The free model that beats its own flagship, wired as your cheap second worker

# The Grunt Work Kit

The Grunt Work Kit is a free download: three plain-text files and one shell script that wire DeepSeek V4 Flash 0731 — a free, MIT-licensed open-weights AI model — into Codex as a *subagent*, so your main model stays exactly where it is and the routine work goes somewhere cheap. It exists because the upstream project is documented Chinese first: the wiring is genuinely about ten minutes, but only if you can read the instructions. The keyword is GRUNT.

## The problem it fixes

You already have a cheap model available. So does everyone else — GPT-5-mini, Claude Haiku, Gemini Flash have been one dropdown away for a year. Almost nobody routes real work to them, for one honest reason: cheap has always meant worse, and a cheap wrong answer that *looks* right is the most expensive outcome there is.

This checkpoint is the first one where that trade reversed on paper. It outscores DeepSeek's own larger V4-Pro-Preview on agentic benchmarks while costing a fraction of it, which is why 785,771 people downloaded it in roughly eight days while already having a cheap tier they weren't using.

Measured first-hand on 2026-08-08:

• 785,771 downloads on Hugging Face. The checkpoint was created 2026-07-31, and HF's downloads field (a rolling 30-day window) is *identical* to downloadsAllTime — which is what proves the whole count accrued inside roughly eight days rather than being a longer average read as momentum.
• $0.09 in / $0.18 out per million tokens for deepseek/deepseek-v4-flash-0731 on OpenRouter, against $0.435 / $0.87 for deepseek/deepseek-v4-pro. That is the ~4.8x figure.
• A 22-item labelling job in 13.6–17.3 s across four runs (mean 15.17 s) on the free tier. An earlier draft of this quoted a single 16.14 s run. One sample is a lottery number, so the figure above is the mean of four.
• ARC-AGI-1 89.0% at $0.02/task and ARC-AGI-2 61.4% at $0.04/task, independently verified by ARC Prize.

Not measured by us, and labelled as such: the Terminal Bench 2.1 figure of 82.7 against the larger model's 72.1 is vendor-reported, from DeepSeek's own model card. We reproduced that the card says it. We did not re-run the benchmark.

1. Try it before you wire it (2 min). Run the model once through OpenRouter's free tier on a real, boring task of your own. If it cannot do your grunt work, none of the rest matters.
2. Put the key in the environment (3 min). Create a DeepSeek API key and export it as DEEPSEEK_API_KEY in the shell or secret manager that launches Codex, then run sh verify-key-present.sh from that same shell. It prints whether the key is set and how long it is, never the value. Never paste the key into a chat, an issue, a screenshot, or a repo.
3. Paste the install prompt into Codex (3 min). SETUP-EN.md step 2 has the exact block. It makes no DeepSeek API call and does not change your main model or provider, and it deliberately stops before the paid smoke test so you can inspect what landed.
4. Inspect the Hook, then trust it (2 min). Run /hooks in Codex and confirm it matches ONLY v4_flash_worker and points at the installed plaintext-handoff script. If it matches anything broader, do not trust it. Then start a NEW Codex task — a running task will not reload the Hook.

The wiring is the easy part. Delegating badly costs more than not delegating at all, because you pay twice: once for the cheap attempt, once for the good model to redo it.

> If a competent person could do this task correctly with no context beyond what is in the prompt, delegate it. If getting it right depends on knowing *why* you are doing it, keep it.

Cost is not the axis. *Context dependence* is. WHAT-TO-DELEGATE.md turns that into two tables — the task classes that pay off (classifying lists, extracting fields from many similar documents, reading long logs for a named pattern, bulk renaming to a stated convention) and the ones to keep on your good model — plus the verification trap that eats the savings when you get it wrong.

A subagent sends prompt content to a different provider. Whatever the parent hands the child leaves your existing vendor boundary. Do not route customer data, credentials, private keys, or anything under an NDA, and do not assume the parent redacts for you — it does not. This is the same consideration as adding any third-party API, and being cheap does not change it.

• `SETUP-EN.md` — the four-step English walkthrough: key handling, the install paste, inspecting and trusting the Hook, and the smoke test. Includes the five conditions that separate a real pass from a false one, and a symptom/cause/fix table for when it doesn't work.
• `WHAT-TO-DELEGATE.md` — the one-line routing test, the two task tables, the verification trap, and a worked timing example with all four raw runs.
• `verify-key-present.sh` — confirms DEEPSEEK_API_KEY is visible to the shell that will launch Codex, the single most common install failure. Prints whether it is set and its length, never the value. Exits nonzero when unset.