Muse Spark

“You can now buy it by the token on a public-preview API, but not license or self-host it — Muse Spark is still a strategy update first.”

The purchasing decision is new and preview-grade: Meta Model API has been self-serve since 2026-07-09 at $1.25 / $4.25 per 1M tokens, without GA terms. If your organization standardized on Llama for self-hosted AI, Muse Spark tells you Meta's frontier investment has moved somewhere you cannot follow: Llama 4 is now the effective end of the open line, Behemoth never shipped weights, and "Llama 5 does not exist." Plan open-weights roadmaps around GLM, Qwen, Nemotron, and Meta's much smaller Muse Glimmer 30B instead, and track the Muse API preview toward GA. The public-preview structure suggests Meta is still testing enterprise appetite; a rate card now exists, GA terms do not.

“The most strategically legible release of 2026: Meta stopped subsidizing the ecosystem and started competing for the frontier itself.”

Muse Spark reorders the market map. Meta spent four generations commoditizing its rivals' complement with open weights; MSL's first act reverses the strategy — closed weights, consumer-first distribution through a three-billion-user funnel, and hardware differentiation via Ray-Ban glasses that no competitor can match. The compute-efficiency claim (Maverick parity at a tenth-plus less compute) is aimed at investors as much as researchers. The vacuum it creates in open weights is already being filled by Chinese labs and NVIDIA, which changes the geopolitics of the open ecosystem. Distribution is the moat here, not benchmarks — and it's a very good moat.

“Free for consumers and $1.25/$4.25 per 1M tokens for builders, with no long-context premium — a real rate card at last, if only a preview one.”

For budget purposes the row now has a rate card: $1.25 input, $4.25 output, and $0.15 cached input per 1M tokens on the Standard tier, with no long-context premium, plus a Contributor tier at $0.10 / $0.20 if Meta may train on your data. No batch pricing, enterprise tier, or committed-use discounts are published. The consumer-free positioning is financed by Meta's ad business, not by you. The finance-relevant angle is indirect: if you built cost models around free Llama weights continuing to improve, that assumption just expired — self-hosted TCO projections should now assume Meta's open line is Llama 4 plus the small Muse Glimmer 30B, with the open-weights alternatives (GLM-5.2, Nemotron) as the new baseline curves. Value score reflects consumer value (high) against enterprise value (zero when scored, before the API rate card).

“Tool use, orchestration, visual chain of thought — now behind an OpenAI- and Anthropic-compatible endpoint, as long as preview terms suit you.”

From a builder's chair the shop is now open, in preview. The capabilities Meta describes — parallel multi-agent Contemplating, visual chain of thought, native tool use — are exactly the primitives agent developers want, and since 2026-07-09 Meta Model API documents them: tool calling, structured output, search grounding, a 1M-token context, and published rate limits (3,000 RPM / 4M TPM on the Standard tier), with endpoints that work with the OpenAI and Anthropic SDKs. What can be assessed: the consumer product is polished, multimodal interactions feel genuinely native rather than stitched, and the reasoning mode produces visibly deliberate answers. Practitioners should build against Meta's own docs at dev.meta.ai rather than third-party tutorials, and plan for preview-stage changes.

“The best free AI on the planet right now — frontier reasoning in my pocket and on my face, for exactly zero dollars.”

For daily consumer use, Muse Spark is a step change: frontier-tier reasoning, image understanding that genuinely reasons over what it sees, and voice interaction, free, inside apps billions already have. Contemplating mode on a hard question produces answers competitive with paid frontier subscriptions. The Ray-Ban integration is the sleeper feature — hands-free multimodal AI that works — and the June glasses launch made it the first frontier model people wear. Frictions: no power-user controls in the apps (no system prompts, no conversation export to speak of; the API is the escape hatch), Meta-grade content policies, and the knowledge that your usage feeds Meta's flywheel. As a free product it's extraordinary; as a power tool it's a walled garden.

“Two cherry-picked numbers in best-case mode, everything else undisclosed — and the everyday index says the ceiling isn't the product.”

The claims are carefully constructed. HLE 58% and FrontierScience 38% are quoted exclusively in Contemplating mode — the compute-maximal configuration — while the independently measured everyday index (AA 43 in early July 2026, before AA rescaled it) sat mid-pack among 2026 flagships, below GLM-5.2's 51 on the same date. The order-of-magnitude efficiency claim is unverifiable without disclosed compute budgets. Architecture, parameters, cutoff: all withheld, from the company that used to publish everything. The GPQA/MMMU-Pro numbers circulating in coverage trace to no Meta source. What survives scrutiny: the HLE result is on Meta's letterhead, the safety report is real and substantive, and the consumer product demonstrably works at scale. Muse Spark is real and good; the framing wants you to believe it is the frontier leader, and the verifiable record says "frontier-adjacent, brilliantly distributed."