Meta's Muse Spark 1.1 is a proprietary multimodal reasoning model designed for coding, tool use, computer operation, long-context analysis, and multi-agent workflows. This guide separates Meta's claims from the evidence, explains the API costs, and shows where the model is competitive and where important limitations remain.
Release details, pricing, regional availability, benchmark scores, and safety findings were checked against Meta's launch materials, Meta's 112-page evaluation report, official developer documentation, and Reuters. Benchmark scores are evidence from specific test setups, not universal proof that one model is best for every workload.
Muse Spark 1.1 is Meta Superintelligence Labs' proprietary multimodal reasoning model for agentic work. It can process supported text, image, video, audio, and document inputs; call tools; operate computer interfaces; coordinate subagents; and work within a published 1,048,576-token context window. It is available through supported Meta AI surfaces and the Meta Model API public preview. Launch pricing is $1.25 per million input tokens and $4.25 per million output tokens. [1] [3] [4]
Muse Spark 1.1 Overview and Quick Facts
Meta announced Muse Spark 1.1 on July 9, 2026 as a major update to the original Muse Spark model released in April. The company positions it as a multimodal reasoning model built for agentic tasks, with improvements in coding, tool use, computer use, multimodal understanding, and multi-agent orchestration. [1] [4]
The key distinction is operational. Muse Spark 1.1 is not designed only to generate a polished answer. It is designed to gather context, plan work, call software tools, interact with interfaces, delegate tasks, inspect results, and continue through a multi-step workflow.
Muse Spark 1.1 is Meta's reasoning and action model. Muse Image is a separate image-generation system. Muse Spark can use supported visual, video, audio, and document information inside a workflow, but it is not Meta's dedicated image-generation model.
Why Muse Spark 1.1 Matters
The release combines three strategic changes: stronger long-running agents, direct paid developer access, and a proprietary hosted model that sits apart from Meta's open-weight Llama strategy.
Answers Become Actions
The model can plan, call tools, operate interfaces, validate results, and revise its approach instead of stopping after a text response.
One Agent Becomes a Team
Muse Spark 1.1 can coordinate parallel subagents, potentially reducing end-to-end time on research, coding, and operational tasks.
Open Weights Become Paid Access
Unlike Llama releases, Muse Spark 1.1 is delivered as a managed, closed-weight API service controlled by Meta.
What Changed From Muse Spark 1.0?
| Area | Muse Spark 1.0 | Muse Spark 1.1 | Why it matters |
|---|---|---|---|
| Developer access | Limited partner access | Public preview through Meta Model API | More developers can evaluate and prototype with the model |
| Multi-agent work | Earlier agent framework | Trained to delegate across parallel subagents | Large jobs can be divided when roles and controls are clear |
| Computer use | More limited extended workflows | Stronger navigation, scripting, adaptation, and context retention | Better fit for workflows that cross multiple applications |
| Coding | Strong but limited on long-horizon work | Large gains on software-agent and coding evaluations | More credible for repository-level analysis and migrations |
| Agent security | More vulnerable to injection and jailbreak attacks | Improved model-level resistance | Reduces risk but does not replace application controls |
| DeepSWE 1.1 | 10.0 | 53.3 | Shows substantial improvement on long-horizon engineering tasks |
Meta reports major gains over Muse Spark 1.0, including JobBench increasing from 17.0 to 54.7, OSWorld-Verified from 53.3 to 80.8, and DeepSWE 1.1 from 10.0 to 53.3. [2]
Core Features and Capabilities
Multi-Agent Orchestration
Muse Spark 1.1 can operate as a lead agent that gathers context, creates a plan, delegates work to parallel subagents, and combines their outputs. It can also act as a narrower subagent and escalate decisions back to a lead agent. [1]
Parallelism does not guarantee quality. Multiple agents can repeat work, use conflicting assumptions, or increase token costs unless roles, evidence rules, conflict resolution, and stopping conditions are explicitly defined.
Tool and Function Calling
The Meta Model API supports developer-defined tools and function calling. Applications can connect the model to databases, custom APIs, search, code execution, enterprise applications, and MCP servers.
