Meta crashes prices for AI programmer, challenging Anthropic and OpenAI

Meta released its first AI agent for writing code — Muse Code — and immediately went all-in: not through technological superiority, but through dumping. The tool entered a market where Claude Code from Anthropic and Codex from OpenAI were already firmly established, and offered a price that would make competitors nervous.

What happened

On Wednesday, August 5, Meta officially introduced Muse Code, a terminal AI agent capable of conducting the full development cycle: from planning changes to writing code and checking the result. Mark Zuckerberg personally announced the launch on social networks, which in itself hints that the company does not consider this an ordinary release, xrust emphasizes.

The project is overseen by Alexander Wang, head of Meta Superintelligence Labs, whom Zuckerberg lured last summer precisely in order to pull the company’s stalling AI strategy out of the hole. Muse Code is his most notable public achievement in office.

The tool runs on top of the latest Muse Spark 1.2 model, trained in parallel with the agent itself — according to Wang, this directly affected the quality of the code. The installation takes one command in the terminal, after which the agent takes over the routine: writes functions, tests them and records every action and edit in a local log — if the system crashes, work can be continued from the same place. Meta claims that in tests the agent implemented six game mechanics in one project in parallel and without conflicts.

Why it's all about price

class=»notranslate»>__GTAG5__ Formally, Muse Code is far from the first agent of this kind: Claude Code from Anthropic and Codex from OpenAI have existed for a long time and have managed to acquire a user base and reputation. Meta entered foreign territory not with new functionality, but with a price list: $1.25 per million input tokens and $4.25 per million output tokens — that’s about 100 and 350 rubles at the current Central Bank exchange rate. And for the “contributor” tariff, according to Wang, the price is more than 10 times lower than the standard one.

The calculation is clear: according to sources surveyed by CNBC, coding today accounts for the lion’s share of consumed tokens at all major AI laboratories. This is the most lucrative segment of the generative AI market, and it is in this segment that Meta has had nothing of its own until now. Dumping here is not a sign of weakness, but a standard late entry tactic: to crush someone else’s margin and force the market to recalculate the economy in a new way.

There is also an internal motive. Investors have been asking Meta for months when its massive investment in AI infrastructure will finally start to pay off, and the company's recent stock slide has only added to that pressure. A public, measurable product with a clear price is an easy way to show the market something tangible.

How Meta tests the product on itself

Before releasing Muse Code to the public, Meta forced thousands of its own engineers to use the agent weekly as part of an internal policy. By the time of release, there were about 7 thousand such active users, and they managed to make over 800 edits, which, according to the company, directly improved the quality of the model. This approach is not uncommon in the industry: both Anthropic and OpenAI tested their agents in a similar way before introducing them to the wider market.

For corporate clients, Meta has also provided a zero data storage option — that is, the code and correspondence with the agent do not end up on the company’s servers. This is a standard requirement of large businesses for any AI tool that is allowed into real production, and its absence from a competitor almost automatically excludes it from the tender.

Among the weak points — at the start, Muse Code does not have a separate application or web interface, only a terminal. Claude Code and Codex by this time had already acquired more friendly shells, so Meta consciously sacrificed convenience for the sake of speed to market.

What does this mean

Over the past year, the market for AI development agents has become the site of a major battle between laboratories — much more fierce than the competition between chatbots. This is where it is measured who actually brings money to the business: a subscription to a chatbot costs tens of dollars per month, and the tokens that are burned by an agent who autonomously rewrites thousands of lines of code are completely different budgets.

The emergence of a third major player with a dumping price will almost certainly accelerate the already ongoing price race in the segment — competitors will have to either reduce the cost of their agents or prove that the difference in code quality justifies the overpayment. For developers and IT companies using such tools, including in Russia through available channels, this primarily means one thing: in the coming months it is worth taking a closer look not only at the capabilities of agents, but also counting how much each thousand lines of generated code costs — the balance of power here may change faster than it seems.

Sources:
cnbc.com
bloomberg.com
siliconangle.com

Xrust Meta brought down prices for AI programmer, challenging Anthropic and OpenAI