Meta entered the AI coding race on August 5 with Muse Code, a terminal-based coding agent, and Muse Spark 1.2, its coding-focused AI model, taking direct aim at Anthropic’s Claude Code and OpenAI’s Codex.

Meta claims that Muse Code is designed for software development on extensive codebases rather than creating single code snippets. Unlike its open-weight Llama models, Muse Code is paid software for macOS and Linux that integrates closely with Muse Spark 1.2.

The CEO Mark Zuckerberg claimed that the agent is able to execute “complete software engineering tasks across large repos: planning changes, writing code, validating the results.” He added that internal testing showed that the agent is capable of creating six different game features at the same time with no issues, while suggesting that there will be other open-weight AI software launches in the future.

Where Meta stands in a market it mostly watched

Meta is making its way into a market where companies like Anthropic and OpenAI are key players, and some startups, like Cursor, have already witnessed a need for AI-based software engineering tools.

Meta has observed that Muse Spark 1.2 may be competitive in some ways, but it is far from being the market leader. It could only score 82.9% on Terminal-Bench 2.1, beating GPT-5.6 Terra and Grok 4.5 but losing to Anthropic’s Opus 5 on this benchmark. Muse Spark 1.2 showed mediocre performance on the DeepSWE 1.1 benchmark as well, achieving third place, with Opus 5 being the winner.

Third-party tests yielded the same result. Muse Spark 1.2 ranks number 14 out of 186 models according to Artificial Analysis in terms of Intelligence Index, scoring 54, which is above the median of 32. It experiences good latency and has a relatively low price, meaning that Meta is focusing more on the balance between performance and price than being the top performance model.

How Muse Spark 1.2 prices against DeepSeek’s V4

Pricing could become one of Meta’s biggest competitive advantages—if developers accept the tradeoff.

Reuters, citing Artificial Analysis, reported that DeepSeek V4-Flash is the least expensive widely used AI model to run in benchmark tests, making the DeepSeek V4 family a key pricing benchmark for competing frontier models.

Muse Spark 1.2 has two different pricing options. Its Standard Pricing for API use is $1.25 per million input tokens, $0.15 for cached input tokens, and $4.25 per million output tokens. Meta offers its Contributor price for approved prompts and responses, bringing the prices down significantly to $0.10 for input tokens, $0.01 for cached input tokens, and $0.20 for output tokens.

DeepSeek charges $0.14, $0.0028, and $0.28 for V4-Flash for one million input, cached input, and output tokens, respectively. As for V4-Pro, the prices are $0.435, $0.0145, and $0.87, respectively, depending on peak and off-peak times as per the API description.

The above comparison shows two different approaches to pricing. When it comes to regular pricing, Muse Spark 1.2 is more expensive in comparison to both DeepSeek models. However, under the Contributor tier, Meta offers a lower price for input and output as compared to both V4-Flash and V4-Pro, which makes it one of the cheapest products for eligible users who are ready to share their data.

The price benefit of DeepSeek might not last forever. Reports say that DeepSeek is planning to apply a major price hike in its AI services that hasn’t been revealed yet. If the new prices are introduced, it might close the gap between DeepSeek and other strong players in the industry.

The bet on persistent background agents

The distinguishing feature of Meta’s agent might be its architecture rather than its benchmark results.

As per the company’s engineering notes, Muse uses persistent background agents to keep the background context of the repositories during each development session, rather than restarting from zero during each request. When dealing with larger projects, Muse conveys the tasks to be carried out in numerous isolated Git worktrees, so that sub-agents can run their tasks without affecting the primary workspace of the developer.

Meta is also focused on reliability. Every model call, tool run, and file edit is logged in a local event log, enabling interrupted sessions to continue from a checkpoint instead of starting over. The company has tested this system on more than 1,000 tool calls throughout a 24-hour GPU kernel engineering job, emphasizing its focus on long-running autonomous software engineering.

The launch also escalates the competition in the worldwide AI sector. Meta’s arrival allows software developers one more dependable marketplace contender in addition to Anthropic, OpenAI, and DeepSeek, putting pressure on AI suppliers to contend as far as coding skills and pricing models are considered. This could lead to quicker product launches and lower inference costs as well as greater availability of advanced AI programming tools that aid software engineering.

If you’re reading this, you’re already ahead. Stay there with our newsletter.