Tool reliability depends on application design as much as model intelligence. Weak schemas, ambiguous tool descriptions, broad permissions, and missing confirmation steps can make a capable model behave unreliably.
Computer Use
Muse Spark 1.1 can inspect interfaces, click controls, type information, and write scripts when automation is more efficient than manual interaction. Meta emphasizes workflows that cross several applications and change while the task is running. [1]
This capability is useful when a clean API does not exist, but it can also misread controls, overwrite data, expose files, or perform an irreversible action.
Agentic Coding
Meta positions coding as one of the model's strongest upgrades. It can diagnose bugs, implement features, navigate large repositories, run tests, inspect screenshots, trace visible failures to source code, and validate changes.
Multimodal Understanding
Muse Spark 1.1 can use supported text, image, video, audio, and document information where perception must lead to action. Meta demonstrates a workflow that analyzes a smartphone video, extracts useful product images, reasons about the item, and operates a browser to create a marketplace listing. [1]
One-Million-Token Context
Meta publishes a 1,048,576-token context window and says the model can retrieve earlier actions and compact older material while preserving important steps. [1] [3]
How an Agentic Muse Spark Workflow Operates
Identify the requested result, constraints, available evidence, and success conditions.
Divide the objective into tasks and decide which files, tools, or subagents are required.
Call APIs, search, write code, operate software, or assign work to specialized agents.
Check tool output, identify failures, compare results with acceptance criteria, and revise the plan.
Deliver the result or pause before a sensitive, expensive, external, or irreversible action.
The hidden weakness is compounding error. A system that succeeds independently on 95 percent of steps has only about a 60 percent probability of completing ten required steps without any failure. Real tasks are not fully independent, but the arithmetic explains why long workflows need checkpoints and validation.
Muse Spark 1.1 Exact Benchmark Results
Meta's evaluation report compares Muse Spark 1.1 with Muse Spark 1.0, Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5. These are the models used in Meta's published evaluation. They should not be confused with a complete list of the newest commercial alternatives available on July 16, 2026. [2]
| Benchmark | Muse Spark 1.1 | Muse Spark 1.0 | Gemini 3.1 Pro | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|---|
| Humanity's Last Exam with tools | 62.1 | 50.4 | 51.4 | 57.9 | 52.2 |
| Humanity's Last Exam without tools | 52.2 | 42.8 | 45.4 | 49.8 | 44.8 |
| MRCR long context, 1M | 54.1 | - | 26.3 | - | 74.0 |
| MCP Atlas | 88.1 | 82.2 | 78.2 | 82.2 | 75.3 |
| Toolathlon-Verified | 75.6 | 49.4 | 61.1 | 76.2 | 73.5 |
| OSWorld-Verified | 80.8 | 53.3 | 76.2 | 83.4 | 78.7 |
| OSWorld 2.0 binary / partial | 14.2 / 47.3 | - | 7.8 / 30.6 | 20.6 / 54.8 | 13.9 / 47.5 |
| WebArena-Verified | 69.0 | 59.0 | 69.0 | 71.2 | 67.0 |
| DeepSearchQA | 84.9 | 76.8 | 71.3 | 84.3 | 87.8 |
| GDPval-AA v2 Elo | 1381 | 1145 | 963 | 1600 | 1494 |
| JobBench | 54.7 | 17.0 | 15.9 | 48.4 | 38.3 |
| Finance Agent v2 | 57.2 | - | 43.0 | 53.9 | 51.8 |
| Terminal-Bench 2.1 | 80.0 | 67.3 | 70.3 | 82.7 | 83.4 |
| SWE-Bench Pro | 61.5 | 55.0 | 54.2 | 69.2 | 58.6 |
| DeepSWE 1.1 | 53.3 | 10.0 | 12.0 | 59.0 | 67.0 |
| HealthBench Professional | 59.3 | 54.1 | 41.6 | 55.8 | 51.8 |
| CharXiv Reasoning with tools | 88.4 | 88.9 | 81.6 | 89.9 | 84.8 |
| BabyVision with tools | 76.3 | 39.9 | 51.5 | 81.2 | 83.6 |
Humanity's Last Exam, MCP Atlas, JobBench, Finance Agent v2, and HealthBench Professional.
