Taskcost
Board ship→
Issue № 001·Pre-run cost previews for LLM flows

Know theline before you ship.

Estimate, replay, and gate AI spend before a single token leaves your repo. Like Infracost — but for prompts, agents, and the long, expensive tail of LLM flows.

Try the playgroundHow it works→
Form 1066 · Cost Manifest
Vessel
agentic-rag-v3
Per-task low
$0.034
Per-task high
$0.187
Monthly forecast
$1,420 → $980
Mix
gpt-4o → llama-3.3-70b
Quality Δ
−2.1pts
Cleared to sail
Savings
31%
In load line
TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕TF—Tropical FreshF—FreshT—TropicalS—SummerW—WinterWNA—Winter North Atlantic✕
¶ The three marks

Three sights.
One waterline.

Most observability tools tell you what your AI cost last week. Taskcost tells you what it'll cost next Tuesday, in the pull request, before the merge.

01 · Preview

Quote before commit.

Drop a prompt or a flow spec into the playground. We price it across 41 models and 11 providers, with low/high cost bands and quality priors. No tokens leave your machine.

See live demo →
02 · Replay

What if we'd run on Haiku?

Take seven days of real production traffic from your request log and re-price it against any candidate model mix. Same prompts, different costs, side-by-side.

Read the replay docs →
03 · Github App

The PR comment that pays for itself.

A .taskcost.yml in your repo. Every PR gets a cost-diff comment: per-flow, per-model, with the cheapest substitute called out. Infracost playbook, prompt edition.

Install the app →
Specimen · PR #1429

It looks like
this in your PR.

No dashboards to babysit. The comment shows up automatically when you change a prompt, a model, or a flow.

taskcost-botcommented · 2 minutes agovia .taskcost.yml
↑ 38%

This PR raises the monthly forecast by $540.

Two flows changed. One added. summarizer swaps to a heavier model; extractor grows its context window; classifier is new.

FlowStatusWas / moNow / moΔ
summarizer▲ heavier model$280$610+$330
extractor▲ context grew$140$220+$80
classifier+ new flow—$130+$130
Total$1,420$1,960+$540
Cheaper substitute available: routing summarizer through groq/llama-3.3-70b would cut $290/mo with a quality delta of −1.4 pts on your benchmark set.
¶ Locally

Or do it from
the terminal.

The same engine that powers the playground and the GitHub App ships as a Python package. tc preview runs offline, shows low/high cost bands, prints the table.

~/agentic-rag-v3
$ pip install taskcost-cli
$ tc preview \
    --input-tokens 200-500 \
    --output-tokens 100-400 \
    --freq 10000 \
    --task coding \
    --mixes default

╭─────────────── Cost preview · n=10,000/mo ────────────────╮
  Mix                       low/call   /month   Δ vs base   Q
  ─────────────────────────────────────────────────────────
  gpt-4o            (base)  $0.0143    $1,430        —      100
  gpt-4o-mini               $0.0011    $  110   −92.3%       91
  sonnet-4-5                $0.0094    $  940   −34.3%      102
  haiku-3-5                 $0.0019    $  190   −86.7%       88
  gemini-2.0-flash          $0.0005    $   50   −96.5%       86
  groq/llama-3.3-70b        $0.0007    $   70   −95.1%       95   ←
╰────────────────────────────────────────────────────────────╯
   ← cheapest with quality ≥ 90
¶ Why the line

In 1876, Samuel Plimsoll convinced Parliament that a single painted line on a ship's hull would save thousands of lives — by making it impossible to overload a vessel and pretend you hadn't. We think prompts deserve the same dignity. Engineers shouldn't ship LLM flows with their eyes closed and apologize on the next invoice. Draw the line. Then sail.

Stop guessing.
Start gauging.

Create a free accountRead the source

Free for personal projects. Open core. Apache 2.0.