Meta Unveils Muse Spark 1.3: A High-Performance Leap in Agentic AI

Meta has officially pulled back the curtain on its latest artificial intelligence advancement, Muse Spark 1.3, an iteration that marks a significant milestone in the company’s ongoing push to dominate the coding and agentic AI landscape. Unveiled yesterday, the new model represents a substantial performance upgrade over its predecessor, Muse Spark 1.2, demonstrating faster processing speeds and superior outcomes on third-party benchmarks. However, the release arrives with a nuanced caveat regarding deployment availability and the evolving nature of Meta’s commitment to open-weight models.

"Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter," Meta co-founder and CEO Mark Zuckerberg declared in a post on X. Zuckerberg highlighted the release as Meta’s most significant leap to date in the domains of coding and complex agentic tasks—workloads that require the AI to not just generate text, but to autonomously plan, execute, and refine multi-step digital processes.

A New Standard for Performance and Utility

The claims made by Meta’s leadership are backed by tangible improvements. Muse Spark 1.3 demonstrates significant gains over the 1.2 release from last month, particularly in its ability to handle long-running agentic tasks. For developers and enterprises, this version stands out as one of the most compelling price-performance offerings currently available near the upper echelons of independent model rankings.

However, the full potential of this release is tiered. Meta’s most impressive benchmark results are tied to the "max" reasoning configuration of the model. According to company communications, this version is currently undergoing final safety testing and is expected to be released "shortly." Independent benchmarking firm Artificial Analysis confirmed that it evaluated the "max" configuration during a limited partner preview, though currently, no major API provider offers access to this specific, high-reasoning variant.

For the time being, the version rolling out across the Muse Code harness and the Meta Model API utilizes the company’s previously established reasoning settings, including the "xhigh" configuration. This reality creates a critical distinction for enterprise architects: the question is not whether Muse Spark 1.3 can reach the current frontier of AI capability, but rather how closely the accessible, deployable model mirrors that performance and what the real-world operational costs will be for companies looking to integrate it into their production environments today.

Evaluating the Shipping Model

While Meta has been transparent by disclosing results for both the "xhigh" and "max" configurations in its internal evaluation reports, the company’s marketing materials have prominently showcased the "max" variant, where the most striking score increases are found.

In terms of raw metrics, Meta reports that the "max" configuration achieves a GDPval-AA v2 score of 1,754, compared to 1,709 for the "xhigh" version. Similarly, on the OSWorld 2.0 evaluation, the "max" variant scores 66.9 against 57.2 for the shipping model, and it holds a 64.9 score on JobBench compared to 61.2. Yet, the performance gap is not uniform across all domains. On some evaluations, such as DeepSearchQA, the models are essentially tied at 89.4, and the "xhigh" version even manages a slight lead on Terminal-Bench 2.1 with a score of 89.2 versus 88.8 for the "max" configuration.

Artificial Analysis, which tracks the global landscape of AI models, currently assigns the Muse Spark 1.3 "max" configuration an Intelligence Index score of 62, while the shipping "xhigh" version lands at 61. This places the shipping model in direct competition with top-tier peers such as GPT-5.6 Sol max, Grok 4.6 high, and Claude Opus 5 high. While Anthropic’s offerings, particularly the Claude Fable 5.1 and Opus 5 models, continue to occupy the very top of the leaderboards, it is clear that Muse Spark 1.3 "xhigh" is now firmly established within the "frontier cluster."

This represents a notable shift from the previous month. When VentureBeat analyzed the launch of Muse Spark 1.2, it was clear that Meta had produced a credible contender that nonetheless consistently trailed behind Anthropic’s best. With the 1.3 update, Meta is no longer merely a participant in the race; it is now actively trading wins with both OpenAI and Anthropic across various coding and agentic evaluations.

Beyond raw intelligence, Meta has focused on usability. The underlying model has been refined to maintain multiple concurrent workflows within long, complex threads. It is now more adept at gathering context through external tools, identifying gaps in its own strategic plans, and proactively seeking clarification from users before initiating consequential actions. Internal data from Meta’s engineering team suggests that these behavioral improvements have led to a 20% reduction in tool calls and a 25% decrease in token consumption during typical coding tasks. For enterprises managing millions of agentic loops, these efficiency gains may prove more valuable than incremental improvements on a leaderboard.