Toolathlon, OSWorld-Verified, WebArena-Verified, Terminal-Bench, and CharXiv.
Long-context MRCR, GDPval-AA, SWE-Bench Pro, DeepSWE, and BabyVision.
How to Interpret the Benchmark Claims
The scores are useful, but the table is not a neutral universal ranking. Meta used high reasoning settings for the compared models. Some results are company-reported, some come from external leaderboards, and others were reproduced using Meta's tools, prompts, and evaluation harnesses. [2]
Meta also acknowledges that competing proprietary models may not be fully optimized for its scaffolds. Different rows use different judges, tools, execution environments, and task limits. The numbers matter, but they are not perfectly controlled laboratory comparisons.
Muse Spark 1.1 trails Claude Opus 4.8 and/or GPT-5.5 on several difficult coding and computer-use evaluations. Meta also states that the model does not yet have enough sustained autonomy to automate AI research and development over a prolonged period. [2]
A Better Evaluation Method
- Select 20 to 50 real tasks from your actual workload.
- Remove confidential data or use synthetic substitutes.
- Give every model equivalent tools, limits, and acceptance criteria.
- Run each task more than once to measure consistency.
- Track completion rate, human correction time, latency, and total cost.
- Record unsafe actions, tool failures, and errors benchmarks may miss.
- Choose the lowest cost per accepted result, not the lowest token price.
Muse Spark 1.1 API Pricing and Cost Calculator
These launch rates cover model tokens. A production agent can also generate costs from reasoning tokens, retries, external search, databases, virtual machines, storage, monitoring, and human review. [3] [4]
| Workload | Input tokens | Output tokens | Estimated cost |
|---|---|---|---|
| Large policy analysis | 200,000 | 20,000 | $0.3350 |
| Repository review | 500,000 | 50,000 | $0.8375 |
| Near-full-context analysis | 1,000,000 | 100,000 | $1.6750 |
| Ten calls at 100K input and 10K output | 1,000,000 total | 100,000 total | $1.6750 |
| Twenty calls at 100K input and 10K output | 2,000,000 total | 200,000 total | $3.3500 |
This estimate excludes taxes, external tool fees, storage, retries, infrastructure, and future price changes.
Availability and Regional Access
Meta AI Consumer Access
Meta says Muse Spark 1.1 is available in Thinking mode in the Meta AI app and on Meta's AI website. Access may differ by account, country, device, product surface, and rollout stage. [1]
Meta Model API
The Meta Model API launched in public preview. Reuters reported that US developers could access the preview at launch. [4]
The initial announcement specifically described US developer access. Developers in the Philippines should verify regional eligibility, billing support, account terms, and current rollout status before using the service as a production dependency.
Consumer access and developer API access are separate. Seeing Muse Spark features in a Meta application does not prove that the same account or region can create and bill a Meta Model API project.
How to Use Muse Spark 1.1 Through the API
- Create an eligible Meta developer account.
- Activate or request Meta Model API access.
- Create an API key and store it in a secure environment variable.
- Install a compatible SDK or call the REST endpoint directly.
- Use the current model identifier shown in Meta's documentation.
- Add strict permissions, logs, budgets, timeouts, and approvals.
- Test using non-sensitive data before connecting production systems.
Basic Python Example
import os
from openai import OpenAI
api_key = os.getenv("MODEL_API_KEY")
if not api_key:
raise RuntimeError("MODEL_API_KEY is not set.")
client = OpenAI(
api_key=api_key,
base_url="https://api.meta.ai/v1"
)
response = client.responses.create(
model="muse-spark-1.1",
input=(
"Review the project requirements. "
"Create an implementation plan, identify major risks, "
"and return a structured checklist. "
"Do not send, publish, delete, or modify external data."
)
)
print(response.output_text)Meta's public-preview endpoints, model identifiers, SDK behavior, and access rules can change. Confirm the latest official documentation before production deployment.
Recommended Agent Prompt Structure
Define the exact result required.
Identify trusted files, records, and sources.
List only the tools the agent may use.
State what it cannot send, change, publish, or delete.
