On July 9, OpenAI launched GPT-5.6 and ChatGPT Work, and most marketing coverage focused on the wrong thing. The headlines talked about benchmarks and model names. Sol, Terra, Luna. Who cares.
The real story is that ChatGPT Work is an agent that produces finished marketing deliverables. Not drafts. Not suggestions. Finished sheets, slides, docs, and web apps. It pulls context from 1,400+ connected plugins, stays with projects for hours, and ships artifacts you can actually use.
Five million people already use it weekly. One million use it for work outside software development. Those numbers came from OpenAI's own launch, and even discounted for marketing spin, they represent a shift that marketing teams need to understand before the invoice arrives.
The trap isn't the technology. It's the pricing model underneath it.
What ChatGPT Work Actually Replaces
For the last two years, AI in marketing meant a smart assistant that helped you write faster. You still had to take the output, clean it up, put it in a template, format it, and ship it. The AI was a drafting tool. The human was the production layer.
ChatGPT Work removes the human from the production layer. You give it a goal, it gathers context from your connected apps, plans the approach, executes it, and returns a finished deliverable. A competitor analysis deck. A content calendar spreadsheet. A campaign brief document. A landing page.
This is what junior marketers spent their first two years doing. The grunt work of taking someone's strategy and turning it into artifacts. That work is now something an agent can handle for the cost of a few thousand tokens.
If you've been following the entry-level marketing squeeze, this is the next chapter. The last wave of AI threatened entry-level jobs. This wave actually has the product to back it up.
The Token Math Nobody Is Running
Here's where it gets expensive. ChatGPT Work runs on GPT-5.6, which ships in three tiers with dramatically different pricing.
Sol costs $5 per million input tokens and $30 per million output tokens. Terra costs $2.50 input and $15 output. Luna costs $1 input and $6 output. That's a 5x spread between the top and bottom tier.
A marketing team running ChatGPT Work for daily operations might process hundreds of tasks per day. Competitor analyses. Content briefs. Social media calendars. Email sequences. Campaign reports. Each task consumes tokens. Each token costs money.
The problem is that most marketing teams will default to whatever tier ChatGPT Work picks for them. They won't know whether a competitor analysis ran on Sol (expensive, high reasoning) or Luna (cheap, fast). They'll see a finished deliverable and assume it was cost-effective.
It probably wasn't.
According to benchmark analysis from Vellum, the quality gap between Sol and Luna is about 3.3 percentage points on standard benchmarks. That sounds small. But in marketing, where a hallucinated competitor name or a fabricated statistic can create real liability, 3 percentage points is the difference between usable and dangerous.
The companies that already struggled with multi-model fragmentation now face a sharper version of the same problem. It's not just "which model do we use." It's "which model tier does our agent use for each specific task, and are we measuring the cost difference?"
Most won't. Most will get a monthly bill, shrug, and pay it. Until the bill gets big enough to matter.
The Usage Trap
ChatGPT Work doesn't price like a software license. It prices like an API. The more you use it, the more you pay. Longer, more complex tasks consume more of your plan's included usage, and once you exceed the included amount, you're paying per task.
This is the part that should scare marketing leaders. Traditional tools (HubSpot, Marketo, Adobe Creative Cloud) charge per seat. You know your monthly cost. It's predictable. You can budget for it.
ChatGPT Work is metered. A quiet week costs almost nothing. A launch week where your team runs 200 competitor analyses, 50 content briefs, and 30 campaign reports could cost more than a full-time junior marketer's salary.
Nobody is building dashboards for this yet. Nobody is tracking per-task token consumption. Nobody is asking "did this competitor analysis need to run on Sol, or would Luna have been fine?" The usage data exists, but it's buried in OpenAI's interface, not in your marketing operations stack.
This is the same cost efficiency illusion that's been haunting AI marketing spend all year. Teams see cheaper per-task costs and assume they're saving money. They're not tracking the cumulative spend, the quality variance, or the hidden cost of re-doing work that came back wrong because it ran on the wrong tier.
The Quality Question Nobody Can Answer
Here's the uncomfortable part. ChatGPT Work produces finished deliverables, but nobody has a reliable way to measure whether those deliverables are good.
A junior marketer's work goes through review. A manager looks at it, gives feedback, and sends it back for revision. The quality loop is built into the org chart.
ChatGPT Work's output goes... where? To the person who requested it. Maybe they review it. Maybe they're too busy and just ship it. Maybe it looks polished enough that nobody questions whether the competitor data is accurate or the campaign logic is sound.
The polished output is the problem. AI-generated deliverables look professional. They have formatting, structure, and confident language. But looking good and being good are different things. A competitor analysis with one fabricated market share number can send a strategy team in the wrong direction for months.
OpenAI says their testing blocked 100% of extraction attempts and that auto-review and red teaming are built in. That's about security, not about marketing accuracy. The model won't leak your data. It might still confidently tell you a competitor's revenue is $2.3 billion when it's actually $1.8 billion.
The agentic ROI theater problem applies here. Teams will point to ChatGPT Work deliverables as evidence of productivity gains. Look at all these decks we produced. Look at all these reports. But if the underlying data is wrong, the productivity is negative. You're producing more bad work, faster.
What Agencies Should Actually Do
If you run a marketing agency or lead a marketing team, three things matter right now.
First, start tracking token costs per deliverable type. If ChatGPT Work runs a competitor analysis, log the tokens consumed and the tier used. Do the same for content briefs, campaign reports, and social calendars. After a month, you'll see which tasks are expensive and which are cheap. Then you can make routing decisions.
Second, build a review layer for AI-generated deliverables before they reach clients. Not a casual "looks good" check. A structured review that verifies claims, checks competitor data, and validates strategic logic. The cost of this review is your insurance policy against the quality cliff.
Third, figure out which deliverables ChatGPT Work genuinely does better and cheaper than humans. Then stop doing those manually. The savings from automation should fund the review layer. If you're not netting positive after both costs, you're doing it wrong.
The teams that win will be the ones who treat ChatGPT Work as a production tool with measurable costs and quality controls. The teams that lose will be the ones who treat it as free labor and discover the bill six months later.
The Question Nobody Can Answer Yet
ChatGPT Work shipped a week ago. We don't know what the actual cost curves look like at scale. We don't know which marketing tasks it handles well and which it fumbles. We don't know whether the tiered pricing will push teams toward intelligent routing or just default laziness.
What we do know is that the unit economics of marketing just changed. A finished deliverable used to cost a human's time. Now it costs a few thousand tokens. The question isn't whether that's cheaper. It almost certainly is. The question is whether anyone will notice when it's more expensive than they thought.
That's the trap. Not the technology. The assumption that cheaper per task means cheaper overall. It doesn't. It means the costs moved somewhere you're not looking yet.
Back to all posts