--- title: "DeepSeek Just Announced Price Hikes, Zuckerberg Immediately Goes All In: Meta’s New Model Offers Steeper Discounts but Charges a “Data Tax”" type: "News" locale: "en" url: "https://longbridge.com/en/news/295104547.md" description: "Meta has released its first coding agent, Muse Code, and the optimized model Muse Spark 1.2, aiming to challenge Anthropic and OpenAI. Built on an asynchronous background agent architecture, the tool supports full software engineering tasks for large codebases, featuring parallel processing, code generation, and debugging capabilities. Mark Zuckerberg personally announced the release, viewing it as a key move in Meta’s revitalized AI strategy" datetime: "2026-08-06T11:46:23.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/295104547.md) - [en](https://longbridge.com/en/news/295104547.md) - [zh-HK](https://longbridge.com/zh-HK/news/295104547.md) --- # DeepSeek Just Announced Price Hikes, Zuckerberg Immediately Goes All In: Meta’s New Model Offers Steeper Discounts but Charges a “Data Tax” Meta has just launched its first coding agent, Muse Code, alongside Muse Spark 1.2, a cutting-edge model optimized for programming. This may represent the company’s most serious push yet into the AI coding tools market. Meta CEO Mark Zuckerberg personally announced the news on social platform X, stating that the currently beta-available Muse Code can “execute complete software engineering tasks within large codebases,” thereby challenging leading AI labs Anthropic and OpenAI. Muse Code comes from Alexandr Wang, head of Meta AI, who leads the Meta Superintelligence Lab and oversees foundational model development. Wang joined Meta last June as part of Zuckerberg’s core personnel restructuring to revitalize the company’s struggling AI strategy. According to reports, this new terminal-based coding tool is powered by Muse Spark 1.2 and offers improvements in “code generation, complex debugging, codebase understanding, and end-to-end developer workflows.” ## Built for Coding, Yet Still Lagging Behind Claude Code “In testing, we had it build six features for a game simultaneously without conflicts,” Zuckerberg stated in his post, emphasizing that its parallel processing capabilities do not interfere with developers’ working copies. Reportedly, the core architectural highlight of Muse Code is what Meta calls asynchronous background agents. Unlike most competing tools that temporarily spawn auxiliary agents for each task, Muse Code keeps a set of specialized background agents continuously active throughout the session. These agents “remain active throughout the session rather than being temporarily generated for individual tasks, thereby avoiding redundant information gathering.” They execute subsequent steps independently and autonomously decide when to report back to the main agent. The practical effect of this architecture is lower latency and less supervision: an agent that already understands the codebase does not need to re-explore it every time a developer makes a new request. When tasks are large enough, Muse Code distributes them to independent sub-agents for parallel processing, with each sub-agent operating in its own isolated git work tree, ensuring the developer’s working copy is never touched. A second notable design choice is auditability. If Muse Code crashes after running a long-duration task for 20 hours, it resumes precisely from where it stopped, without losing work or requiring re-prompting. Every model call, tool execution, approval, and edit is appended to a local event log before execution. Meta states that this single source of truth gives the runtime the property of being “precisely replayable and restart-safe.” Notably, the latest Muse Spark model was developed and trained alongside Muse Code, enhancing overall coding performance. Meta released benchmark charts comparing Muse Spark 1.2 with other coding models. On Terminal-Bench 2.1, Muse Spark 1.2 running within Muse Code scored 82.9%, slightly higher than OpenAI’s GPT-5.6 Terra (81.8% in Codex) and xAI’s Grok 4.5 (81.6% in Grok Build), but trailing Anthropic’s Opus 5 best-effort score in Claude Code, which led at 86.7%. On DeepSWE 1.1, Muse Spark 1.2 achieved 59.3%, ranking third behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). Most strikingly, on Meta’s own internal coding benchmarks, Muse Spark 1.2’s 70.6% easily beat GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%), but still lagged behind Opus 5’s 79.4% by nearly 9 percentage points. Even in tests designed by Meta itself, Anthropic’s model emerged victorious. In fact, Claude topped all three charts. Nevertheless, the generational improvement is tangible: Muse Spark 1.2 improved by 6.7 percentage points over version 1.1 on Terminal-Bench and by 6.3 percentage points on DeepSWE. Meta also showcased an impressive long-running case study: running Muse Spark 1.2 for up to 24 hours on NVIDIA Hopper hardware, involving over a thousand tool calls to optimize GPU kernels. The agent worked in Triton without wrapping existing third-party libraries, ultimately achieving “substantial improvements” on KDA and MLA kernels, including non-intuitive optimizations such as recentered gated accumulation decay. Zuckerberg noted, “It continued to discover substantial improvements even after the initial exploration phase.” ## Winning on Price? Trading Data for Discounts **Affordable pricing appears to be Meta’s key differentiator for Muse Code against popular products from Anthropic and OpenAI.