Meta's Muse Spark 1.3 Debuts No. 6 of 636 on AI Intelligence Index: Spark (SPK) in Focus

Meta Ships Muse Spark 1.3

Meta released Muse Spark 1.3, its proprietary frontier reasoning model, on September 2, shipping it through the Muse Code coding agent and the Meta Model API. The model carries a 1 million token context window, accepts text, image and video inputs, and places sixth among 636 models on the Artificial Analysis Intelligence Index, per the benchmark's own release page for the model. That placement matters because the index is the most widely watched independent yardstick for frontier model quality, and a top-six position puts Muse Spark 1.3 alongside the strongest closed models in production. Meta's internal comparisons against version 1.2 show roughly 20% fewer tool calls in agentic and coding workflows and about 25% lower token consumption, while long-context retrieval scores of 98.5 and 98.1 sit near the ceiling of the evaluation. The 1 million token window means the model can reference long documents, entire codebases and full conversation histories in a single pass. Developers can feed multiple files and task logs together, and agentic services retain earlier instructions and outputs across long multi-step jobs — a property that matters for autonomous workflows where one dropped instruction can derail an entire chain of actions. Meta has kept the Spark line behind its own API rather than publishing open weights, a deliberate contrast with Muse Glimmer, the open model the company shipped earlier that runs on a single graphics card. The xhigh variant is live in production, the Max version remains in limited preview, and detailed sub-scores for the xhigh configuration have not been disclosed beyond the published index position. For trackers of the Spark (SPK) token, this is the dominant development under the Spark name this session; no protocol-side announcement from the SPK project has accompanied the Meta release, and none should be assumed from timing alone.

Pricing and the Data Trade-Off

The commercial structure is where Muse Spark 1.3 becomes directly relevant to builders. The standard xhigh endpoint prices at $1.25 per million input tokens and $4.25 per million output tokens — identical to version 1.2 — with cached input at $0.15 per million tokens. Cache pricing at that level rewards repeated context reuse, which is exactly the pattern long-context agent workflows produce. Meta also operates a separate “contributor” tier at roughly $0.10 per million input tokens and $0.20 per million output tokens, a discount of 10 to 20 times with one condition: data sent through that endpoint may be used to train Meta's models. That two-tier split — confidentiality at standard rates, steep discounts in exchange for training data — is one of the industry's clearest expressions of the trade-off between cost and data acquisition. Enterprises handling sensitive source code, proprietary algorithms or customer records need to weigh the data-usage terms before the sticker price: a 20-fold cheaper endpoint that exposes proprietary logic is not cheap at all. The contributor structure also sharpens the strategic picture. Because Muse Spark 1.3 is API-only, users never download the weights; Meta operates the access point, controls rate limits and retains visibility into usage patterns. Open-weight releases let teams run models in their own environments and tune them freely, which is why infrastructure-sensitive organizations — exchanges, custodians and wallet providers managing every wallet address in their custody stack — typically treat API-only frontier models as a convenience tier, not a backbone. For high-volume agent workloads, the economics compound: an agent making thousands of tool calls per hour sees the per-call reduction and the per-token discount multiply rather than add. Readers tracking the market in real time can follow live spot and futures prices on Binance.

Cheaper Inference for Crypto Builders

COINOTAG's read is that the two threads — benchmark placement and pricing architecture — point at a single shift: frontier-grade reasoning is becoming dramatically cheaper per unit of work. For crypto development teams building exchanges, node infrastructure, on-chain analytics and security tooling, that changes experimentation budgets in a meaningful way, since agent-heavy workflows are precisely where the 20% tool-call and 25% token savings compound. Spark (SPK), whose spot price moved 13.8% over the past 24 hours, sits in the AI-agent token narrative that feeds on exactly this kind of headline — but the Meta release carries no confirmed SPK-specific impact, and positioning on name adjacency alone is textbook FOMO rather than a fundamentals-based thesis.