20%
Of savings — our only fee
The efficiency layer for AI.
Trimml sits between your application and the model, striking the tokens your prompts never needed — cutting input spend by half while preserving intent and output quality. You pay only a share of what you save.
Install the CLI for Claude Code & Codex:
Install
curl -fsSL https://trimml.com/install.sh | bashDemo: Trimml strikes the redundant tokens from example prompts, keeping a verbatim subsequence — cutting them by roughly 54% with about 41 milliseconds of added latency (internal benchmark).
−54.2%
Reduction · internal benchmark
41ms
p95 added latency · internal benchmark
$0
Base fee — ever
Benchmark figures measured on Trimml's internal corpus, June 2026; your workload will vary.
How it works
One hop. Half the tokens.
Point your prompts at Trimml before they reach the model. We return the same prompt with the waste removed — nothing added, nothing rephrased.
Your application
Verbose prompts, repeated context, boilerplate instructions
Trimml
Scores every token, strikes the redundant ones in ~41ms
Any LLM
Same answer quality, a fraction of the input bill
Send the prompt
One call to POST /v1/compress with your raw prompt and a target ratio. No SDK lock-in, no model changes.
We strike the waste
A purpose-built encoder scores every token for information value and drops the filler — politeness, padding, duplicate context.
Pocket the difference
Forward the compressed prompt to any model. Every token we remove is metered, priced, and reported as savings.
The extractive guarantee
Compression you can audit.
Never
Paraphrased
Every output token is one of your input tokens, in order.
Never
Reordered
The compressed prompt is a strict subsequence of the original.
Never
Hallucinated
Nothing is generated, so nothing can be invented. Diff it yourself.
Please carefully review the following information and provide a detailed summary of the contents while making sure not to miss any important details.
// kept tokens highlighted — a verifiable subsequence, 23 → 6
Platform
Prompts first. Then the entire loop.
Prompt waste is only the visible layer. Agents burn far more on raw tool output, filesystem noise, and logs no human reads. Trimml is built to optimize all of it.
Live
Prompt optimization
Token-level compression for prompts, system messages, and RAG context. One API call, metered savings on every request.
In development
Agent runtime
An execution layer between agents and the OS that compresses tool output — directory listings, git, logs — before models ever see it.
Roadmap
Efficiency analytics
Org-wide visibility into tokens, spend, and waste — budgets, governance, audit trails, and optimization recommendations.
Pricing
We only earn when you save.
20% of measured savings
Original tokens minus compressed tokens, priced at a transparent blended rate ($3.00/MTok). That's the whole formula.
No subscription, no minimums
If we save you nothing, you owe us nothing. Incentives stay pointed the same direction as yours.
Transparent metering
Every request logs original and compressed token counts. Your invoice is reproducible from the raw usage table.
$61,000
Assumptions — typical workloads, not guarantees
Input-token share of spend60%
Typical prompt reduction54%
Gross savings$19,837/mo
Trimml fee — 20% of savings$3,967/mo
You keep$15,870/mo
// estimate only — actual savings are metered per request and depend on your prompts
Developers
Integrate in minutes.
One endpoint, one header. Works in front of every provider — Anthropic, OpenAI, Google, open weights — because the output is just a better version of your prompt.
$ curl https://api.trimml.com/v1/compress \
-H "Authorization: Bearer dk-9f2c41…" \
-d '{"model": "claude-sonnet-4-5", "input": "..."}'
{
"output": "review the following — summary — important details.",
"original_input_tokens": 23,
"output_tokens": 6,
"tokens_saved": 17,
"compression_ms": 38.2
}
Stop paying for waste.
Connect a key, route your prompts, watch the savings meter run.
Get started — free