More tasks solved than the heaviest model, at a sixteenth of the cost.
On 66 real feature tasks graded by hidden tests, KLP solved 59 at $0.046 each. The heavy model alone solved 56 at $0.721. Every figure recomputes from public receipts.
A low-cost model did the coding. A frontier model checked the exact result. Together they solved more tasks than the most expensive model working alone, for a sixteenth of the price.
We ran Krystal Loop Protocol on a public benchmark of 66 real feature tasks in three open-source Python projects: Click, Jinja and dirty-equals. Every result was graded by tests the models never saw.
KLP solved 59 of the 66 tasks at $0.046 each. The heavy model working alone solved 56 at $0.721 each. The standard frontier model working alone solved 56 at $0.123 each.
Three more tasks solved. 94% less spent.
Cost per task on 66 feature tasks, graded by hidden tests
- Heavy model alone (gpt-6-astra): 56 solved
- $0.721
- Frontier model alone (gpt-6.1-sol): 56 solved
- $0.123
- KLP Standard (deepseek-flash builds, gpt-6.1-sol reviews): 59 solved
- $0.046
94% less than the heavy model and 63% less than the frontier model, with three more tasks solved than either.
The whole KLP run cost $3.02. The heavy model's run cost $47.57. Across three separate KLP runs the loop solved 54 to 59 tasks at $0.046 to $0.061 each, so the saving held every time.
Where the saving comes from
It comes from who reads what. The low-cost model makes about 94% of the calls: searching, reading, writing and testing. The frontier model is called once per review and reads one compact packet: the task, the exact diff and the check output. That review costs about one and a half cents.
You still get frontier judgement on every change. You stop paying frontier prices for the typing.
Findings settled by running a test
A reviewer's opinion does not start a repair in KLP. Every blocking finding arrives with a small probe test, and the runner executes it. If the probe fails, the problem is real and the worker fixes it. If the probe passes, the finding is discarded.
In the best run the critic confirmed 77 problems this way. 70 were fixed and verified before handover, and no repair broke a task that was already passing. Every run ends with a patch and a plain-English receipt showing what was checked, what was found and what was fixed.
Already paying for a coding subscription? Use it as the reviewer.
The critic does not have to be a metered API. With Gemini 4 Argon reviewing through its command-line tool on a subscription, the loop solved 56 tasks, the same as the metered frontier critic on the same candidates, for $0.028 a task in metered spend.
Two agents, conflicting features
We also took 30 pairs of features whose patches conflict with each other. A plain merge delivered 2 of the 30. Two KLP workers, joined by the Semiotic Protocol (Semio), delivered 28 with both hidden test sets passing, at $0.103 a pair. The frontier model building both features alone delivered 22 at $0.196 a pair.
Semio carries the handover between agents as short typed messages that software can settle. Against the same two workers coordinating by chat, it delivered the same 28 pairs with 333 model calls in place of 581, and 61% lower integration cost.
The earlier gains still stand
This benchmark sits on top of results we have already published, and they all come with the same release:
- 75% fewer coordination input tokens. Semio turned four coordination calls into one, with every scenario handled correctly. Read that test.
- 24.6% lower total cost on a real product repair, with a worker that cost 95% less. Read that test.
- 21% fewer input tokens with compact briefs, and 3 update reads where polling took 21. Read that test.
- A large-repository task solved for $0.60. On a SWE-bench Pro task in Ansible, a repository of 5,501 files, a DeepSeek worker under KLP was scored solved by the official verifier.
Check every number yourself
Every figure in this post recomputes from receipts in the public repository. One command reprices every recorded model call and rebuilds each table. It needs no API keys and takes about a minute:
git clone https://github.com/KrystalUnity/krystal-loop-protocol.git
cd krystal-loop-protocol
python3 -m bench.verifyYou can also rerun the KLP arm of the benchmark with your own keys for about $3.
Run it on your own repository
pipx install git+https://github.com/KrystalUnity/krystal-loop-protocol
cd your-repository
klp init
klp run "Fix the pagination bug in list_jobs" --check "python -m pytest -q tests/test_queue.py"You get a patch and a receipt. Merging, pushing and deploying stay with you.
KLP 0.5.0 is free for individuals and small businesses. Anyone can evaluate it and rerun the benchmarks. Larger organisations need a commercial licence to use it in their work.
Read the receipts. Run the loop.
Give the repository to your coding agent. It has a page written for agents, with the commands, the facts it can cite and a summary for you.