TL;DR — On July 9, 2026, Meta Superintelligence Labs shipped Muse Spark 1.1, its first model behind a paid API. Pricing sits at $1.25/M input, $4.25/M output with $20 free credits — undercutting Anthropic’s Opus 4.8 ($5/$25) and OpenAI’s GPT-5.5 (~$5/$30) by 6–7× on output. It tops JobBench (54.7) and MCP Atlas (88.1) for professional and scaled tool use, but trails on pure coding: SWE-Bench Pro (61.5 vs Opus’s 69.2) and DeepSWE (53.3 vs GPT-5.5’s 67.0). The catch: US-only public preview with a waitlist, no OpenRouter, and no release date for the open-source variant. Meta is planting a flag — but the access gates and mixed coding scores make this a “watch closely, prototype when you can” launch, not a “switch now” moment.
Introduction
Three months after unveiling Muse Spark under AI chief Alexandr Wang, Meta has fired its opening shot in the coding-agent war. Muse Spark 1.1, released July 9, 2026, is a strategic pivot: for the first time, Meta is charging developers for API access to its own frontier model.
The original April 2026 Muse Spark was a select-partners preview, openly weak on coding and agentic work. Version 1.1 targets exactly those gaps with what Wang calls a “step-change” in capability and a price tag designed to pressure incumbents (CNBC).
The honest read: Muse Spark 1.1 wins the agentic tool-use race but concedes raw coding accuracy to Opus 4.8 and GPT-5.5. And the public preview is US-only with a waitlist — developers in Europe and elsewhere can’t touch it yet.
What Shipped
Muse Spark 1.1 comes out of Meta Superintelligence Labs (MSL), the unit assembled under Alexandr Wang. Internally code-named Avocado, it’s the first model Meta has ever put behind a paywall (CNBC).
| Spec | Detail |
|---|---|
| Model type | Multimodal reasoning (Thinking mode) |
| Context window | 1M tokens, self-managed with active compaction |
| Modalities | Text, images, video, documents |
| Agent roles | Primary agent + subagent orchestration |
| Tooling | Native MCP, custom skills, computer-use mode |
| SDK compatibility | Drop-in OpenAI + Anthropic SDKs |
| Availability | US-only public preview with waitlist |
The 1M-token context window is notable for its active compaction: the model manages the window itself, deciding what to keep in active memory and what to compress as sessions grow long. For agentic runs stretching across hours, this matters — naive context-stuffing is where most long-session agents fall apart on cost and coherence (Lushbinary).
The dual SDK compatibility is a quiet distribution win. Point an existing OpenAI or Anthropic client at https://api.meta.ai/v1, set the model to muse-spark-1.1, and keep your code — near-zero-friction swap for A/B testing (Digital Applied).
Pricing: Aggressive, With a Reasoning-Token Catch
Meta priced Muse Spark 1.1 to undercut incumbents dramatically (CNBC):
| Input (per 1M tokens) | Output (per 1M tokens) | |
|---|---|---|
| Muse Spark 1.1 | $1.25 | $4.25 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GPT-5.5 | ~$5.00 | ~$30.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
New accounts get $20 in free credits — enough for evaluation, not production. Wang called the pricing “very aggressive and attractive” meant to “scale with immense consumption usage” (Reuters via CNBC).
The reasoning-token trap. Muse Spark 1.1 thinks before it answers. Thinking tokens appear as reasoning_tokens and are billed as output at $4.25/M. The reasoning_effort parameter runs from minimal to xhigh — matching effort to task complexity is the main lever on your bill. A heavy reasoning workload costs significantly more than the headline suggests (AIReiter).
At sticker price, Muse Spark output is roughly 6× cheaper than GPT-5.5. But third-party relays already deliver Sonnet-class Claude at ~$0.60 in / $3.00 out — below Muse Spark’s own price, with no waitlist.