** To this end, Meta has offered two tiers for its model API. First is the standard tier, where the tool defaults to pay-as-you-go pricing identical to Muse Spark: $1.25 per million input tokens and $4.25 per million output tokens. This price is lower than Anthropic’s Sonnet 5 model, which typically charges $3 per million input tokens and $15 per million output tokens. Meta promises that prompts and generated content at this tier will not be used for training, with rate limits of 3,000 requests and 4 million tokens per minute per team. Wang stated that the agent also features a “contributor tier accessible at significantly lower costs,” “even more than 10 times cheaper than the pay-as-you-go tier.” This tier is priced at $0.10 per million input tokens and $0.20 per million output tokens. Notably, this price is even cheaper than DeepSeek-V4-Flash. That model costs 1 yuan per million tokens when input cache misses occur, 0.2 yuan per million tokens for input cache hits, and 2 yuan per million tokens for output. Moreover, DeepSeek just announced: “We plan to generally increase API service pricing in the near future, with expected significant hikes.” In exchange, developers must “opt in to help improve the model,” referring to Meta leveraging third-party data to enhance its underlying technology. Furthermore, this tier has much stricter rate limits (60 requests per minute), clearly targeting individuals and small-scale experiments. Meta emphasized in its blog post that a zero-data-retention option is available to enterprises, meaning the company will not retain developer data to improve models, though this comes at an additional cost. He described this as “a significant enterprise-grade feature important to business customers.” Last week, the company reported disappointing earnings, with weak revenue forecasts and shrinking free cash flow, causing its stock price to plunge. Meta derives 98% of its revenue from online advertising, a market it dominates by targeting consumers based on user data. This new coding agent is another avenue for Zuckerberg to generate revenue from AI, even as his company continues to heavily invest in data centers and related computing infrastructure. ## A Complete Pivot in Open Source Strategy? Zuckerberg Leaves Suspense Compared to Meta’s usual approach, the most conspicuous absence in this launch was any mention of “open source.” Responding to developer questions on X, Zuckerberg stated, “I will have more to share soon,” seemingly hinting at potential open-source plans in the future. For three years, Meta positioned itself as the champion of open AI through its Llama series: from the consumer AI boom triggered by the LLaMA leak in February 2023, to the commercial licensing of Llama 2, to the 405-billion-parameter Llama 3.1 and Zuckerberg’s declaration that “open-source AI is the way forward.” Meta consistently promised developers that frontier weights should be freely downloadable, self-hostable, and fine-tunable. This strategy yielded significant results—by early 2026, the Llama family had been downloaded approximately 1.2 billion times, averaging a million daily, with self-hosting saving enterprises up to 88% in costs. Cracks began to appear with Llama 4. Launched in April 2025, Llama 4 received mediocre reviews and eventually admitted to inflated benchmark results. Meanwhile, Chinese open-source models from DeepSeek, Alibaba, and Zhipu AI rapidly rose, accounting for about 41% of Hugging Face downloads by the end of 2025, eroding Llama’s leadership. This setback prompted Zuckerberg to restructure the AI business into the Meta Superintelligence Lab (MSL) in the summer of 2025 and recruit Scale AI co-founder Alexandr Wang. On April 8 this year, the Llama era effectively ended. MSL released its first proprietary model, Muse Spark, with no downloadable weights, no self-hosting, and limited to cloud API access. At the time, a Meta spokesperson only stated that “current Llama models will continue to be open source,” remaining silent on future models; although Wang mentioned plans to “open-source future versions.” But four months later, the launch of Muse Code carries no open-source commitments—neither weights nor licenses—and the term “open source” never appeared in the launch blog post or Zuckerberg’s updates. This pivot is particularly sharp as competitors move in the opposite direction. OpenAI open-sourced Codex CLI under the Apache 2.0 license and launched the gpt-oss open-weight model; Google’s Gemini CLI also adopts the Apache license. Meta is now closer to Anthropic’s proprietary stance, whereas the company that once proclaimed open source as the way forward now requires developers to pay per token to use its models. The newly introduced contributor tier somewhat inherits the Llama strategy: instead of exchanging free weights for market share, it exchanges cheap tokens for training data, requiring users to “subsidize” the model with their own data. ### Related Stocks - [META.US](https://longbridge.com/en/quote/META.US.md) - [METU.US](https://longbridge.com/en/quote/METU.US.md) - [FBL.US](https://longbridge.com/en/quote/FBL.US.md) - [METW.US](https://longbridge.com/en/quote/METW.US.md) - [METD.US](https://longbridge.com/en/quote/METD.US.md) - [DPSK.NA](https://longbridge.com/en/quote/DPSK.NA.md) - [FBY.US](https://longbridge.com/en/quote/FBY.US.md) - [FBYY.US](https://longbridge.com/en/quote/FBYY.US.md) - [MAGX.US](https://longbridge.com/en/quote/MAGX.US.md) - [ANTH.NA](https://longbridge.com/en/quote/ANTH.NA.md) - [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [BABA.US](https://longbridge.com/en/quote/BABA.US.md) - [09988.HK](https://longbridge.com/en/quote/09988.HK.md) - [89988.HK](https://longbridge.com/en/quote/89988.HK.md) - [HBBD.SG](https://longbridge.com/en/quote/HBBD.SG.md) ## Related News & Research - [Meta Asks Engineers to Use MetaCode as It Tries to Catch Up with Anthropic and OpenAI](https://longbridge.com/en/news/294998546.md) - [Meta says its AI model hacked another company, adding to worries about bots going rogue](https://longbridge.com/en/news/295141974.md) - [FACTBOX-What we know about the rogue AI-agent security breaches](https://longbridge.com/en/news/295127514.md) - [LiveRamp adds Meta data to Cross-Media Intelligence measurement platform](https://longbridge.com/en/news/294824633.md) - [Stableford Capital II LLC Raises Holdings in Meta Platforms, Inc. $META](https://longbridge.com/en/news/294944023.md)