Muse Spark 1.3 Tops OpenCode Daily Usage
Muse Spark 1.3 topped OpenCode daily token usage two days after launch, converting Meta's Contributor strategy and massive AI spending into live adoption.
Table of Contents
DeepSeek's reign atop the OpenCode leaderboard has ended. The model that unseated it is Meta's newly released Muse Spark 1.3, launched just two days earlier. On September 4, OpenCode founder dax posted a 22-second leaderboard video announcing that “Meta Muse Spark has dethroned DeepSeek.”
Meta Chief AI Officer Alexandr Wang quote-posted the clip with a single reaction: “A M E R I C A.” According to OpenCode Data, Muse Spark 1.3 processed 2.7 trillion tokens on September 4 as of 1:54 p.m. UTC, surpassing DeepSeek V4 Flash's 1.9 trillion. It marks the first time an American model has taken first place on the platform.
Muse Spark 1.3 Takes First in an Arena Dominated by Low-Cost Rivals
The milestone carries weight because OpenCode is not Meta's proprietary ecosystem. OpenCode Zen evaluates model-and-provider pairings before selecting those best optimized for coding agents. Its leaderboard is dense with price-competitive alternatives, including GPT-5.6 Luna, DeepSeek V4 Flash, GLM 5.3, Kimi K3, and MiMo-V2.5.
Cumulative token metrics illustrate how steep that competition is. In recent tallies, DeepSeek V4 Flash led overall volume with 25 trillion tokens, followed by Muse Spark 1.2 Contributor at 13 trillion and MiMo-V2.5 at 10 trillion alongside GLM and Kimi variants. Muse Spark 1.3 accumulated 6.1 trillion tokens in its first two days before claiming the daily number-one spot.
This surge was not the result of a captive audience. Developers actively evaluate cost against latency and performance, switching providers with minimal friction. For the first time, an American frontier model has captured substantial volume in a coding-agent market previously dominated by Chinese cost-efficient models.
Contributor Pricing: Low Cost Upfront, Training Data in Return
Meta's distribution strategy laid the groundwork for this rapid ascent. The company split access into two tiers—Standard and Contributor—steeply discounting rates for developers who allow their interactions to be used for model training. OpenCode lowered the threshold further by making Muse Spark 1.3 Contributor free for a promotional window.
| Tier | Input per 1M tokens | Output per 1M tokens | Data terms |
|---|---|---|---|
| Standard | $1.25 | $4.25 | Not used for model training |
| Contributor | $0.10 | $0.20 | Prompts and completions may train future models |
| OpenCode Zen | Free | Free | Contributor terms apply |
Meta's pricing documentation defines this arrangement clearly. Prompts and outputs from Standard users are excluded from training pipelines, whereas the Contributor tier grants Meta permission to use that data for future models. Developers receive near-free access to a high-end coding model, while Meta collects dense telemetry and completions from active software engineering.
The policy acts less like a temporary discount and more like a distribution and data flywheel. OpenCode recorded 50,000 unique users and 1.15 million completed sessions for Muse Spark 1.3 Contributor, with 92% of input tokens handled through prompt caching. By removing adoption barriers, Meta secured instant market share while generating high-value training data for future model generations.
$145 Billion and Alexandr Wang: Meta's AI Spending Translates Into Product
This breakthrough was the product of deliberate structural investments. In June 2025, Meta acquired a 49% stake in Scale AI for $14.3 billion and appointed founder Alexandr Wang to lead Meta Superintelligence Labs. An acquisition initially scrutinized for its premium valuation has, just over a year later, produced a flagship model driving developer workflows.
The scale of Meta's underlying infrastructure spending was even larger. The company projected 2026 capital expenditures of $130 billion to $145 billion, having deployed $31.08 billion in the second quarter alone. Backed by massive facilities like the Hyperion data center and an overhauled training stack, the original Muse Spark delivered performance comparable to Llama 4 Maverick at less than one-tenth the compute footprint.
Release velocity matched the pace of capital deployment. After the initial release in April, Meta shipped version 1.1 in July, 1.2 in August, and 1.3 in September, marking four major iterations in five months. Meta reported that version 1.3 reduces tool calls by roughly 20% and token usage by 25% relative to 1.2, turning engineering efficiency into tangible productivity gains.
Developer Adoption, Not Benchmarks, Propels Meta to Number One
Low cost alone cannot account for 1.15 million completed sessions. Within hours of release, developers posted a Lies of P mechanical-heart demo powered by multi-image inputs alongside a Counter-Strike recreation built for $1.75. Alexandr Wang amplified both demonstrations, reflecting how rapidly practitioners embraced the model's practical output.
The outcome demonstrates Meta's research, infrastructure, and distribution operating in tandem. Live developer traffic and sustained token throughput validated the architecture as it overtook DeepSeek within two days. As Meta expands into consumer coding agents, Muse Spark 1.3 provides concrete usage metrics demonstrating the strength of its internal models.
Meta had long faced investor skepticism that heavy AI spending produced ambitious roadmaps rather than market-leading products. That capital has now become specialized compute, new leadership has delivered an efficient architecture, and Contributor distribution has put it into active developer workflows. The 2.7 trillion daily tokens are the first clear evidence that this investment system works in the market. What now bears watching is whether Muse Spark 1.3's early surge turns Meta's long AI bet into durable returns.