Benchmarks: The Split-Decision Story
Meta’s own launch table tells a clear story: Muse Spark 1.1 leads on agent and tool-use benchmarks, trails on pure coding and multimodal (Digital Applied). All figures are vendor-reported, with rivals in their strongest modes (Opus 4.8 at max, GPT-5.5 at xhigh, Gemini 3.1 Pro at high). Leading scores in bold.
| Benchmark | Category | Muse Spark 1.1 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| MCP Atlas | Agent · scaled tool use | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench | Agent · professional tool use | 54.7 | 48.4 | 38.3 | 15.9 |
| Humanity’s Last Exam | Agent · reasoning with tools | 62.1 | 57.9 | 52.2 | 51.4 |
| OSWorld-Verified | Agent · computer use | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro | Coding | 61.5 | 69.2 | 58.6 | 54.2 |
| DeepSWE 1.1 | Coding · long-horizon | 53.3 | 59.0 | 67.0 | 12.0 |
| Terminal-Bench 2.1 | Coding · CLI tasks | ~55 | ~72 | ~65 | — |
| BabyVision | Multimodal | 76.3 | 81.2 | 83.6 | 51.5 |
Sources: Meta launch blog; compiled from (Digital Applied, Lushbinary)
Reading the Pattern
On the four agent rows — MCP Atlas, JobBench, Humanity’s Last Exam, and Finance Agent v2 (57.2, not shown) — Muse Spark 1.1 is first. The JobBench lead (54.7 vs Opus’s 48.4 vs GPT’s 38.3) measures professional tool use in realistic workflows. MCP Atlas 88.1 on scaled tool use across multiple MCP servers is the capability that matters when building agents that run real tools against real APIs.
On computer use (OSWorld), it slips behind Opus at 80.8 vs 83.4 — close but not leading. Across coding (SWE-Bench Pro, DeepSWE, Terminal-Bench), it’s a solid third. This is a model that wins the tool-orchestration race and concedes raw accuracy — for agent builders, that’s arguably the right trade-off.
Independent benchmarks for 1.1 are still sparse. Artificial Analysis scored the original April Muse Spark at Intelligence Index 52, behind Gemini 3.1 Pro and Claude Opus 4.6. The 1.1 update tracks as roughly a 43-point gain across Meta’s internal suite, but third-party SWE-Bench Verified numbers weren’t published at launch (AIReiter). The pattern is credible because Meta shows the rows it loses — treat exact numbers as a starting hypothesis, not a verdict.
Agentic Capabilities
The architectural centerpiece is native multi-agent orchestration. The same model functions as both primary agent and subagent — a lead instance breaks tasks into pieces, delegates to parallel subagents, then gathers results. This orchestrator-worker pattern is baked into the model, not bolted on by a framework (Lushbinary).
Four supporting building blocks:
- Computer use. Drives a real Linux desktop from plain-language goals — screenshots, reasoning, clicks, repeat. OSWorld 80.8 puts it within striking distance of Opus 4.8’s 83.4 (AIReiter).
- Native MCP support. Speaks the Model Context Protocol natively. Combined with MCP Atlas 88.1, this is what separates an agent model from a chat model.
- Custom skills and built-in web search. Reusable capabilities as first-class primitives; add
{"type": "web_search"}to any Responses API call for real-time cited answers.
Early partners reinforce the positioning: Replit’s Amjad Masad called it “a complete agentic foundation,” Cline’s Saoud Rizwan highlighted “tool use strong enough to run real coding workloads at a viable price point” (Digital Applied).
Access: The Gates That Matter
Muse Spark 1.1 is available through the Meta Model API in public preview:
- Sign up at Meta’s developer portal; $20 in free credits per account
- US-only for now — early partners have access; new users join a waitlist
- Not on OpenRouter — Meta is deliberately limiting API access to its own properties (CNBC)
For consumers, it’s live in Thinking mode in the Meta AI app. Wang expects it to eventually replace Llama models across WhatsApp, Instagram, Facebook, and smart glasses. The US-only waitlist is the biggest practical barrier — European and Asian teams needing frontier agentic models today must use Claude, GPT, or Gemini, which are globally available.
Market Analysis
Meta’s $1.25/$4.25 pricing is the most aggressive from a hyperscaler in the coding-agent space, undercutting Opus 4.8 output by 6× and GPT-5.5 by 7× on sticker price. This is deliberate price compression: force competitors to match or justify their premium. When Anthropic launched Sonnet 4.6 at $3/$15, it created a mid-tier; Muse Spark undercuts even that while claiming Opus-class agentic performance.
The benchmark split reveals a bifurcating market. Pure coding (SWE-Bench, DeepSWE) is where Opus 4.8 leads and raw accuracy is the only metric. Agentic coding (MCP Atlas, JobBench) is where tool orchestration and multi-step planning matter more. Muse Spark 1.1 targets the second market. Wang framed it explicitly: “You kind of have to build coding capabilities as part of that in service of overall agentic capabilities” (CNBC).
