Cut your AI bill by 40% without touching the model: here is how one company did it

Key points

• Writer's research team cut the average cost per AI task by 41%, from 21 cents to 12 cents, by redesigning the orchestration layer around the model rather than the model itself.
• Token consumption, the volume of text processed in each task, fell 38%, from 14,200 tokens to 8,800 tokens per task.
• Task success rates held steady or improved slightly, moving from 78% to 81%, while median task time dropped 44%, from 48 seconds to 27 seconds.
• The efficiency gains only held reliably on the two most capable models tested: Writer's own Palmyra X6 and Anthropic's Claude Sonnet 4.6.
• The researchers published their findings in a paper that any engineering team can apply today, because it requires no changes to the underlying AI model.

AI is expensive. Any company that has moved beyond a small demo and started running AI tools at real scale has felt that in the billing dashboard. A new research paper from Writer, an enterprise AI software firm, now puts hard numbers on why costs spiral, and shows a practical fix that does not require swapping out your AI model or retraining anything.

The short version: the problem is not the model. It is the plumbing around it.

What exactly is an "AI harness" and why should you care?

An AI harness, sometimes called an orchestration layer, is the software that sits between your application and the AI model itself. Think of it as a stage manager. It decides what instructions to send the model, what documents to hand it, how to handle errors, and when to call in specialist tools. Most development teams build this layer quickly and treat it as throwaway glue code. Writer's researchers argue that is the single most expensive mistake in AI engineering right now.

The paper tests six different AI models, including Anthropic's Claude Sonnet 4.6, Google's Gemini Flash 3.5, and Writer's own Palmyra X6, on 22 standardised business tasks. They kept the models and tasks identical, then compared a standard harness against a carefully optimised one. The results surprised even the researchers.

Costs fell by 41%. Task completion time dropped by 44%. Quality did not drop at all.

So where was the waste going before? The culprit has a name inside the industry: "tokenmaxxing." A token is a small chunk of text, roughly three-quarters of a word, that an AI model reads and generates. Providers charge per token. Tokenmaxxing is the habit of dumping enormous amounts of text into every AI request, hoping the model figures things out, rather than engineering the request carefully.

Writer's CTO Waseem AlShikh put it plainly to VentureBeat: "Your invoice is tokens-per-task times price-per-token, and most teams only watch the second number."

Price cuts from AI providers feel like relief. But if your system is burning three times as many tokens as necessary, cheaper tokens still means a bloated bill.

The optimised harness fixes this through four moves: caching the parts of the prompt that never change so you are not re-paying for the same text on every call; compressing conversation history so old context does not pile up; sending search tasks to small specialist sub-agents rather than stuffing raw documents into the main conversation; and cutting unconstrained retry loops that send token counts through the roof when the model hits an error.

One caution worth flagging. The sub-agent approach, where you delegate tasks to smaller specialist models, only worked reliably on the two strongest models tested. Lighter models like Gemini Flash 3.5 scored below a usable reliability threshold on those delegation tasks. If your team runs on a smaller, cheaper model, some of these gains will be harder to capture.

But the broader lesson holds regardless of which model you use. Structure your prompts so the static parts come first and the dynamic parts come last. That single change makes prompt caching actually work and stops you paying twice for the same instructions.

The one honest takeaway: before you spend another dollar on a more powerful AI model, audit how many tokens your current system is burning per task. That number, not the per-token price, is where your real AI costs hide.