The Economics of "Almost Too Cheap to Meter"

Despite the ambitious rhetoric surrounding the release, Meta has not implemented a broad price cut for its API. The Standard pricing tier remains identical to that of Muse Spark 1.2: $1.25 per million input tokens and $4.25 per million output tokens, with cached input tokens priced at $0.15 per million. Consequently, Zuckerberg’s assertion that the model is "almost too cheap to meter" should be interpreted not as a claim of reduced token pricing, but as a statement of faith in the productivity developers can extract from those tokens.

Meta says Muse Spark 1.3 has frontier performance — but its best results come from a model developers can’t broadly use yet

Artificial Analysis provides data that both supports and complicates this view. While the "xhigh" model generates 235.2 output tokens per second, the estimated cost per task is approximately $0.55. At an Intelligence Index of 61, this gives it the lowest cost-per-task ratio of any currently measured model at that performance tier. However, this also reveals an interesting trend: despite the per-token price remaining static, the cost to complete an average task has actually increased compared to the previous generation. Artificial Analysis attributes this primarily to heavier input-token consumption required for more sophisticated, complex agentic evaluations.

This does not necessarily contradict Meta’s internal data regarding token efficiency, as the two entities are measuring different types of work. Meta is focusing on its own specific coding workflows, whereas third-party evaluators are testing a broader, more taxing suite of reasoning and agentic behaviors. This highlights the inherent complexity of AI economics: when models function as agents, the cost of an operation is no longer just a function of token rates. It becomes a combination of reasoning effort, the number of turns required, the frequency of tool calls, and the necessity for retries.

Additionally, Meta continues to offer its "Contributor" tier, priced at $0.10 per million input tokens and $0.20 per million output tokens. This remains a highly attractive option for prototyping, provided the user agrees to allow Meta to utilize their prompts and completions for model training. This, of course, introduces a distinct set of data-governance considerations for any enterprise handling sensitive proprietary code or confidential internal information.

Executive Competition and Strategic Shifts

The intensity of the current AI landscape was underscored by Alexandr Wang, Meta’s chief AI officer. Following the publication of the independent benchmark results, Wang took to social media to post a pointed, "Gemini who?" in response to the data.

The jab was particularly sharp because it coincided with Google’s release of Gemini 3.8 Flash, which is aimed at the exact same category of workloads—long-horizon software engineering and autonomous agents. Google has framed 3.8 as its most capable reasoning and coding model in its "Flash" series.

The competitive landscape is incredibly tight. While Meta’s Muse Spark 1.3 "xhigh" holds a slight lead in intelligence and cost-per-task metrics according to Artificial Analysis, Google’s Gemini 3.8 Flash offers significantly higher output throughput—approximately 30% faster—and a lower raw sticker price during its current promotional period. However, that promotional pricing is set to expire at the end of the year, at which point Google’s costs will increase substantially.

Ultimately, for an enterprise looking to select a model today, the choice is far from binary. Gemini offers superior speed for high-volume tasks, while Muse provides a slightly more robust, high-effort agentic experience based on current independent benchmarks.

Perhaps the most lingering question for the developer community concerns Meta’s broader philosophy regarding open weights. When Meta launched the Muse ecosystem in August, the move toward a proprietary, API-first strategy was a notable departure from the company’s previously lauded open-weight approach, which had been instrumental in the widespread adoption of the Llama family of models.

While Meta did pivot shortly after that launch to release the 30-billion-parameter Muse Glimmer model under an Apache 2.0 license, and had indicated plans to open the weights for the Spark 1.2 model, the arrival of Muse Spark 1.3 as a proprietary model has left the roadmap ambiguous. Meta’s current communication refers to "the Muse Spark open weights release" without identifying specific versions, release dates, or licensing details.

For many teams that standardized their infrastructure around Llama specifically for the control offered by self-hosting and local customization, this lack of clarity is significant. Muse Spark 1.3 undeniably proves that Meta possesses the capability to iterate at an industry-leading pace, delivering a fast and highly competitive model. However, the next true test for Meta will be its ability to translate that development speed into a coherent, predictable roadmap that allows enterprises to plan their future infrastructure with confidence—and to fulfill its outstanding commitment to delivering the promised open-weight versions of its most advanced technology.

Share:

Siti Muinah writes for Tech Maze.