using Sol to produce and check search content. The other is making a business visible in search and in AI-generated answers. A new model does not do the second job for you.

For SEO, Sol can organise research notes, compare page intent, find unanswered customer questions, draft an outline and check a page against its brief. It can also review title options, internal-link opportunities and inconsistencies between pages. A search specialist still chooses the target query, checks demand in a keyword tool, studies the competing results and decides what the business can say with authority.

Volume is the trap. Google’s guidance on generative AI content says that using AI “to generate many pages without adding value for users” may break its spam policy on scaled content abuse. A cheaper model makes that mistake cheaper to make, not safer.

For answer engine optimisation (AEO) and related AI-search work, Sol can help with prompt research, page audits, source gathering and clearer answers to buying questions. Start with your customers’ real questions and the evidence you can publish: service detail, prices where appropriate, limits, credentials, examples and policies. The model can help structure that material. It cannot make a third-party system recommend a business.

Google’s guidance on AI features in Search says standard SEO best practices “remain relevant” for AI Overviews and AI Mode. It says there are “no additional requirements” and no special schema.org markup to add, and that indexing and serving are not guaranteed. Treat any promise of guaranteed AI visibility with care. For wider context, see our guides to GEO and AI search visibility.

How can GPT-6 Sol help with paid media, social, PR and creative work?

Sol is most useful for preparation and supervised review in PPC, social and PR, not for unsupervised publish or spend decisions.

In PPC, Sol can turn an approved offer into tightly specified ad variants, compare a landing page with its ad, and draft a change log for the account manager. Ask it to flag unsupported claims and missing exclusions. A person decides the budget, targeting, final copy and launch. The token bill tells you nothing about media efficiency.

For social media, Sol can adapt one approved idea for different formats, prepare response guidance and spot where a draft drifts from the brand voice. Give it real examples of the organisation’s writing. OpenAI says the new models give answers with “more clarity, less jargon” and fewer odd phrases. That is OpenAI’s assessment, not proof that Sol can match a particular client’s voice. A named editor still decides whether a post sounds like the business.

PR punishes a plausible mistake. Sol can organise source material, test whether a release answers the obvious journalist questions and prepare interview notes. Check every name, date, quote, statistic, partnership and commitment against the original record. Never let a model invent a spokesperson’s view because the paragraph reads well.

The same limit applies to graphic design and motion graphics. Sol can help write a creative brief, organise feedback, or check that captions and on-screen claims match the approved copy. OpenAI’s model page lists text and image input and text output, with no audio or video support. Image generation is available as a separate API tool. Designers and editors still make the visual decisions.

How should you use GPT-6 Sol for web development and connected agents?

Paid ChatGPT users can select Sol in Codex for coding help; connected CMS or ad-platform agents still need narrow permissions and human approval.

GitHub also added Sol to Copilot for Pro+, Max, Business and Enterprise plans, on a gradual rollout. In either tool, Sol can help inspect a WordPress issue, draft a small change, compare a landing page with its approved design and prepare a test checklist. OpenAI reports better coding and computer-use results than GPT-5.6 Sol. Artificial Analysis measured a two-point gain on its Coding Agent Index. Those results justify a trial on representative tasks, followed by normal code review and testing.

An agent connected to a CMS, analytics account or ad platform needs a narrower brief than “fix the campaign”. Decide which records it may read, which draft changes it may prepare and who approves publication. Check the result in the live interface. A well-written draft is not finished if the form, tracking, mobile layout or published page is wrong.

GPT-6 Sol vs Astra and Luna: which should you use?

Use Sol first for substantial but repeatable work; escalate hard failures to Astra; push simple, checkable volume to Luna.

OpenAI lists these standard API prices per million tokens (short context; cache-write rates shown where the API publishes them):

Use Sol first for substantial but repeatable work: source-led drafts, report commentary, campaign variants, coding help and supervised work across tools. Test it on a real brief with a fixed review standard.

Use Astra when the task is complex, long-running or high-stakes, or when Sol keeps missing requirements that Astra meets. Astra costs five times as much as Sol per token on the list rates above. On Zapier’s benchmark it scored higher than Sol at a higher cost per task. Route the difficult exceptions to it, not every headline variation. See our GPT-6 Astra guide.

Use Luna for simple, high-volume steps with an answer you can check, such as extraction, tagging and short summaries. It costs a twentieth of Sol’s rate. Keep Sol or a person on synthesis and judgement.

Test other suppliers in the same way. Give Claude Opus 5.5, Grok 4.7 or any other current model the same brief, source material and review criteria as Sol. No launch chart tells you which model writes the best brief in your organisation’s voice.

What should a marketing team test this week?

Start by confirming Sol appears in Work or Codex (not Chat), then run one recurring UK agency job against a written pass checklist.

1. Check your access first. Look for Sol in ChatGPT Work and Codex, not ordinary Chat. Note which paid users can select it. Enterprise admins: confirm the model is enabled. Free/Go: Sol is not available; Luna is desktop only.
2. Choose one recurring job. Examples that UK agencies already run: a monthly Google Ads / Meta performance commentary from exported CSVs; a WordPress service-page refresh against an approved brief; a HubSpot email sequence draft from a fixed offer sheet; a GA4 change log for the account manager.
3. Collect the approved inputs only. Include source links, brand examples, exclusions and the owner of each claim. Remove material the model does not need, especially anything that pushes you over 272,000 input tokens.
4. Write a pass checklist. Cover factual accuracy, missing requirements, brand voice, accessibility or technical issues, platform policy (Ads/Meta), and named final approval.
5. Run the same brief through Sol and your current method. Start at the default medium effort, then try a higher one if the draft is thin. Record model charges, review time and rework. Compare approved results, not first drafts.
6. Watch Fast and long-context separately. If someone enables Fast in Work/Codex, remember the 2.5× multiplier. If a request exceeds 272,000 input tokens, OpenAI charges the higher long-context rate on the whole request.
7. Add one controlled connection only if it helps. A read-only data source or a draft destination (for example a staging WordPress draft, not live publish) teaches you more than an agent with permission to publish or spend.
8. Decide the next route. Keep Sol for the tasks it passes. Move simple steps to Luna. Test Astra or another model on the failures that remain.

GPT-6 Sol gives businesses a credible, lower-cost option for useful AI work. The result still depends on the workflow, the quality of the inputs and the standard of the final review. AIWIZ’s AI adoption services cover AI strategy, workflow automation and integration. If you want to test where Sol belongs in your marketing production, talk to AIWIZ. Bring one process, and we will help you design a trial that measures approved work before you scale it.