Meta has structural advantages. Muse Spark 1.1 runs on infrastructure already powering Facebook and Instagram — combined with Meta Compute (July 1 cloud launch), this gives cost leverage pure AI labs lack. Wang confirmed an open-source variant is “in development” with no release date — a potential Llama-moment for coding agents, but not a product yet. Meta is also training Watermelon, a larger model using “an order of magnitude more compute than Avocado,” with internal reports suggesting GPT-5.5 parity on key benchmarks.
Should You Build on Muse Spark 1.1?
| Use case | Recommendation | Why |
|---|---|---|
| Agentic / computer-use prototypes | Join the waitlist | Multi-agent recipes and computer-use are where it pushes the field; $20 credits make exploration free |
| Production coding | Stay on Claude or GPT | Independent coding evidence is thin; a US-only waitlist is a poor production dependency |
| Outside the US | Use Claude / GPT / Gemini | The models it benchmarks against are globally available today |
| Cost-sensitive agentic workloads | Watch the bill | Reasoning tokens billed as output raise effective cost; match reasoning_effort to task |
| MCP-heavy architectures | Strong candidate (when accessible) | 88.1 on MCP Atlas is best-in-class for scaled tool use |
Muse Spark 1.1 is not a “switch now” for most teams. It’s a “watch closely, prototype when you get in” — and a reminder that Meta is back in the frontier race with pricing that pressures the field.
FAQ
What is Muse Spark 1.1?
A multimodal reasoning model from Meta Superintelligence Labs, released July 9, 2026. It features a 1M-token self-managed context window, native primary-agent and subagent orchestration, MCP support, custom skills, and computer-use mode. It is the first Meta model available through a paid developer API.
How much does the API cost?
$1.25/M input, $4.25/M output, with $20 free credits for new accounts. Internal reasoning tokens are billed at the output rate. Use reasoning_effort (minimal to xhigh) to control cost.
How does it compare to Claude and GPT on benchmarks?
It leads on agentic tool use (JobBench 54.7, MCP Atlas 88.1) but trails on pure coding (SWE-Bench Pro 61.5 vs Opus’s 69.2, DeepSWE 53.3 vs GPT-5.5’s 67.0). It is a tool-orchestration specialist, not a coding-accuracy leader.
Can I use it outside the US?
Not through the official Meta Model API. The public preview is US-only with a waitlist. Use Claude, GPT, or Gemini for immediate access.
Is Muse Spark open source?
No — it’s proprietary and paid. An open-source variant is in development with no release date.
Is it available on OpenRouter?
No. Meta is deliberately limiting API access to its own properties.
What SDKs are compatible?
Drop-in compatible with both the OpenAI SDK (Chat Completions + Responses) and the Anthropic SDK (Messages format). Set base URL to https://api.meta.ai/v1, model to muse-spark-1.1.
What’s next for Meta’s model lineup?
A larger model code-named Watermelon is training, using an order of magnitude more compute than Avocado (Muse Spark). Internal reports suggest GPT-5.5 parity but no release date.
Further Reading
- Meta’s official Muse Spark 1.1 launch blog — vendor benchmarks, API docs, Cookbook patterns
- CNBC: Meta jumps into AI coding market — Alexandr Wang interview, pricing, Watermelon
- AIReiter: Muse Spark 1.1 API Pricing Guide — pricing breakdown, competitive comparison, access guide
- Digital Applied: Muse Spark 1.1 Agentic Model API — benchmark deep-dive, tool-use analysis
- Lushbinary: Muse Spark 1.1 Developer Guide — architecture, agentic design patterns
- AI Weekly: Meta prices Muse Spark 1.1 API — concise pricing analysis
- The Agent Report: Meta’s Watermelon Catches Up to GPT-5.5 — prior coverage of Meta’s model roadmap
- The Agent Report: Meta’s Muse Spark and the Llama Abandonment Strategy — Meta’s shift from open-source to proprietary
Benchmark data sourced from Meta’s vendor-reported launch table and independent reporting as of July 9, 2026. All figures should be treated as Meta’s claims until third-party evaluations are published. Pricing and availability are accurate as of launch date and may change.