Meta Cuts Coding Agent Prices 21x for Developers Who Give Up Their Data

Meta released Muse Code in beta on Wednesday, Aug. 5, offering developers a contributor tier with output tokens priced roughly 21 times cheaper than its standard rate in return for permission to train future models on their prompts and completions. The terminal agent, powered by Muse Spark 1.2, installs on macOS or Linux with one command. CNBC reported that Meta is challenging Anthropic and OpenAI, while the Wall Street Journal reported investor pressure for returns on its AI spending.

Contributor pricing

The standard tier costs $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached-input tokens as of Wednesday. Meta states that it will not train on prompts or completions submitted at that level. Teams receive limits of 3,000 requests and 4 million tokens per minute.

Contributor pricing falls to $0.10 per million input tokens, $0.20 per million output tokens and $0.002 per million cached-input tokens. That is roughly 12 times cheaper for input and 21 times cheaper for output, with a limit of 60 requests per minute. Both tiers require a Meta account and billing details before Muse Code will run.

Mark Zuckerberg directed new users toward the data-sharing option. “It’s easy and low-cost to get started,” he wrote on X. “Install Muse Code with one line and you can start on our contributor tier.”

Alexandr Wang, Meta’s chief AI officer, told CNBC the contributor tier costs less than one-tenth of pay-as-you-go pricing and requires developers to “opt-in to help improve the model.” He declined to disclose Muse Spark user numbers, calling adoption “exciting and strong.” Meta is only beginning to accept requests for zero-data retention, separate from the standard tier’s no-training commitment. It has not explained how the two differ. Wang called zero-data retention “a big enterprise feature that is important for folks.”

Muse Code and Muse Spark are proprietary, a departure from the open-weight Llama family downloaded roughly 1.2 billion times by early 2026. OpenAI’s Codex CLI and Google’s Gemini CLI harness are available under the Apache 2.0 license.

The benchmark gap

Meta’s Wednesday charts put Claude Opus 5 first in all three coding tests it published. On Terminal-Bench 2.1, Muse Spark 1.2 in Muse Code scored 82.9%, behind Claude Opus 5 in Claude Code at 86.7% but ahead of GPT-5.6 Terra in Codex at 81.8% and Grok 4.5 in Grok Build at 81.6%.

Muse trailed Claude Opus 5 on DeepSWE 1.1. On Meta’s internal test, Muse scored 70.6%, against 79.4% for Claude Opus 5. In a kernel-optimization test exceeding 1,000 tool calls, Muse improved performance by roughly 61% to 69%, against roughly 74% to 75% for Claude Opus 5.

Every benchmark figure is vendor-run and published by Meta, with each model paired to its own harness. Muse Spark 1.1 ran in the generic mini-swe-agent harness, while version 1.2 ran in Muse Code. Some of the reported 6.7-point Terminal-Bench gain and 6.3-point DeepSWE gain therefore reflects the new harness, not solely the model. Meta also tested GPT-5.6 Terra rather than the higher-priced GPT-5.6 Sol.

The runtime

Muse Code, which has no graphical interface or IDE integration, keeps its background agents alive throughout a session. Sub-agents work in isolated git worktrees, while a local append-only event log records each model call, tool run, approval and edit. Meta says the design is “replay-exact and restart-safe,” allowing work to resume after a crash.

“In testing we had it build six features for a game simultaneously with no collisions,” Zuckerberg wrote on X.

The business test

Meta shares fell 10% last week after a quarterly earnings call brought a light revenue forecast and dwindling free cash flow. Online advertising supplies 98% of its revenue.

Some Meta employees already use Muse Code, Wang told the Wall Street Journal. “We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective.”

Zed and Warp Both Ship AI Coding. They Bet on Opposite Surfaces.

One product anchors the workday to the file under your cursor. The other anchors it to the agent sessions running in the background. Zed and Warp both moved AI coding toward the center of their produ

The Implicator

Professional Tutorial: Build a Pi-to-Pi Agent Communication Bus

The useful version of multi-agent coding does not start with ten agents. It starts with two Pi Coding Agent sessions, using the terminal coding harness to hold different context: one session sits near

The Implicator

Pi Is Not a Claude Code Rival. It Is a Harness Rebellion

Flask creator Armin Ronacher published a technical essay on his personal site on January 31 that endorsed a yellow terminal coding agent called Pi as "the minimal agent within OpenClaw." His company E

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai