Meta Muse Spark 1.1 Review (2026): The 75%-Cheaper Frontier API, and When It Beats GPT and Claude

Meta's first paid frontier model runs $1.25 per million input tokens and $4.25 per million output, roughly a quarter of what OpenAI and Anthropic charge for their top tiers. Muse Spark 1.1 leads every tool-use benchmark it was tested on and still trails Claude Opus 4.8 on the hardest coding, the exact thing its name is selling.

The verdict

Muse Spark 1.1 is the best value in frontier-class tool use right now, and it is not the model to reach for when the job is hard, long-horizon coding. Those two facts sit side by side, and the marketing only tells you the flattering half.

The price is real. At $1.25 in and $4.25 out per million tokens, an output-heavy agent loop that costs $30 per million on GPT-5.6 Sol costs $4.25 on Muse Spark, about seven times less. On Meta's launch benchmarks it tops MCP Atlas and JobBench, the two scaled-tool-use tests, by wide margins over Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. If your product spends most of its tokens orchestrating tools, calling MCP servers, and coordinating subagents, this is the cheapest way to run a frontier-tier model at that job today.

The catch is what "Spark" implies. Meta named the line for coding and puts most of its marketing weight there, yet on independent coding benchmarks it lands mid-pack: third on agentic terminal coding, third on long-horizon agentic coding, and behind Opus 4.8 on diverse software engineering. It is also closed-weight, text-only on output, and available in public preview to US developers only, behind a waitlist.

Switch to it now if you are shipping tool-use or agent-orchestration workloads at volume and the per-token bill is your constraint. Stay on Opus 4.8 or GPT-5.6 Sol if your agents do sustained autonomous coding where a wrong step is expensive. Below, the pricing decoded, the benchmarks Meta's post underplays, and the one rule that decides it.

What Muse Spark 1.1 actually is

Meta AI's Muse Spark 1.1 is the second model from Meta Superintelligence Labs, released July 9, 2026, and the first Meta model outside developers can pay to use through an API. It is a multimodal reasoning model built for agentic work: it plans multi-step tasks, drives external tools, and delegates to subagents rather than just answering a single question. It ships with a 1,048,576-token context window (a full million tokens) and actively compacts earlier work to keep room for the steps it needs later. Much of its pitch runs through MCP servers, the open standard that lets a model call outside tools and data sources through one interface, so a single agent can drive a spreadsheet, a browser, and an internal API without custom glue for each.

Two things about it break from Meta's past. First, it is proprietary and closed-weight, meaning you cannot download or self-host it, unlike the open Llama family that made Meta's name in open models. Zuckerberg was blunt about the shift: "Since this is not an open source model, this is I think the first time that we're doing a real serious API." If you need open weights to self-host or fine-tune, Muse Spark is out, and the best open-weight models are a different shortlist. Second, consumer access is free: it runs in "Thinking" mode inside the Meta AI app and at meta.ai with a Meta login, so anyone can try the model before deciding whether the paid API earns a place in a build.

One spec matters more than the marketing: the model accepts text, image, and audio input but returns text only. There is no image or audio generation here. That is a separate Meta line (Muse Image and Muse Video). Muse Spark is a brain for agents, not a media generator, and reading it any other way sets up the wrong expectations.

On the agent side, the capability list is genuine: planning mode, goal conditioning, subagent delegation, and context compaction, plus a claim that it zero-shot generalizes to new native tools, MCP servers, and custom skills without per-integration tuning. Early partners vouch for the direction. Saoud Rizwan, CEO of Cline, said Meta is "clearly building for serious agentic coding, strong tool use at a price point that makes it viable to run real coding workloads at scale." Replit's Amjad Masad praised its "top-tier coding abilities (particularly frontend and design)" and its OpenAI-compatible packaging. Hold those quotes against the independent benchmarks below, because the numbers are more mixed than the endorsements.

The pricing, and the "75% cheaper" claim decoded

The Meta Model API charges $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits on every new account. Cached input drops to $0.15 per million, and the built-in web-search grounding tool runs $2.50 per 1,000 queries. The interface is OpenAI-SDK compatible and supports structured output, parallel tool calling, and prompt caching, so wiring it into an existing stack is close to a base-URL swap.

"Roughly a quarter of what rivals charge" is accurate for the frontier tier. Here is the ladder against the models it is pitched against, with today's numbers.

Read the top of that table and the claim holds: Muse Spark's output token is 5.9 times cheaper than Opus 4.8 and 7 times cheaper than GPT-5.6 Sol, while its input token is a flat 4 times cheaper than both. That is where the 75%-off headline comes from, and for a frontier-class model it is a real break in the price curve.

The bottom of the table is where the real nuance lives. Muse Spark is not the cheapest model on the market. Its $1.25 input token is higher than Claude Haiku 4.5's $1 and GPT-5.6 Luna's $1, and its output undercuts both of those budget tiers only slightly. So the honest framing is not "the cheapest AI API." It is frontier-class capability priced near the budget tier on output. If your workload is short-context extraction or classification you can retry cheaply, Haiku 4.5 or Luna still win on raw input cost and you do not need Muse Spark's ceiling. The value shows up specifically when you want a frontier-grade agent brain and your bill is dominated by output.

The benchmarks Meta's own post underplays

Meta's launch post leans on internal evaluations like its Meta Internal Coding Bench, which no one outside Meta can reproduce. The useful numbers are the third-party comparisons, and they tell a sharper story than either the marketing or the partner quotes. The tested OpenAI model here is GPT-5.5; GPT-5.6 Sol carries the same $5/$30 rate, so the pricing comparison above holds either way.

The pattern is clean once you group the rows. On the four tool-use and reasoning-with-tools tests, Muse Spark 1.1 leads three outright and misses the fourth by six tenths of a point to Opus. Its JobBench margin over Gemini (54.7 to 15.9) and over GPT-5.5 (to 38.3) is not close. This is a model tuned to plan and orchestrate across many tools, and the numbers back the design.

On the three coding rows, the story inverts. It sits third on Terminal-Bench 2.1 (80.0, behind GPT-5.5's 83.4 and Opus 4.8's 82.7) and third on DeepSWE 1.1, the long-horizon test, at 53.3 against GPT-5.5's 67.0 and Opus 4.8's 59.0. The one coding win is SWE-Bench Pro, where its 61.5 beats GPT-5.5 but still trails Opus 4.8's 69.2. DataCamp's read matches the table: its long-horizon agentic work "is still weak compared to GPT-5.5 and Opus 4.8." For a line named "Spark" and marketed on coding, that is the honest asterisk. It is a strong coder that beats one frontier rival and loses to the other, wrapped in tool-use scores that genuinely lead the field.

Against Gemini 3.1 Pro specifically, which is the comparison people keep asking, it is not a fair fight on tools: Muse Spark leads every tool-use row, and DeepSWE (53.3 to 12.0) is a rout. Gemini's case has to be made on price and ecosystem, not on these tasks.

When Muse Spark 1.1 is the right call, and when it is not

The decision is not "is it good." It is good. The decision is whether your specific workload lands in the zone where its lead and its price overlap.

The explicit rule: if your token bill is dominated by output and your agents spend their time using tools rather than writing large, novel codebases, Muse Spark 1.1 is the switch. If a single wrong step in a long autonomous run is expensive, pay for Opus 4.8 or Sol and keep Muse Spark for the cheaper, tool-heavy parts of the pipeline. Many production systems will end up routing: Muse Spark for the high-volume orchestration, a top coder for the hard implementation calls. Its OpenAI-compatible API makes that routing a config change, not a rewrite.

FAQ

Is Muse Spark free?

Consumer use is free: Muse Spark 1.1 runs in Thinking mode inside the Meta AI app and at meta.ai with a Meta login, subject to server-side rate limits. The Meta Model API is paid, at $1.25 per million input tokens and $4.25 per million output, with $20 in free credits when you sign up.

Is Muse Spark 1.1 open source?

No. It is proprietary and closed-weight, unlike Meta's open Llama family. Zuckerberg called the paid API Meta's first "real serious API," and the model is served only from Meta's own properties, not released for self-hosting.

Is Muse Spark better than Gemini 3.1 Pro?

On tool use, decisively: it leads MCP Atlas (88.1 to 78.2) and JobBench (54.7 to 15.9), and crushes Gemini on long-horizon coding (DeepSWE 53.3 to 12.0). Gemini's case rests on price and ecosystem, not these tasks.

Is Muse Spark 1.1 good for coding?

Good, not best. It beats GPT-5.5 on SWE-Bench Pro (61.5 to 58.6) but trails Claude Opus 4.8 there (69.2), and it lands third on both Terminal-Bench 2.1 and long-horizon DeepSWE. Strong for tool-driven coding, weaker for sustained autonomous work.

What can Muse Spark 1.1 do?

It plans and orchestrates agentic work across external apps, tools, and MCP servers, delegates to subagents, and manages a 1M-token context window with compaction. It accepts text, image, and audio input and returns text only.

Working out which models belong in your stack, and where the cheap tier actually pays, is the whole game now. The free AI Tools Map for business owners lays out which model fits which job without the hype. Want the sharpest picks in your inbox as they ship? Join the newsletter.