Model spend / task
$0.37
49% lower · from $0.72
Cost / correct answer
$0.45
78% lower · from $2.01
Configuration and scoring
The 23 September 2026 study used Claude Opus through Claude Code for 50 tasks: 40 required retained knowledge and 10 could be answered from the repository. Deterministic keyword checks scored 18 baseline answers and 41 lean answers correct.
The lean configuration replaced a roughly 23,000-token static context file with a three-line pointer and a six-verb MCP profile. It changes the context strategy as well as enabling retrieval.
| Mean per task | Baseline | Enkeon · lean |
|---|---|---|
| Input tokens | 503,541.2 | 202,258.94 |
| Model spend | $0.7232 | $0.3708 |
| Runtime | 26.2675 s | 16.8827 s |
| Agent turns | 6.60 | 4.76 |
| Correct answers | 18 / 50 | 41 / 50 |
Scope and limitations
Means were reproduced from results-final.jsonl (off arm) and results-lean-on.jsonl in the internal study. Input tokens include cache reads and creation. Spend and runtime are CLI-reported. Model spend excludes Enkeon fees.
Keeping the original context file produced about 16% fewer input tokens and approximately unchanged model spend. The larger reduction belongs to the lean configuration shown here.
One run per configuration does not establish run-to-run uncertainty, human time saved or long-session performance. Keyword scoring is a narrow correctness criterion. These figures are internal workflow results, not a customer productivity benchmark.