AI/Fin Research
Guide

Muse Spark puts an automated research pipeline within any researcher's budget

Meta's model costs about a quarter of the top models by the token, and its coding agent installs with one command. The steps, the prices and the one tier to avoid.

AI Fin ResearchStarter4 min read
A full paper pipeline run costs 68 cents on Muse Spark. Being broke is no longer a reason to stay out.
$1.25per million input tokens on Muse Spark's standard tier. Opus 5.5 charges $4
$4.25per million output tokens. Opus 5.5 charges $20
$0.18per hour of audio transcribed with Meta's Muse Voice Transcribe

Prices are from Meta’s and Anthropic’s pricing pages on 9 October 2026.

Six steps for a first-time user

Muse Code is Meta’s coding agent, in early beta. Like Claude Code, it is a program you install and talk to in a terminal in plain English.

  1. Install it. Open a terminal and paste this line.

    curl -fsSL https://dev.meta.ai/install.sh | bash

  2. Make a project folder and start the agent.

    mkdir sec-panel && cd sec-panel && git init
muse

  3. Sign in. A browser window opens at dev.meta.ai. Sign in there when the agent asks.

  4. Choose the standard model. Set the model to muse-spark-1.3. Leave the one whose name ends in -contributor alone until you have read the warning further down.

  5. Ask for a plan first. Type /plan and then the job. The agent reads your files, writes a plan into .agents/plans/ and stops for your approval before it changes anything.

    /plan Read https://aifinresearch.com/guides/sec-filings-with-an-agent.md and
create sec.py and tests/test_panel.py exactly as given there. Build
data/panel.parquet for Apple, Microsoft and JPMorgan Chase (CIK 320193, 789019
and 19617) with the concepts Assets, Liabilities and NetIncomeLoss. Run pytest
and show me that it passes. Commit after every step that passes.

  6. Approve the plan and let it work. If you close the terminal, muse resume picks the session up again.

Terminal~/sec-panel
$ muse                    start the agent
 
> /plan Read https://aifinresearch.com/guides/... and create sec.py ...
● Plan saved to .agents/plans/. Waiting for your approval.
> approved, go ahead
● Write(sec.py)
● Bash(pytest -q)
  ⎿ 5 passed
An illustration of a session, not a recording. The grey notes are ours.

The same job at three prices

One full run of a ten-stage paper pipeline uses about 200,000 input tokens and 100,000 output tokens.

Model and tier Input, per million Output, per million One pipeline run Fifty runs
Opus 5.5, by the token $4.00 $20.00 $2.80 $140
Muse Spark, standard $1.25 $4.25 $0.68 $34
Muse Spark, contributor $0.10 $0.20 $0.04 $2

The model is also sold through OpenRouter, so one key can reach it alongside others.

Use it where volume matters

Scoring text in bulkA million filings or headlines scored once is mostly input tokens. That is where a price of $1.25 against $4 adds up.
Transcribing callsMuse Voice Transcribe is listed at $0.18 for an hour of audio. A thousand one-hour earnings calls cost $180.
Rerunning a pipelineFifty reruns on new samples or specifications cost $34 on the standard tier.

Keep the top model for the work that has to be right

No independent test found compares Muse Code with Claude Code or Codex on research tasks. The benchmark figures in circulation are Meta’s own. Muse Code is in early beta, and Meta’s announcement does not describe an auto mode or a finish-line command like the ones the SEC filings guide relies on for unattended runs.

So split the work. Build and debug the pipeline with the best model you can pay for, and write the tests yourself. Then run the volume on the cheap one, with the same tests deciding whether each run passed. A text measure should be run on two models anyway, and the second one now costs almost nothing.

References