Define actions requiring human confirmation.
Specify format, depth, and acceptance criteria.
Safety Findings and Minimum Deployment Controls
Meta's launch post says Muse Spark 1.1 operates within safe margins. The detailed evaluation report adds an essential qualification: Meta distinguishes the model's capabilities before safeguards from the residual risk after safeguards are applied. [2]
Meta reports improved resistance to prompt injection and direct misuse compared with Muse Spark 1.0. It also recommends system-level safeguards, tool allowlists, and workspace isolation for API deployments. [2]
Minimum Controls for a Muse Spark Agent
Provide only the tools and records required for the current task.
Use a sandbox, test account, virtual machine, or restricted workspace.
Pause before sending, publishing, purchasing, deleting, or changing data.
Record prompts, tool calls, outputs, costs, approvals, and errors.
Treat websites, documents, email, and tool output as injection sources.
Set token, call, retry, time, and spending limits.
Check calculations, schemas, file changes, and code using fixed tests.
Maintain backups and a tested process for reversing agent actions.
Muse Spark 1.1 Competitor Comparison
Meta's benchmark report compares Muse Spark 1.1 with GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro. By the article's publication date, GPT-5.6 Sol and Claude Fable 5 were also current commercial alternatives. The benchmark table and current-market table therefore answer different questions.
Models Used in Meta's Benchmark Report
Strong low-cost agentic, tool-use, and computer-use proposition.
Benchmark comparator used in Meta's July 9 evaluation report.
Benchmark comparator with strong coding and computer-use results.
Benchmark comparator with multimodal and long-context capabilities.
Current API Alternatives Verified July 16, 2026
| Model | Context | Max output | Input per 1M | Output per 1M | Key condition |
|---|---|---|---|---|---|
| Muse Spark 1.1 | 1,048,576 | Not clearly stated in reviewed public sources | $1.25 | $4.25 | Public preview with regional eligibility limits |
| GPT-5.6 Sol | 1,050,000 | 128,000 | $5.00 | $30.00 | Current GPT-5.6 frontier API model |
| Claude Fable 5 | 1,000,000 | 128,000 | $10.00 | $50.00 | Anthropic's highest-capability widely released model |
| Claude Opus 4.8 | 1,000,000 | 128,000 | $5.00 | $25.00 | Still available for complex agentic and enterprise work |
| Gemini 3.1 Pro Preview | 1,048,576 | 65,536 | $2.00 at 200K or less; $4.00 above 200K | $12.00 at 200K or less; $18.00 above 200K | Preview model with length-tiered pricing |
Current specification sources: Meta [3], OpenAI GPT-5.5 [5], Anthropic Opus 4.8 [6], Google Gemini 3.1 Pro [7], OpenAI GPT-5.6 Sol [8], and Anthropic current models [9].
Muse Spark 1.1 is substantially cheaper per published standard token than the premium alternatives in this table. That can matter for agents that produce long plans, reports, code, and tool arguments. It does not prove lower total cost. A cheaper model can still be more expensive if it needs more retries or more human correction.
When Should You Choose Muse Spark 1.1?
- You need low published token prices for agentic workloads.
- Your workflow depends on tools, MCP servers, or computer use.
- You need multimodal input inside a longer operational process.
- You can tolerate public-preview maturity and regional restrictions.
- Your own evaluation shows a low cost per accepted result.
- You require local, offline, or on-premises deployment.
- You need downloadable weights or deep model customization.
- Your organization is outside the supported API region.
- You need a mature commitment for a mission-critical system.
- A competitor performs materially better on your actual tasks.
Recommended Four-Week Pilot
Define tasks, test data, safeguards, and acceptance criteria.
WEEK 2Run Muse Spark and at least one current competitor on the same tasks.
WEEK 3Measure quality, consistency, latency, total cost, and correction time.
WEEK 4Approve limited deployment only after controls and rollback are tested.
TecTack Verdict
Muse Spark 1.1 is a serious frontier-model release rather than a minor Meta AI update. Its strongest case is the combination of low token pricing, strong tool-use evidence, capable computer use, multimodal perception, long context, and major gains over Muse Spark 1.0.
