Updated May 28, 2026
Your Prompts Are Too Short: The Data Proves It
OpenRouter's data shows average prompt length grew 4x in a year, driven by developers loading context. Here's why marketers should follow their lead.
More context, better output. The numbers back it up.
OpenRouter published a state-of-AI report with one stat that stopped me cold: average prompt length grew 4x over the year. The growth isn’t the interesting part. Who’s driving it is.
Programming is doing the heavy lifting
That 4x jump in average prompt length? Developers are carrying it. They’re feeding entire codebases into AI, averaging more than 25,000 tokens per request. Call it 50-plus pages of context per prompt.
Marketing prompts, meanwhile, sit flat at around 5,000 tokens. Same story in finance, health, legal, and most other categories. Everyone outside engineering is still typing one-liners.
The good news, and the catch
The good news: you don’t need to write a novel. 5,000 tokens is roughly four pages of context, and that’s plenty for most marketing work.
The catch: developers already worked out that more context produces better output, and they’re running circles around the rest of us while we catch up.
What the developers are actually doing
Programmers aren’t writing longer prompts because they’re chatty. They’re loading the model up:
- Uploading whole files: style guides, existing code, documentation
- Showing examples: “here’s working code, make mine work like this”
- Stating constraints: the framework to use, how to handle errors, the edge cases that bite
What you get back is AI that behaves like a junior developer who actually read the docs before opening their mouth.
What this means for marketers
You don’t need 25,000 tokens. You almost certainly need more than you’re using right now.
Instead of: “Write email copy for our sale.”
Try this: upload your brand guide, your three top-performing emails, the offer details, your audience segment notes, a sharp ICP, and a short list of things you want the AI to steer clear of. That’s maybe 8,000 tokens of input. It’s also the entire gap between “meh” and something you’d actually ship.
The developers aren’t smarter. They’re just handing the model more to work with. That’s the whole idea behind how I write prompts that actually earn their output: context is the lever, and most marketers are barely leaning on it. It’s the same reason I treat prompting as a core skill in how I think about AI for marketing, not a nice-to-have.
So go load one up.
Five more things worth knowing from the report
For the data people, a handful of other findings from the full OpenRouter report stuck with me:
- Open source hit 30%. DeepSeek alone processed 14.37 trillion tokens. “No budget” stopped being an excuse.
- Reasoning models crossed 50%. People are using AI to think through problems now, not just to look things up.
- Programming is more than half of all usage. Claude owns most of that market, and the real ROI is sitting in automation.
- The “glass slipper” effect is real. Whoever solves a problem first tends to keep the user. DeepSeek shows a boomerang too: people try an alternative, then come back.
- Asia doubled its usage. North America dropped to 47%, Asia climbed to 29%, and Singapore alone now accounts for 9.2% of global usage.
One caveat to keep in your pocket: OpenRouter represents under 1.7% of total AI tokens. This is a slice, not the whole pie. We have to assume from the data we do have that Google, OpenAI, and Anthropic are trending along similar lines. Directional, not gospel. But the direction is hard to argue with.
Stop feeding AI two sentences
You hand it two lines, then wonder why the output reads like a two-line answer. Once a week I send one short email that breaks down the exact context I load before I hit enter: the files, the examples, the constraints, the before-and-after. If you’re tired of writing prompts that come up short, it’s free.