Quick answer. Muse Spark is Meta's first proprietary, closed AI model — a multimodal reasoning LLM built by Meta Superintelligence Labs (led by Alexandr Wang). It launched April 8, 2026, and marked the end of Meta's open-weight Llama era. The July 9, 2026 Muse Spark 1.1 update opened Meta's first paid, OpenAI-compatible API at $1.25 per million input tokens and $4.25 per million output tokens — roughly a quarter of what OpenAI and Anthropic charge. It's free for consumers inside Meta AI. The story isn't "smartest model"; it's aggressive price plus strong agentic tool-use, with pure coding as a known weak spot.
For years, Meta's AI strategy was open weights: the Llama models anyone could download and run. That era is over. On April 8, 2026, Meta Superintelligence Labs launched Muse Spark — Meta's first closed, proprietary frontier model — and VentureBeat summed up the shift bluntly: "Goodbye, Llama." Three months later, the Muse Spark 1.1 update opened a paid developer API and dropped Meta into the AI price war with rates that undercut everyone.
Here's what Muse Spark actually is, what it costs, how it benchmarks against the current frontier, and where it fits — with vendor claims and independent numbers kept clearly separate, because Meta has a benchmark-credibility history worth remembering.
What is Muse Spark?
Muse Spark is a natively multimodal reasoning model from Meta Superintelligence Labs (MSL) — the division Meta stood up in 2025 under Alexandr Wang (the former Scale AI CEO, now Meta's Chief AI Officer) following Meta's multibillion-dollar investment in Scale AI. "Muse" is the series; "Spark" is the first model line, with larger variants reported to be in development.
The strategically important part isn't the architecture — it's that Muse Spark is Meta's first proprietary, closed model with no open weights. Meta built the Llama brand on downloadable, open-weight models; Muse Spark is a hosted service you access through Meta's products and API instead. Meta positions it as "personal superintelligence" for a "people-first" era, purpose-built to run across the apps billions already use — Instagram, WhatsApp, Facebook, Messenger, and Meta's AI glasses. Meta also claims the original v1 matched Llama 4 Maverick's capability at "over an order of magnitude less compute" (a vendor claim).
What's new in Muse Spark 1.1?
The April launch was consumer-only; the real developer moment was Muse Spark 1.1 on July 9, 2026. It brought:
- A paid API (the Meta Model API), in public preview — Meta's first metered, pay-as-you-go API, with an OpenAI-compatible interface. Reported at launch as US-only, with $20 in free credits for new accounts.
- A 1M-token context window (up from ~262K in v1), with active context management that compacts earlier steps while keeping the critical ones.
- Video and PDF input in a single call, on top of the existing text, image, and speech input.
- An agentic and coding focus — Meta highlights enterprise-codebase support, a planning mode, and goal conditioning. CNBC framed 1.1 as Meta finally entering the AI coding market to chase Anthropic and OpenAI.
The API model id has been reported as muse-spark-1.1; confirm the exact string and base URL in Meta's live developer console before wiring it into production, as those specifics come from secondary coverage rather than Meta's primary docs.
What can Muse Spark do?
Muse Spark exposes three reasoning modes that trade speed for depth:
- Instant — fast, direct answers.
- Thinking — extended reasoning for harder problems.
- Contemplating — a multi-agent parallel-reasoning system that spawns several agents, reasons in parallel, and synthesizes the results (Meta's answer to the "spawn sub-agents" pattern also seen in OpenAI's Sol Ultra and Grok 4.5).
In hands-on testing, Simon Willison found roughly 16 built-in tools — web search, a Python code interpreter (with pandas, numpy, matplotlib, OpenCV), image generation, HTML/SVG rendering, and visual grounding (bounding-box, point, and count outputs) — and likened the ecosystem to Claude Artifacts, calling it "quite strong." Muse Spark is also multimodal on perception (visual STEM, entity recognition) and was tuned on health tasks with input from over a thousand physicians, which is why Meta leans on health and multimodal use cases in its demos.
How much does Muse Spark cost?
Two very different answers, depending on how you use it:
- Consumers: free. Muse Spark powers Meta AI at meta.ai and in the Meta AI app at no charge (Artificial Analysis lists the consumer tier at $0/$0).
- Developers (Muse Spark 1.1 API): $1.25 per million input tokens and $4.25 per million output tokens, pay-as-you-go, with $20 in free credits to start.
That pricing is the headline. Here's how it lands against the flagships it's chasing:
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Muse Spark 1.1 | $1.25 | $4.25 |
| xAI Grok 4.5 | $2.00 | $6.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GPT-5.5 / 5.6 (Sol) | $5.00 | $30.00 |
| Claude Fable 5 | $10.00 | $50.00 |
At roughly a quarter of Anthropic's and OpenAI's output rates — and cheaper than even Grok 4.5 — Muse Spark 1.1 is the cheapest near-frontier API in the set above. One important caveat for cost planning: Muse Spark is a reasoning model, so its hidden chain-of-thought tokens are billed at the $4.25 output rate. On tasks that reason a lot, the effective per-task cost is higher than the sticker price suggests — run your own token accounting before assuming the headline number.
How good is Muse Spark? (Benchmarks)
Independent numbers matter more than usual here, because Meta drew benchmark-gaming accusations during the Llama 4 / LMArena episode. So lead with the third parties:
- Artificial Analysis currently scores Muse Spark at an Intelligence Index of 43, ranked #35 of 572 models, with a 262K context window on the page. Note that AA re-versioned its index (v4.0 → v4.1) and recalibrated every model downward — Muse Spark's earlier "52" and today's "43" reflect that reversioning, not a regression.
- VALS-AI ranked Muse Spark 1.1 4th overall and called it "particularly fast and cost-effective."
On Meta's reported 1.1 benchmarks (relayed by independent press such as The Decoder), the picture is a clear split (all figures vendor-reported and not yet broadly reproduced): Muse Spark 1.1 leads on agentic and reasoning — MCP Atlas ~88.1 and Humanity's Last Exam ~62.1, both reported ahead of Claude Opus 4.8 and GPT-5.5 — but trails on pure coding, with SWE-Bench Pro ~61.5 versus Opus 4.8's ~69.2, and it's a known weak performer on Terminal-Bench, behind Anthropic's and OpenAI's coding models. Meta itself admits long-horizon agentic workflows are still a gap. Treat the exact vendor figures as claims pending broader independent replication.
How does Muse Spark compare with GPT-5.6, Claude Fable 5, Gemini 3.1 Pro, and Grok 4.5?
Muse Spark isn't trying to take the intelligence crown — it's running a value play. Here's the practical framing:
- vs GPT-5.6 and Claude Fable 5 — those lead on raw coding correctness and top the aggregate intelligence indexes; Muse Spark undercuts both dramatically on price and matches or beats them on some agentic tool-use and reasoning benchmarks. If your workload is coding-correctness-heavy, GPT-5.6 or Fable 5 win; if it's high-volume agentic or cost-sensitive, Muse Spark is compelling.
- vs Claude Opus 4.8 — Opus is stronger on SWE-Bench-grade code correctness; Muse Spark is roughly 6× cheaper on output and reportedly ahead on MCP Atlas and HLE.
- vs Gemini 3.1 Pro — Gemini offers a larger 2M context and broad knowledge; Muse Spark is cheaper and stronger on some agentic tests.
- vs Grok 4.5 — the closest price rival. Muse Spark 1.1 undercuts even Grok's $2/$6; both are "near-frontier at a fraction of the cost" plays.
The honest summary: Muse Spark is "good enough on most, best on cost, and genuinely strong on agentic tool use — but not the model you pick for the hardest single-shot coding." For a broader map of the open-vs-closed landscape, see our open-source LLMs landscape and the Llama 4 guide Muse Spark supersedes.
How to access Muse Spark
- Free, for consumers: open meta.ai in a browser or the Meta AI app (iOS/Android), and it's also rolling into WhatsApp, Instagram, Facebook, Messenger, and Meta's AI glasses. Access requires a Meta login, and Meta trains on public user data.
- Developers: the Muse Spark 1.1 Meta Model API is in public preview (reported US-only at launch), with an OpenAI-compatible interface and $20 in free credits — so existing OpenAI-style clients can point at it with minimal changes. Confirm the exact model id and endpoint in Meta's developer docs.
What are the risks and controversy around Muse Spark?
Muse Spark is as much a strategy story as a model story, and it comes with real friction:
- The open-source pivot. Abandoning open-weight Llama for a closed model is a reputational risk with the developer community that Llama's openness built. It's the single biggest debate around the launch.
- Benchmark skepticism. Given the prior Llama 4 / LMArena benchmark-gaming accusations, independent evaluations (Artificial Analysis, VALS-AI) carry more weight than Meta's own charts — weigh them accordingly.
- Privacy. Consumer use requires a Meta account and training on public user data, and a separate Meta image feature that applied AI edits to other users' Instagram photos without permission drew public backlash. Read the data terms before routing sensitive work through the free consumer tier.
Should you use Muse Spark?
Use it when cost is the deciding factor and your workload is high-volume, agentic, reasoning-heavy, or multimodal — Muse Spark 1.1's pricing is the most aggressive at the near-frontier, and its tool-use is genuinely strong. Look elsewhere when you need best-in-class single-shot coding correctness (GPT-5.6 or Claude Fable 5), the longest contexts (Gemini), or when data-governance concerns rule out a Meta-hosted, closed model with no self-host option.
As always, don't switch on benchmarks alone — run your own eval set on your real prompts and compare quality and total token cost (chain-of-thought included). Our AI coding agents guide covers how these models slot into day-to-day agentic workflows. For the earlier head-to-head, see our Muse Spark vs ChatGPT, Claude & Gemini comparison.
FAQ
Who makes Muse Spark?
Meta — specifically Meta Superintelligence Labs (MSL), led by Chief AI Officer Alexandr Wang. It's Meta's first proprietary, closed model, replacing the open-weight Llama line as Meta's flagship.
Is Muse Spark free?
For consumers, yes — it powers Meta AI at meta.ai and in the Meta AI app at no cost. For developers, the Muse Spark 1.1 API is paid: $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits to start.
How much does the Muse Spark API cost?
$1.25 per million input tokens and $4.25 per million output tokens — roughly a quarter of OpenAI's and Anthropic's rates, and cheaper than Grok 4.5. Because it's a reasoning model, hidden chain-of-thought tokens bill at the output rate, so the effective per-task cost runs higher than the headline.
Is Muse Spark open source?
No. Unlike Meta's earlier Llama models, Muse Spark has no open weights — it's a closed, hosted model accessed through Meta's apps and API. VentureBeat framed the launch as the end of Meta's open-weight era.
Is Muse Spark good at coding?
It's mixed. Muse Spark 1.1 is strong on agentic tool-use benchmarks (MCP Atlas, JobBench) and reasoning (Humanity's Last Exam), but trails Claude Opus 4.8 and GPT-5.6 on pure code-correctness tests like SWE-Bench Pro and Terminal-Bench — a gap Meta itself acknowledges on long-horizon agentic work.
Choosing between models like Muse Spark, GPT-5.6, and Claude for your stack, and need engineers who already build and evaluate this way? Codersera helps you hire vetted remote developers fluent in agentic AI tooling — start with a risk-free trial.