▲ All images in this article are from the internet. Please contact us for removal if there is any infringement.

The Bidding Model for Agent Programming

This article was first published on Shadow Memo by Mo Yingsheng

On August 5, Zuckerberg announced in a high-profile manner that Meta had launched its first programming AI agent, Muse Code, with an open beta version available for download. Powered by the Muse Spark 1.2 model, it offered an eye-popping price point—just $4.25 per million output tokens. If users were willing to provide data to help Meta improve the model, the cost could drop more than tenfold.

The very next day, DeepSeek dropped a bombshell in its developer backend: API services would see an overall price hike, described as "significant."

One slashes prices aggressively to gain market share; the other quietly raises prices to prepare for battle.

Few would believe this was mere coincidence. The two models debuted in quick succession, both focusing on long-duration agent tasks. Zuckerberg’s move clearly targeted DeepSeek’s rhythm.

But the real intrigue lies not just in timing but in the starkly different approaches: one slashes prices to the floor while the other raises them amid soaring demand.

Several developer friends in my circle discussed this. Some said Zuckerberg aims to dominate the AI programming market, while others claimed DeepSeek had gotten carried away with its price hike.

Whether DeepSeek has gotten carried away is debatable, but the positions and calculations of both companies warrant a closer look.

What Exactly Is Muse Code?

Let’s start with Muse Code.

This isn’t some toy that merely helps complete a few lines of code. Running in the terminal, it installs with a single command and can handle staggering tasks—independently executing "complete software engineering tasks" in large code repositories, from planning changes to writing code and verifying results, covering the entire workflow.

For particularly large tasks, it splits into multiple sub-agents, each working independently without interference, before merging results.

Zuckerberg cited an example in his post: during internal testing, Muse Code simultaneously built six independent features in a game project without any conflicts.

Six features advancing in parallel without clashing—that’s impressive.

But what truly excites developers isn’t the functionality—it’s the price.

What does $4.25 per million output tokens mean? Let’s compare: competitor GPT-5.6 Terra charges $12 for output, while Claude’s top models are even pricier. Muse Code’s pricing is a fraction of its rivals’.

If developers opt for the contributor subscription—allowing Meta to use their code data to improve the model—the cost can drop as low as $0.20 per million output tokens. In Beijing, $0.20 won’t even buy a bottle of mineral water.

Alexandr Wang, head of Meta’s Superintelligence Lab, put it bluntly: "The Differentiated Positioning (differentiated positioning) of Muse Code and Muse Spark series lies in price, not capability."

In plain terms: I know my model might not be as smart as yours, but it’s far cheaper.

This admission is candid but reveals Zuckerberg’s deeper anxieties. Meta plans to invest $130–145 billion in AI infrastructure in 2026, nearly doubling its 2025 spending of $72 billion.

But Wall Street is impatient. A week before Muse Code’s launch, Meta’s stock dropped 10% due to weak revenue outlook and sharply declining free cash flow.

Investors grill Zuckerberg daily: When will all this spending pay off?

Muse Code is part of his answer—letting AI generate revenue, even if incrementally.

His strategy is straightforward: use ultra-low prices to lure developers, subsidize market share, and once users rely on your toolchain, gradually raise prices.

Netflix and Uber did this. But here’s the catch: switching costs in AI are far lower than in ride-hailing or streaming—migrating APIs essentially means changing a few lines of code.

Many see Muse Code as just a price war, overlooking a critical question: Why did Zuckerberg choose programming agents?

Meta’s Llama series has pursued a "general-purpose" approach in large models, with AI products focused on consumer-facing scenes like chat assistants and image generation.

Suddenly pivoting to programming agents—a geek-centric field—and directly competing with Claude Code and DeepSeek suggests strategic intentions worth probing.

The answer may lie in two numbers.

The first is developers’ willingness to pay. Among all AI segments, programmers are the most willing to spend on efficiency tools—by a wide margin.

