Hystax Releases Open-Source Token Compressor That Cuts AI Coding-Agent Costs by 22.7% Without Changing the Agent

Hystax has released OptScale Kompressor on GitHub under the MIT license, opening up the deterministic token compressor behind its OptScale AI platform. On Terminal-Bench 2.1, it cut a coding agent's real provider bill by 22.7%, about twice Headroom's savings, with no significant change in the number of tasks solved.

Hystax, a cloud infrastructure, FinOps, and AI governance software company, today announced the open-source release of OptScale Kompressor, the token compressor built into its OptScale AI platform. In benchmark tests on AI coding-agent workloads, Kompressor lowered the provider bill by 22.7% with no statistically significant change in task success.

Hystax is making OptScale Kompressor available on GitHub as open-source software. Kompressor shortens the build logs, compiler messages and other machine-generated output that AI coding agents send to a language model, without changing the agent or its prompts. In Hystax's tests, it cut the provider bill by 22.7% and input tokens, the units of text that providers charge for, by 25%, compared with no compression.

AI coding agents re-send their entire working history to the model at every step, including command output they have already read. In Hystax's traces, this tool output makes up about half of all prompt data, and much of it is repetitive. Through prompt caching, providers bill repeated input at a steep discount, but only as long as it stays identical. A compressor that changes its output from one turn to the next loses that discount and can cost more than it saves.

How OptScale Kompressor Works

Kompressor sits between the agent and the model provider and rewrites only machine-generated tool output. Its design follows four principles:

• Deterministic: Kompressor is rule-based and uses no machine learning models. The same input always produces the same output, which keeps the provider's cache discount intact.
• Recoverable: Every removed passage is replaced with a labeled marker, and the model can fetch the original back.
• Format-aware: Each type of output, such as logs, compiler messages, JSON, diffs or source code, goes to a compressor built for it.
• Fail-safe: If compression fails or does not shrink the output, the original passes through unchanged.

Kompressor ships as a single Rust binary for Linux, macOS and containers and processes one kilobyte of output in about 0.2 milliseconds. Its compression engine makes no outside network calls, such as telemetry or model downloads.

Benchmark Results

Hystax ran the Codex CLI agent with OpenAI's GPT-5.5 model on Terminal-Bench 2.1, a public benchmark of 89 command-line tasks. It compared actual provider bills, including cache discounts, for no compression, Kompressor, and Headroom at its default and most aggressive settings. Headroom is an open-source compressor whose algorithms informed Kompressor.

• No significant change in task success: Kompressor solved 81.1% of tasks, against 82.2% without compression.
• About twice Headroom's savings: Headroom cut costs by 8.2% at its default settings and by 11.6% at its most aggressive. Against Headroom's default settings, Kompressor cost 15.8% less.
• Largest savings in log-heavy work: Bills fell 26% in machine learning training and evaluation tasks and 28% when building software from source.
• Smaller requests: Requests above 200,000 tokens fell from 1.9% to 0.13% of the total, keeping agents further from the context window limit, the most text a model can process at once.

"Coding agents spend a large part of their budget re-reading their own build logs. Kompressor removes that repetition without hiding anything from the model, because every line it removes can be fetched back," said Nick Smirnov, CEO at Hystax. "We are making it open source because any component that sits between a company's agents and its models should be open to inspection, line by line."

Availability

OptScale Kompressor is available on GitHub at https://github.com/hystax/optscale-kompressor under the MIT license.

It also runs in OptScale AI as part of the platform's cost optimization features. The full benchmark report, with per-task results and methodology, is available on the OptScale AI blog.

About Hystax and OptScale AI

Hystax is a cloud infrastructure, FinOps, and AI governance software company. Its products help enterprises migrate, protect, and optimize cloud environments, and govern AI usage across teams and models.

Hystax's OptScale is an open-source FinOps platform used by thousands of organizations worldwide to manage and reduce cloud costs across multi-cloud environments.

OptScale AI, released in May 2026, is an enterprise AI ROI and governance platform. It ties AI spend to delivered work, such as merged pull requests, cuts costs through smart routing and token compression, screens every prompt, and keeps AI agents within their limits. Learn more at: optscale.ai

Media Contact

Xenia Lavrova, Hystax, 44 753716973, [email protected], https://hystax.com

SOURCE Hystax