Meta Muse Spark 1.1: Meta's First Paid AI Model

Meta has done something it has never done before: put a price tag on its best model. On July 9, 2026, Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model built for agentic work, and made it available to developers through a new, paid Meta Model API. After years of shipping open-weight models for free under the Llama and Muse banners, Meta is now charging for access the same way Anthropic and OpenAI do — a strategic pivot as much as a product launch.

What Meta shipped

Muse Spark 1.1 is a closed-weights, multimodal model with a 1 million-token context window and, according to Meta, state-of-the-art performance on agentic benchmarks. The company positions it squarely at the tasks that define the current frontier: long-horizon tool use, computer use, coding, and multimodal understanding across images, video, and PDFs.

The headline capability is context management. Rather than treating its million-token window as a passive buffer, Muse Spark 1.1 actively curates it — deciding what to keep in active memory, what to retrieve from earlier in a session, and what to compact once a conversation runs long. Meta frames this as the difference between a model that has a large context window and one that can actually use it over a multi-hour task without losing the thread.

Meta also leaned into the AI agent story. The model supports parallel tool calling, structured output, and built-in search with citations, and it can act as an orchestrator — a main agent that delegates subtasks to parallel subagents. That is the architecture most serious agent frameworks have converged on in 2026, and Meta is now shipping it as a native model capability rather than something developers have to wire up themselves.

The benchmarks

Meta published results claiming Muse Spark 1.1 rivals GPT-5.5 and Opus 4.8 across agentic evaluations while being roughly 10x cheaper and twice as fast. On specific tests, Meta says the model leads MCP Atlas at 88.1 and Humanity’s Last Exam at 62.1, placing it ahead of Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on those particular measures. Meta also cited strong showings on JobBench and FinanceBench.

Vendor-published benchmarks always deserve a skeptical read — labs pick the evaluations that flatter them, and MCP Atlas (a test of how well a model drives tools over the Model Context Protocol) rewards exactly the agentic behavior Meta optimized for. But the direction is clear enough: Meta believes it has a model good enough to charge for, and it is willing to be measured against the frontier labs on their own turf. That is a notable change in posture from a company whose entire AI brand was, until recently, “free and open.”

The pricing is the strategy

The Meta Model API launched in public preview for US developers on July 9, priced at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for every new account before pay-as-you-go kicks in. You can also use the model for free in Thinking mode inside the Meta AI app and at meta.ai with a Meta login.

Put those numbers next to the field and the intent is obvious. Just this week, OpenAI launched GPT-5.6 in Sol, Terra, and Luna tiers, with Terra at $2.50/$15 and Luna at $1/$6. Anthropic has been running Sonnet 5 at promotional $2/$10 pricing. Meta’s $1.25/$4.25 undercuts the mid-tier of both while claiming near-flagship performance. It is the same playbook Chinese labs have used to pressure Western incumbents — lead with capability-per-dollar — now run by a US hyperscaler with effectively unlimited compute.

Meta jumps into the coding market

The launch also marks Meta’s formal entry into the AI coding market, the most lucrative and competitive slice of the API business. Muse Spark 1.1 can autonomously write and debug software, interact with external coding environments, and — Meta emphasizes — is particularly strong at frontend and design work. That puts it in direct competition with the tools and models that have defined the state of AI coding assistants in 2026, a category Anthropic and OpenAI have dominated.

Coding is where API revenue concentrates because coding agents burn tokens: long files, long tool traces, long reasoning chains. A model that is genuinely competitive on coding quality and meaningfully cheaper per token is the kind of thing that moves a procurement decision. Meta knows this, which is why the pricing and the coding pitch arrived together.

Why Meta is doing this now

For most of the past two years, Meta’s AI strategy was to commoditize the model layer — release capable open weights, deny rivals a pricing moat, and keep the value in Meta’s own apps and ad business. Charging for a closed-weights model is a reversal of that logic, and it reflects two pressures.

First, frontier models have gotten genuinely expensive to train and serve, and giving away the best one starts to look like leaving money on the table when competitors are booking billions in API revenue. Second, Meta has spent heavily to rebuild its AI organization and needs to show that spending produces a product the market will pay for, not just internal features. Reports through the first half of 2026 painted a picture of restructuring and uneven progress inside Meta’s AI efforts; a paid frontier API is the clearest possible signal that the labs are shipping again.

It is worth noting what Meta did not do. Muse Spark 1.1 is closed-weights — there is no download, no open license, no self-hosting. Meta has effectively split its lineup: open models to keep the ecosystem tilted its way, and a closed flagship to compete for revenue. That is the same two-track approach several labs have adopted as open-source models close the gap on the mid-tier while the frontier stays proprietary.

What it means

Muse Spark 1.1 turns a two-horse API race into a genuine price war. With Meta pricing a frontier-class agentic model below the mid-tiers of both GPT-5.6 and Sonnet 5, the pressure on per-token economics is now structural, not promotional. The labs have been counting on developers to accept premium pricing for premium capability; Meta is betting that “90% of the frontier at a fraction of the cost” wins the majority of real workloads.

Who wins: developers and enterprises, unambiguously. A credible frontier agentic model at $1.25/$4.25 with a self-managing million-token context is a strong default for agent and coding workloads, and it hands buyers leverage in every negotiation. It also validates the shift enterprises have been making from “tokenmaxxing” toward cost-efficient inference — cheaper capable models make that discipline easier to sustain.

Who feels it: the pure-play labs. OpenAI and Anthropic can out-spend Meta on almost nothing, but they cannot out-subsidize a company whose model business does not need to stand alone to justify itself. For labs racing toward public markets — Anthropic and OpenAI both filed confidentially for IPOs in June — a subsidized hyperscaler competitor pressuring API margins is exactly the risk factor public investors will scrutinize.

What to watch: whether Muse Spark 1.1’s real-world quality holds up outside Meta’s chosen benchmarks, especially on the long-horizon coding tasks where reliability matters more than headline scores; how fast Meta moves the API out of US-only preview; and whether OpenAI, Anthropic, or Google respond with price cuts of their own. The direction of travel is set: in the second half of 2026, capability is table stakes and price is the battlefield.