Ask average users to pay $20 monthly for ChatGPT, and many hesitate. But ask a coder to pay $50 for an agent that handles tedious tasks, and they’ll sign up without blinking.

Developers also amplify spread (dissemination): a great programming tool spreads virally in tech communities, with extremely efficient word-of-mouth marketing.

The second number is Meta’s dilemma. Zuckerberg has spent billions on computing power, chips, and model training but hasn’t found a product form that makes AI self-sustaining.

Llama, despite being open-source and free, attracts mostly researchers and academics, with a long commercialization path.

Programming agents are different. API calls are metered—every line of code generated generates revenue. The commercialization path couldn’t be clearer.

In short, Zuckerberg chose programming agents to find the "closest-to-money" AI application.

Muse Code’s mission isn’t to be the smartest programming model but to be the first product that lets Meta’s AI business see actual revenue.

What’s Behind DeepSeek’s Price Hike?

Now, DeepSeek.

After announcing the price hike on August 6, the developer community erupted. Some joked, "The butcher is sharpening his knife," while others recalled DeepSeek’s May move to make V4-Pro’s 2.5-fold discount permanent—only to reverse course less than a quarter later.

But a closer look at DeepSeek’s recent moves reveals a trail.

In mid-July, DeepSeek introduced peak-valley pricing: API prices doubled during weekday peak hours (9 AM–12 PM, 2 PM–6 PM) compared to off-peak times.

Goldman Sachs noted in a report that this reflected sustained domestic demand for AI models, tightening computing resources, and the industry shifting from price wars to rational pricing.

Simply put: demand is so high that servers are strained.

By early August, peak-valley pricing wasn’t enough, prompting an overall hike. The core reason is straightforward: DeepSeek’s models are effective, developers rely on them, call volumes surge, and computing costs rise accordingly.

Unlike Meta’s billion-dollar capital budget, DeepSeek had to raise prices to maintain service quality.

But there’s a deeper layer: Liang Wenfeng’s strategic choices.

In late July, Liang held a four-hour closed-door investor meeting—long for a founder. Many of his remarks, viewed alongside the price hike, carry heavy weight.

He made a notable remark: "Restraint is a strategy. Sometimes you sacrifice some things to gain others."

What does restraint entail? Liang listed specifics: no 3D generation, no video generation, no world models, no pursuit of user volume, and no ambition to become the next super app.

DeepSeek concentrates all resources on one path: AGI technology.

Given DeepSeek’s current influence and technical reserves, venturing into video or multimodal applications would attract funding easily.

But Liang chose to bet everything on the most distant, difficult, and uncertain basket.

His words: "Now isn’t the time to maximize product returns. Building lower-level tech from a technical high ground is dimensional reduction. Products are byproducts on the AGI journey."

Bluntly: If I achieve AGI, products follow effortlessly. But if I chase quick profits and distract attention (divert resources), failing AGI would be the real loss.

Pricing-wise, Liang adheres to "earning a reasonable profit, not maximizing it."

He revealed DeepSeek’s API pricing aims to recover costs in ~10 months, unlike peers pricing at "the highest the market can bear."

Note the focus on "cost recovery," not "market share" or "subsidies."

This diverges sharply from Zuckerberg’s logic. While Zuckerberg fights for market share and AI self-sufficiency via price wars, Liang prioritizes pouring resources into AGI, with commercialization secondary. The hike stems from demand exceeding supply, not exploitation.

Two Logics, One Offense, One Defense

Viewed together, the contrast sharpens.

What does Zuckerberg have? A trillion-dollar capital budget, the world’s largest social user ecosystem, and a history of advertising-funded growth. His strategy: use money to erect a price barrier—you charge high, I undercut you by three-quarters; you can’t match subsidies, I push you out.

This is classic "internet playbook": burn cash to monopolize, then harvest later. It worked in ride-hailing, food delivery, and e-commerce.

