Zuckerberg is going all out to compete with DeepSeek

Zhidx, August 6 report: Today, Meta launched its first programming Agent — Muse Code (beta version), as well as the new-generation model Muse Spark 1.2 carried by this Agent.

The model sparked heated discussions overseas as soon as it was released, with the core focus directly falling on its pricing strategy: it is cheaper than DeepSeek-V4-Flash. If users allow Meta to use their data to train the model, they can enjoy a 5% discount.

Interestingly, on the very same day, DeepSeek released an announcement stating: "We plan to raise the overall API service pricing in the near future, with an expected significant increase." Other leading domestic models also show signs of loosening in the price war. For example, the price of Zhipu's GLM model will also be raised significantly after the discount period ends.

When overseas giants OpenAI and Meta have all dived into the price war, domestic models such as DeepSeek seem to not follow the usual rules, making this global large model melee more and more interesting.

01. Up to 75x price cut, Meta targets user data

Mark Zuckerberg posted on X: "This model is positioned as easy to get started and low-cost, users only need one line of code to install it and start using the 'Contributor Plan'."

The so-called "Contributor Plan" means Meta provides two API versions: one version allows data sharing, and the other does not. The input price of the former is $0.10 per million Tokens, which is 12.5 times cheaper than the latter; the cached input is $0.002 per million Tokens, which is 75 times cheaper; the output is $0.20 per million Tokens, which is more than 21 times cheaper.

Some netizens joked: "It's hilarious that the input and output costs of Muse Spark are lower than those of DeepSeek-V4-Flash, and it also has a uniquely optimized cache hit rate. They don't care about profits, they only want data."

Earlier on July 31, DeepSeek-V4-Flash was officially launched on API. The price of this model when the input cache is not hit is 1 yuan per million tokens, the price when the input cache is hit is 0.2 yuan per million tokens, and the output price is 2 yuan per million tokens.

02. DeepSeek is "so strong that it doesn't look like a Flash model", can Meta keep up?

Right now, DeepSeek-V4-Flash is triggering a new wave of global enthusiasm. Since its API launch, its reputation for ultra-high cost performance has continued to rise.

Platzi, the leading online education platform in Latin America, co-founder and CEO just posted today and exclaimed: "I have no idea how DeepSeek-V4-Flash achieves this speed and quality, it's amazing."

Some netizens even said bluntly: "Claude is dead. There is no reason for the world to continue to be subject to a supplier whose price is 100 times higher but whose performance is comparable to its competitors."

But the good reputation of DeepSeek-V4-Flash not only comes from its low price, but also from its performance.

The benchmark test results of this model far exceed its own "big brother" DeepSeek-V4-Pro-Preview launched in April. Even today, many overseas netizens are still excitedly praising: "It doesn't look like a Flash model at all now." In the Almanbench test, the 0731 version of DeepSeek-V4-Flash surpasses Kimi K3, and its score is almost comparable to GPT-5.6 Sol xhigh.

So for Meta, which is trying to fight a price war with DeepSeek, what is the performance of its latest Muse Spark 1.2 and beta version of Muse Code?

03. Programming evaluation ranks second only to Opus5, Muse Spark 1.2 still has a gap with cutting-edge models

In terms of model performance, Muse Spark 1.2 has been greatly improved compared with its previous generation model, but there is still a certain gap with cutting-edge models.

Muse Spark 1.2 is an updated version optimized specifically for programming. In code capability benchmark tests such as Terminal‑Bench 2.1 and DeepSWE 1.1, its overall results are second only to Claude Opus5, and better than GPT‑5.6, Grok4.5 and Gemini3.6.

In the complex reasoning GDPVal‑AA V2 test, Muse Spark 1.2 is second only to Opus5; on the MCP Atlas benchmark for Agent tool calling, Muse Spark 1.2 ranks first.

Meta tested the model's ability to iteratively optimize GPU kernels in more than 1000 tool calls (up to 24 hours). With the help of Muse Code's agent coding environment, the model outperforms Gemini 3.6 Flash and GPT-5.6 Terra, but is inferior to GPT-5.6 Sol and Opus5.

Muse Code is positioned as an end-side coding Agent, capable of handling complete software engineering tasks in large code bases, including planning changes, writing code, and verifying results. Meta has released a series of cases, but whether it can really compete with Claude Code and Codex remains to be tested in developers' real scenarios.

As shown in the case below, Muse Code uses a simple Agent loop, supplemented by a set of asynchronous background Agents to enhance the capabilities of the main Agent. Their continuous operation reduces latency and lowers the need for human intervention in complex multi-step tasks.

Photon sphere simulation generated by Muse Spark 1.2

Game shooting interface generated by Muse Spark 1.2

Avocado lawn game generated by Muse Spark 1.2

The figure below shows that the user inputs the roaming video of the house into the terminal in the form of an MP4 file. Muse Code will parse the video and generate a visually rich vacation home marketing and booking page.

In addition, Muse Spark 1.2 scored 54 in the Artificial Intelligence Index (AII), and its agent knowledge work capability has been significantly improved compared with previous versions, making Meta tied for the third place among US labs with SpaceXAI.

04. Conclusion: Meta cuts prices to grab customers, DeepSeek raises prices to prepare for battle, behind the AI price war is a tough Agent competition

Following OpenAI's price cut, Meta has also launched a price war in the US AI circle. Meta is trying to take a share of the market under the dominance of OpenAI and Anthropic; and it is entering the end-side programming Agent market for the first time, trying to get back to the game. The recent major internal adjustment of Google DeepMind just provides a window of opportunity for Meta.

Low price may be a good way to attract traffic, but to open up the situation, higher model performance and Token quality are also needed. This is also the reason why DeepSeek-V4-Flash has become a hit overseas again.

It is worth noting that just today, DeepSeek released an announcement that it plans to raise the overall API service pricing in the near future. It can be seen that DeepSeek does not limit competition to the price war, but continues to focus on the competition of model quality. Recently, DeepSeek is