The evidence does not support calling it universally dominant. It leads several reasoning, professional-agent, finance, health, and MCP evaluations, while competing systems lead several difficult coding, long-context, computer-use, search, and multimodal tests.
Frequently Asked Questions
What is Muse Spark 1.1?
Muse Spark 1.1 is Meta's proprietary multimodal reasoning model for agentic work, including coding, tool use, computer use, and long-running workflows.
Who developed Muse Spark 1.1?
It was developed by Meta Superintelligence Labs.
When was Muse Spark 1.1 released?
Meta announced Muse Spark 1.1 on July 9, 2026.
What is its context window?
Meta publishes a 1,048,576-token context window. Large capacity does not guarantee perfect retrieval or memory.
How much does the Muse Spark 1.1 API cost?
Launch pricing is $1.25 per million input tokens and $4.25 per million output tokens. Confirm current pricing before deployment.
Is Muse Spark 1.1 open source?
No. It is a proprietary closed-weight model hosted by Meta.
Can Muse Spark 1.1 run locally?
Meta has not released downloadable Muse Spark 1.1 weights for ordinary local deployment.
Is the Muse Spark API available in the Philippines?
The launch public preview was reported for eligible US developers. Philippine developers should verify current regional eligibility directly with Meta.
Is Muse Spark 1.1 good for coding?
Yes. It shows major gains over Muse Spark 1.0 and strong coding results, but competing models lead several demanding software-engineering tests.
Can Muse Spark 1.1 operate a computer?
Yes, in supported agent frameworks. Computer use should be isolated and supervised because interface errors can cause real changes.
Can it understand video and audio?
Meta describes multimodal workflows that use visual and audio information, including product-video analysis followed by browser actions.
Is Muse Spark 1.1 better than GPT-5.6 Sol?
There is no universal answer. Muse Spark has much lower published token prices, while GPT-5.6 Sol is a newer current OpenAI frontier model than the GPT-5.5 version used in Meta's benchmark report. Test both on the same real workload.
Is Muse Spark 1.1 better than Claude?
Muse leads some published agent and reasoning evaluations. Claude models lead other coding and computer-use evaluations. The correct choice depends on reliability, quality, cost, controls, and ecosystem requirements.
What is the biggest operational risk?
The main risk is giving an imperfect agent broad access to tools, files, accounts, or systems. Prompt injection and multi-step errors make unrestricted autonomy unsafe.
Should a company deploy Muse Spark 1.1 now?
A controlled pilot is reasonable. Mission-critical deployment should require task-specific evidence, security controls, monitoring, fallback options, approvals, and rollback procedures.
Sources and Verification Notes
-
Meta - Introducing Muse Spark 1.1, July 9, 2026
Release date, product positioning, Thinking mode, context, orchestration, computer use, coding, multimodal workflows, and API announcement.
-
Meta - Muse Spark 1.1 Evaluation Report, July 9, 2026
Exact benchmark values, methodology, safety findings, prompt-injection results, autonomy limitations, and risk qualifications.
-
Meta Model API - Getting Started
API endpoint, model identifier, compatibility, context documentation, and developer guidance.
-
Reuters - Meta debuts Muse Spark 1.1 with preview open to developers
Independent confirmation of the launch, initial US preview, $20 credit, and launch token pricing.
-
OpenAI - GPT-5.5 API Model Documentation
Specifications for the OpenAI model used in Meta's benchmark comparison.
-
Anthropic - Claude Opus 4.8
Context, output, pricing, and positioning for the Anthropic model used in Meta's benchmark comparison.
-
Google AI for Developers - Gemini 3.1 Pro Preview
Context window, output limit, modalities, preview status, and model details.
-
OpenAI - GPT-5.6 Sol API Model Documentation
Current OpenAI frontier-model context, output limit, and standard pricing verified July 16, 2026.
-
Anthropic - Current Claude Models Overview
Current model lineup, Claude Fable 5 and Opus 4.8 pricing, context, maximum output, and availability.
July 16, 2026: Initial guide published. Added a distinction between Meta's benchmark comparators and current commercial alternatives, including GPT-5.6 Sol and Claude Fable 5.