But this tactic has a fatal flaw in AI. Switching costs for ride-hailing apps are high—you won’t install a new app for $2 savings. But API migration? A few code changes suffice. If DeepSeek’s models clearly outperform Muse Code in programming, will developers sacrifice code quality for $1 differences?

No. Programmers hate compromising on tool quality.

What does Liang have? A model widely recognize by developers, an AGI-focused R&D system, and a patience for commercialization.

His strategy: concentrate all resources on "making models smarter," avoiding diversification, and ignoring short-term returns. As Liang put it: "The time to fully pivot to commercialization is far off."

This confidence stems from DeepSeek’s stable funding and technical accumulate (accumulation). It doesn’t need AI to sustain itself yet.

Meanwhile, Wall Street pressures Meta.

One attacks with money and anxiety; the other defends with technology and patience.

Who Faces Greater Anxiety?

One background detail completes the story.

DeepSeek is also developing programming agents. In May, it formed the Harness team to compete with Claude Code in this space, with products reportedly in active development (internal codename: DeepSeek Code).

Liang explicitly stated in the closed-door meeting that Coding Agent is the top priority now.

In other words, Muse Code is charging into a Track (sector) where DeepSeek has already fortified defenses and is preparing counterattacks.

More intriguing are the product positioning differences. Muse Code pursues "full-process automation": give it a task, and it plans, executes, and verifies alone.

DeepSeek’s tool, per internal sources, emphasizes "human-AI collaboration": AI executes, humans decide, working in tandem.

Which route is better? Unclear. Full automation saves effort but may falter with complex logic; human-AI collaboration ensures precision but demands more developer input.

Each suits different scenes and users.

One thing is certain: the clash between these companies in this sector has just begun.

In AI, who faces greater anxiety?

On the surface, DeepSeek seems anxious. A rival with a multiple-fold budget invades its core market with drastically lower prices—anyone would tense up.

However, looking deeper, the one who is truly anxious might be Zuckerberg.

Meta has invested way too much in AI. With capital expenditures in the hundreds of billions of dollars, GPUs being snapped up globally, and computing clusters expanding exponentially, all of this is testing the patience of shareholders.

If Muse Code fails to deliver impressive data in a short period, such as user numbers, usage volume, retention rates, etc., then the pressure on Zuckerberg from investors will only continue to grow.

In contrast, DeepSeek's price hikes are due to supply falling short of demand and insufficient computing power. This kind of 'sweet dilemma' is something many AI companies can only dream of.

When a product needs to rely on price increases to regulate demand, it means it has established irreplaceable value in the eyes of users. What does it mean to succeed? This is it.

To put it bluntly, Zuckerberg is buying time today, using subsidies to gain market share and prices to acquire users, so that the AI business can generate revenue quickly and silence Wall Street.

Liang Wenfeng, on the other hand, is betting on tomorrow's technology by sacrificing today's potential earnings to increase the likelihood of achieving AGI. If he truly succeeds, today's price war will be insignificant.

In Conclusion

Free things are often the most expensive, but when it comes to the AI sector, this saying might need to be revised to 'cheap things may be the least reliable.'

Today, you might feel like you've scored a deal using Muse Code for $4.25.

But if it consistently fails to catch up to DeepSeek in terms of capabilities, if its performance ceiling is always lower than its competitors, then the few dollars you saved are actually coming at the expense of code quality.

For product makers, this is never a good trade-off.

Of course, the Premise (Chinese for 'prerequisite') is that DeepSeek can maintain its technological lead. This is the biggest headache for all challengers—after all, prices can be replicated, but intelligence is hard to copy.

Two companies, two paths. One is harvesting users today, while the other is sowing the seeds of tomorrow's technology.

Short-term wins and losses are not hard to determine, but long-term outcomes are anyone's guess.

In the AI race, burning money has never been a moat. The only real moat is having a smarter model than your competitors.

And that is something money can't buy. It relies on time, focus, restraint, and a bit of luck in getting it right.