Mark Zuckerberg Unveils Muse Code in Beta, Amidst Competitive Disparities
On Wednesday, Mark Zuckerberg introduced Muse Code in its beta phase, marking Meta’s inaugural foray into the realm of artificial intelligence coding agents.
However, it appears that Anthropic’s Claude Opus 5 outshines Muse Code across all four comparative charts released by Meta at the time of launch.
Despite this, Meta proceeded to unveil the charts, seemingly opting to market an economically accessible tool rather than one that demonstrates superior performance.
Independent assessments suggest that the performance gap is more substantial than what Meta has disclosed publicly.
Releasing Muse Code in beta today. It’s a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results.
Powered by Muse Spark 1.2, a coding-focused model update. pic.twitter.com/xqavk41w6v— Mark Zuckerberg (@finkd) August 5, 2026
Muse Spark 1.2 underpins Muse Code and achieved a score of 82.9% on the Terminal-Bench 2.1 assessment.
In stark contrast, Claude Opus 5 garnered an impressive 86.7% on the same benchmark, developed by researchers from Laude Institute and Stanford, which comprises 89 distinct tasks covering system repairs, data manipulations, and security measures.
While securing second place is commendable, Muse Code did surpass OpenAI’s Codex, which scored 81.8%, and Grok Build at 81.6%.
However, the next comparative analysis was particularly telling. In the DeepSWE 1.1 evaluations, which encompassed 113 coding tasks with internet access disabled during scoring, Muse Spark 1.2 plummeted to a third-place finish at 59.3%.
Meta subsequently released its own assessment, derived from 440 authentic pull requests executed by its engineers, where Muse Spark 1.2 scored 70.6%, approximately nine points adrift of Opus 5.
The assessment reflects merely a 2.3-point improvement over the previous version, Muse Spark 1.1, which was dispatched in July.
Notably, Meta positioned itself against OpenAI’s GPT-5.6 Terra; however, they omitted OpenAI’s more potent model, Sol, from all three coding comparisons.
On the independent Terminal-Bench 2.1 leaderboard, Sol reigns supreme at 89.5%, followed closely by Opus 5 at 89.1%.
Both figures surpass the 86.7% purported by Meta for Opus 5 — an instance wherein the company deliberately selected a less formidable setting for its closest competitor.
When compared with Sol, the leading model, Muse Spark 1.2 actually trails by 6.6 points instead of 3.8.
In a specific GPU kernel task involving over 1,000 tool calls, Sol outperformed the baseline by 71.2%, whereas Muse Spark 1.2 registered an improvement of 68.7%, placing it fourth among six contenders.
A significant caveat exists; Muse Spark 1.2 has yet to feature on the public leaderboard, which only includes 26 of the 183 monitored models. Hence, its 82.9% rating remains a figure promoted by Meta.
“Muse Spark 1.2 is our next step as we push toward frontier, with larger, more capable models on the way,” Zuckerberg declared in a post.
Zuckerberg Poised to Partner with Competitor in AI Leasing Deal
Reports indicate that Meta is engaged in discussions to lease computation resources to Anthropic, an agreement that could culminate in a staggering $10 billion over a two-year period.
In such a scenario, Meta’s data centers would play a pivotal role in operating the Claude models that Muse Code endeavored to surpass.
The financial implications of Muse Code’s development were substantial, with Zuckerberg disbursing $14.3 billion in June 2025 to acquire Scale AI and its founder, Alexandr Wang, who now oversees Meta Superintelligence Labs.
Pricing remains the sole viable strategy for Wang. Costs align with the July introduction of Meta’s inaugural paid API, which is available at $1.25 per million input tokens and $4.25 per million output tokens.
An alternative contributor tier is offered at over ten times less; developers can qualify by permitting Meta to train on their code. Wang, however, did not disclose adoption statistics for the Muse Spark series.
Meta’s financial reports elucidate the rationale for the pricing strategy. Revenue surged by 28% to $60.8 billion last quarter, yet operating profit declined by 8% to $18.8 billion.
The operating margin diminished considerably, sliding from 43% a year prior to 31%. Notably, Meta allocated $31.08 billion toward capital endeavors in the past quarter alone, with projections reaching as high as $145 billion for the fiscal year.
Muse Code does provide engineering capabilities that Claude Code lacks, featuring background agents that maintain contextual awareness throughout a session. Moreover, sub-agents operate independently within distinct copies of a repository.
In essence, Meta has developed a respectable second-tier coding assistant, pricing it as a budget-friendly option. The beta phase will elucidate whether developers are willing to exchange marginal accuracy for a significantly reduced cost.
Source link: Kucoin.com.
