AIBRIEF

Meta Launches Muse Code, Its First AI Coding Agent, Undercutting Rivals on Price

N
NathanTechnology Editor · Technical Lead
Published · Updated
According to CNA (中央社), Meta launched its first agentic coding tool, Muse Code, in public beta on macOS and Linux on August 5, 2026, under AI chief Alexandr Wang. Inside (Inside.com.tw) reports its Contributor tier is priced more than 10 times cheaper than standard pay-as-you-go rates, in exchange for Meta using submitted data to train future models.

What Is Muse Code? Features, Design, and Product Format

Meta introduced Muse Code as its first agentic coding tool, positioning it against Anthropic and OpenAI on the basis of "price advantage," according to CNA. The tool was built under Alexandr Wang, Meta's AI chief. CNA quotes Wang describing the product directly: "It can be installed with a single command, and then used to handle complete software engineering tasks for a variety of purposes, including planning changes, writing code, and verifying results."

According to Inside, Muse Code is a coding agent that runs inside the terminal, and it opened to public beta on macOS and Linux on August 5. Inside also details the tool's underlying design logic: rather than simply generating code on command, Muse Code first plans the scope of a change, then verifies its own results after completion. For complex tasks, it can spin up multiple sub-agents to divide the work, and it can run asynchronously in the background, so users are not required to stay at the terminal waiting for output.

When Was It Released, Who Built It, and What Model Powers It?

Per Inside, the public beta launched on August 5 on macOS and Linux. CNA reports that development was led by Alexandr Wang, who Meta recruited last year from his role as CEO of Scale AI to become a central figure in CEO Mark Zuckerberg's push to revive the company's enterprise AI strategy. Wang leads Meta Superintelligence Labs and oversees foundation model development, according to CNA.

CNA further reports that Muse Code is currently available as a preview version, paired with Meta's newest AI model, Muse Spark 1.2.

How Does Muse Code Undercut Rivals on Price?

According to CNA, Wang said Muse Code offers a "Contributor" plan priced more than 10 times cheaper than pay-as-you-go rates — but it requires users to "opt in to help improve the model," letting Meta use third-party data to strengthen its underlying technology.

Inside breaks down the two tiers in detail. The Standard tier charges $1.25 per million input tokens and $4.25 per million output tokens; data on this tier is not used for model training, and the rate limit is capped at 3,000 requests per minute and 4 million tokens per minute. The Contributor tier charges just $0.10 and $0.20 per million tokens respectively — input is 12 times cheaper and output is more than 21 times cheaper than Standard — with cached input priced as low as $0.002 per million tokens. The trade-off, per Inside, is that Contributor-tier users must agree to let Meta use their prompts and generated outputs to train its next-generation models, and the rate limit drops sharply to 60 requests per minute and 2.1 million tokens per minute.

Inside also benchmarks Muse Code's pricing against DeepSeek's V4 Pro, which recently made its promotional pricing permanent at $0.435 per million input tokens and $0.87 per million output tokens — at one point the cheapest frontier model globally. However, Inside reports DeepSeek has since announced a peak-hour surcharge: V4 Pro's output price during peak hours will rise from 6 yuan to 12 yuan per million tokens, effectively doubling, though the effective date has not been announced.

PlanInput (per 1M tokens)Output (per 1M tokens)Rate limitData used for training?
Muse Code – Standard$1.25$4.253,000 req/min, 4M tokens/minNo
Muse Code – Contributor$0.10$0.20 (cached input: $0.002)60 req/min, 2.1M tokens/minYes
DeepSeek V4 Pro – standard$0.435$0.87
DeepSeek V4 Pro – peak surcharge (announced, undated)6→12 CNY per 1M

Source: Inside (Inside.com.tw), CNA (中央社)

Why Now? Meta's Financial Backdrop

CNA reports that Meta posted earnings the week before the Muse Code launch, and its stock fell sharply after revenue guidance came in below market expectations and second-quarter free cash flow declined significantly. CNA also reports that as much as 98% of Meta's revenue comes from online advertising, which the company drives by using user data to deliver targeted ads to consumers.

What This Means

The numbers reported across both sources outline a clear structure rather than a single narrative: Muse Code's Contributor tier is priced at roughly one-twelfth to one-twenty-first of its own Standard tier (per Inside), and its Standard rates already undercut DeepSeek's V4 Pro peak-hour output pricing, which CNA and Inside's reporting place at up to 12 yuan per million tokens versus Muse Code Standard's $4.25. That aggressive pricing structure — and the data-sharing trade-off attached to its cheapest tier — was disclosed by CNA in the same week Meta reported a stock decline tied to weaker free cash flow, and against a revenue base that CNA reports is 98% dependent on advertising. The evidence from both outlets ties these facts together in time without either source drawing an explicit causal line between them.

📊 Evidence

FAQ

What operating systems does Muse Code support at launch?

According to Inside, Muse Code's public beta launched on August 5, 2026, on macOS and Linux.

What do users give up to get Muse Code's cheapest pricing tier?

Per CNA and Inside, the Contributor tier — priced at $0.10 per million input tokens and $0.20 per million output tokens — requires users to opt in and let Meta use their prompts and generated outputs to train future models, and it comes with a lower rate limit of 60 requests and 2.1 million tokens per minute, versus 3,000 requests and 4 million tokens per minute on the Standard tier.

How does Muse Code's pricing compare to DeepSeek's V4 Pro?

Inside reports DeepSeek V4 Pro's standard rate is $0.435 per million input tokens and $0.87 per million output tokens, with an announced but undated peak-hour surcharge that would double output pricing from 6 to 12 yuan per million tokens — compared with Muse Code's Standard tier at $1.25 input and $4.25 output per million tokens.

📎 Sources

  1. cna.com.tw
  2. inside.com.tw
N
NathanTechnology Editor · Technical Lead

Related

BRIEF

Taiwan Stock Exchange Cuts Disposition-Stock Period to 5 Days, Speeds Matching to Every 2 Minutes From August 10

According to a Central News Agency (CNA) report, the Taiwan Stock Exchange (TWSE) will implement new disposition-stock rules on August 10, 2026, shortening the standard disposition period from 10 to 5 business days and quickening intraday order matching from roughly every 5 or 20 minutes to about every 2 minutes. Stocks also flagged for a high day-trading ratio see their disposition period cut from 12 to 7 business days, per CNA and Commercial Times (CTEE) reports.

林紀旭 James Lin ·
BRIEF

SpaceX and Tesla Commit $16.8 Billion to Build AI Chip Plant Terafab in Texas

According to iThome and CNA reports, Texas Governor Greg Abbott announced on August 6, 2026 that SpaceX and Tesla will build the AI chip plant Terafab in Grimes County, Texas, with a first-phase investment exceeding $16.8 billion expected to create 3,000 jobs, with Intel joining as a process partner.

林紀旭 James Lin ·