BackContext is live: one context layer for every agent you run.
Explore BackContextPer-token prices look small. Per-task costs at production volume do not. Here is the arithmetic we walk every customer through, using list prices for the four frontier models we benchmark against.
A typical business task, such as extracting line items from an order email, matching a candidate to a requisition, or routing a support ticket, runs about 4,000 input tokens and 500 output tokens. The input covers instructions, examples, the document and any retrieved context. The output is the answer. That is the basis we use throughout.
List prices per million tokens, input and output:
One task on GPT-6 Astra: 4,000 input tokens at $10 per million is $0.040. 500 output tokens at $50 per million is $0.025. Together that is $0.065 per task, or $65 per 1,000 tasks. The same arithmetic for the others:
$65.00 divided by $1.40 is where our 46× figure comes from.
Now run 1M tasks a month, a realistic load for an operations team with agents in the loop:
Against GPT-6 Astra, that is about $763K saved per year on a single workload. Most customers run more than one. Even against GPT-6.1 Sol, the cheapest of the four, the fine-tuned model saves about $139K a year at the same volume, before counting the difference in accuracy.
List price is not the whole bill. A wrong extraction or a bad match has to be caught and fixed by a person. That is why our eval reports show cost per 1,000 successful tasks next to cost per 1,000 tasks.
GPT-6.1 Sol looks cheap at $13.00 per 1,000 tasks. At 88.9% accuracy on our held-out customer tasks, 111 of every 1,000 answers are wrong. If each fix takes a person three minutes, that is more than five hours of rework per 1,000 tasks, which costs far more than the tokens. At 97.4%, a fine-tuned model leaves 26 of every 1,000 for review.
Some tasks still go to a frontier model when our model's confidence is low. We report the fallback rate alongside cost so the routing cost is visible, not buried.
DataFoldIT runs candidate matching at 52× lower cost per task than GPT-6 Astra, with higher accuracy on its own held-out set. Hytech Seeds projects $310K in annual savings, and order entry time dropped from 6 minutes to under 40 seconds.
If a task runs a few hundred times a month, or changes every week, pay per token and move on. Fine-tuning pays off when the same task runs thousands of times a day and the rules are stable enough to learn. Most businesses have a handful of tasks like that, and those few usually carry most of the bill. Start there, and leave the long tail of one-off requests on a frontier model.
The formula is simple: tokens per task, times price per token, times tasks per month. The part that needs your data is accuracy on your own held-out outcomes. Book a call and we will run both.
