On September 2, 2026, Meta introduced Muse Spark 1.3 as its largest performance jump yet in the Spark family, available in Muse Code and Meta Model API with a “max reasoning” mode. The official blog emphasized long-horizon agentic and coding tasks: fewer unnecessary turns, cleaner style and, per internal comparisons, about 20% fewer tool calls and 25% fewer tokens versus Muse Spark 1.2.
Meta Chief AI Officer Alexandr Wang told Bloomberg and downstream outlets that 1.3 is “competitive” with Anthropic’s Claude Fable 5.1 and “better than” OpenAI’s GPT-5.6 Sol at coding, while also outperforming current Chinese models on his claim. Neural Pulse labels those statements as executive positioning—useful signal, not an independent verdict.
An Artificial Analysis evaluation reported in public coverage gave Muse Spark 1.3 an Intelligence Index score of 62, placing it behind Claude Fable 5.1 and Claude Opus 5 while ahead of some OpenAI models in that cut. The index aggregates multiple benchmarks; without a full open table it is directional evidence, not a definitive market ranking.
The question developers care about most is not only ELO: will there be weights? Meta has not decided whether to publish 1.3’s. A roadmap item still lists releasing 1.2’s weights, but by mid-September that promise had not materialized. The first Muse Spark arrived closed in April. Without weights, the ecosystem depends on paid API tiers (cheaper “contributor” rates if data use is allowed) and Muse Code.
European law enters here. Article 53 of the AI Act partly exempts genuinely free and open-source general-purpose models from certain technical-documentation duties toward the AI Office and downstream developers—but that exemption switches off if a model is classified as carrying systemic risk. Frontier plus open source is not “no duties”: copyright policy and a public training-content summary still apply.
To qualify as open source for the exemption, parameters (weights), architecture information and usage information must be public, without non-commercial clauses or user thresholds that gut openness. Meta knows the terrain: it has already been skeptical of parts of the European code of practice. The 1.3 weights decision is therefore also a compliance and “open” narrative decision.
Published Meta pricing illustrates the business model: on the order of $1.25/$0.15/$4.25 per million tokens (input/cache/output) on the standard path, versus $0.10/$0.002/$0.20 on the contributor path. That anchors Spark 1.3 in the frontier API market, not in the downloadable-weights market of earlier Llama years.
The contrast with the Muse consumer agent (covered in prior Neural Pulse cycles) is deliberate in this piece: the new angle is the 1.3 model, claims versus Fable/Sol, undecided weights and the Article 53 collision. We are not reopening the agent VM as the main story.
For European CIOs, the checklist is prosaic: is the model offered in the EU? Is there an authorized representative? Is there a training summary? Is there a systemic-risk evaluation? Opening weights might ease part of Article 53… or not, if the Commission treats the model as systemic. Closing weights simplifies commercial control and complicates Zuckerberg’s open narrative.
Neural Pulse fixes facts as of September 15, 2026: September 2 launch; internal efficiency metrics vs 1.2; Wang’s claims vs Fable 5.1 and GPT-5.6 Sol; Artificial Analysis score 62 per media; undecided 1.3 weights; 1.2 still unreleased; Article 53 framework with a limited open-source exception that excludes systemic-risk models. We do not claim universal parity or superiority on every axis.
The next catalyst is binary: a weights announcement (1.2 or 1.3) or prolonged silence while Watermelon—the larger still-undated model—absorbs internal attention. Meanwhile Spark 1.3 competes in the agentic API market with one hand tied behind its open-source back… or protected, depending on your view.
In short: Meta says it closed the coding gap with Anthropic/OpenAI; Europe replies that an “open” label does not erase systemic-risk duties; developers wait for a weights tarball that may never arrive. That triple tension defines Muse Spark’s mid-September state.
Contributor versus standard pricing reveals another tension: the cheap path subsidizes product improvement with user data; the expensive path promises not to use that data. In Europe, that duality interacts with GDPR and AI Act transparency duties; a discount for telemetry does not erase systemic-risk obligations or training-summary duties.
Watermelon, the still-undated larger model, lurks as a hidden variable: if Meta saves its best for a later closed launch, Spark 1.3 can be both Meta’s best open-ish API and a strategic decoy. Developers should plan for portability—own evals, provider abstraction—rather than marry a single September week’s claims.
