Introducing the cheapest code agent subscription

Get the code agent · @rawmodels/pi on npm · How antislop was trained

We trained antislop to write. It also reads a repository, calls tools, edits files and runs tests, so we serve it to pi, the open-source terminal agent, for a flat monthly price. Below are the setup, our production measurements and the weak spots.

npm i -g @earendil-works/pi-coding-agent
pi install npm:@rawmodels/pi
pi
/login rawmodels

/login opens a page on rawmodels.ai with a short code. Approve it in the browser and pi gets its key. Then pick rawmodels/antislop with /model and give it a task.

With an existing key you can skip the login and set export RAWMODELS_API_KEY=skrm-….

We gave it a small Python repository with failing tests and one instruction: read the code, fix the bugs in calc.py, run the tests until they pass. We ran it three times against the public API, the endpoint you get. Each run fixed the code and ended on “all tests passed” after five to nine turns.

An agent spends its tokens on reading. Each turn resends the conversation plus the latest tool output, and the model replies with a short plan and a tool call. Output made up 8% of tokens in these runs, and the longest answer ran 721 tokens. For an agent, prompt caching matters more than decode speed.

One agent on production:

MetricValue
Decode speed, one stream95–100 tokens/s
Wait + prefill, 12–15k-token prompt1.8–2.8 s
Prompt served from cache on the next turn12.6k of 15.2k tokens
Share of agent input read from cache (real pi tasks)87%
Tokens moved per GPU-slot-second (real pi tasks)857

The gateway keeps a prompt cache per account. Your conversation stays on the GPU between turns, and the server processes only the new tool output. A turn that resends 15k tokens starts generating in about two seconds.

Then we ran many users at once from inside our network, each on its own account, six turns each. Agents got a growing code context and a three-second pause per turn for tools. Roleplay users got an 8k-token character card and twenty seconds to read and type.

WorkloadUsers at onceErrorsWait p50Wait p90
Coding agents20 of 122.0 s3.0 s
Coding agents, 4k output cap60 of 362.1 s10 s
Roleplay chats100 of 601.1 s1.4 s
Roleplay chats200 of 1201.1 s1.6 s
Roleplay chats301 of 1801.2 s5.5 s
Wait is the time from sending a request to the first generated token: queue time plus prompt processing.

The server generates four requests at once and queues up to eight more for up to 60 seconds. You give up your slot while you read or your agent runs a tool, so twenty roleplay chats share the box with waits under two seconds. An agent keeps its slot for most of a turn. Two agents run with no wait today. At four, the slowest turns waited about 26 seconds.

Agents run out of reserved memory before they run out of GPU. pi asks for up to 32,768 output tokens per turn, and the gateway reserves that much KV cache before the turn starts so a long answer can't run out of room. Most turns use a few hundred. With a 4k cap, six agents ran with no errors. Our next gateway change reserves what a turn uses and extends the reservation for the rare long ones.

One RTX 5090 runs SGLang. The weights are NVFP4. The KV cache is FP8 and holds 147,456 tokens on the GPU, and another 24 GB of host memory keeps older prefixes warm. EAGLE speculative decoding drafts three tokens ahead. Context is 128K. Tool calls go through the qwen3_coder parser and come back as standard OpenAI tool_calls, so pi and other OpenAI-compatible clients work.

A small gateway sits in front. It admits requests in order, runs one request per account and keeps each account's cache separate.

$11.30 a month: $10 for the plan and $1.30 for card processing and sales tax handling. You get $20 of usage a month at our list price, about 100 million tokens of coding-agent traffic. Any five hours caps at $3, so a few heavy users can't crowd out the rest. The plan also covers roleplay on rawmodels.ai. Past the allowance, requests draw on your prepaid credit or wait for the window to reset.

Launch code: RAWPI50 takes 50% off for the first three months.

It is a 27B model. It handles small and medium tasks: read a module, find a bug, write a test, refactor a function. A frontier model will beat it on large changes across many files. With one request per account, an agent that fans out parallel tool calls runs them one after another. One server also means a queue in a busy hour, and we add a second box once that queue stops emptying.


Test GLM Raw and frontier models in our interactive playground

Open playground

More from the blog

All articles
Open antislop model

Open antislop model

We trained Qwen3.8-27B against a calibrated judge that tells human writing from model writing. It now ranks fourth of 29 on EQ-Bench 4, above GPT-5.5. The weights, the GGUF builds and the reward are open.

October 16, 2026 · 6 min read

OpenJev. System One model ready for any workload: images, videos, texts

OpenJev. System One model ready for any workload: images, videos, texts

Five decision models on OpenRouter cannot see an image or a video. OpenJev can. It is also faster, free for 1,000 requests a day, open-weight, and right on tasks from Jev's own benchmark that Jev gets wrong.

September 29, 2026 · 10 min read