Muse Spark 1.1

The flagship multi-modal reasoning model launched by Meta Super Intelligence Laboratory, specially built for Agent tasks and programming, with millions of token contexts, and the price is only a quarter of competing products at the same level.

large language model LLM Meta

https://ai.meta.com/

Visit Website

In-depth Report

Muse Spark 1.1 is a new flagship multi-modal reasoning model released by Meta Super Intelligence Laboratory on July 9, 2026. It is specially built for Agent tasks and focuses on tool invocation, computer operation, programming and multi-modal understanding, with a context window of 1 million tokens. It also brings Meta’s first paid developer interface, the Meta Model API, which is priced at US$1.25 per million input tokens and US$4.25 per million outputs, which is about a quarter of competing products at the same level. It is widely interpreted as a price war against OpenAI and Anthropic. Zuckerberg returned to the X platform in person after more than three years, marking Meta's major shift from the open source Llama route to the closed source charging model. The model won many firsts in tool calls and professional scenario tests, but it was also accompanied by controversy over Terminal-Bench rankings and complaints about client stability.
Muse Spark 1.1 comes from Meta Superintelligence Labs and is Meta’s first landmark achievement after reorganizing its AI team and investing heavily in poaching. On the day of the release, Zuckerberg posted his first post on X in more than three years, calling it "a powerful agent and programming model at an extremely low price." Musk responded immediately, and the interaction between the two became news in itself. Analysts believe that choosing to make the official announcement on a competitor’s platform is a deliberate communication strategy. The strategic significance of this release is greater than mere model iteration. Meta was known for its open-weighted Llama series in the past, but Muse Spark 1.1 is a closed-source model and is commercially available for external use for the first time through a paid API, which means that Meta has officially joined the enterprise-level model charging market where OpenAI and Anthropic are located. According to media reports such as Sina Finance, Meta also launched a self-developed AI chip "Iris" plan during the same period, cooperating with Broadcom design and TSMC manufacturing in an attempt to reduce dependence on Nvidia and AMD and support the goal of doubling computing power next year. The model will gradually replace Llama-powered chatbots in WhatsApp, Instagram, Facebook and smart glasses, covering Meta’s product matrix of billions of users.
Muse Spark 1.1 is a multi-modal reasoning system that supports text, image, video, audio and file input, with a context window of 1,048,576 tokens, four times that of the previous generation. It can plan tasks around user goals, call external tools, write and debug code, and maintain context throughout multi-step long-flow tasks. Meta said the model "knows when to write scripts to automate and when to click directly on the interface." It can shuttle between multiple applications to perform tasks, and requires minimal manual intervention when facing unfamiliar interfaces. For multi-agent automated workflows, the model introduces a context compression mechanism: the main Agent is responsible for generating the project plan, and the sub-Agents execute separately. The massive data generated during the execution process will be compressed to retain the most critical details. When necessary, previous work results can be traced back, and information can be transferred between different sub-tasks. This design directly addresses the pain point of "context explosion" in Agent workflow. In Meta's internal demonstration, engineers asked it to generate a chat application based on a prompt. The model could automatically screenshot the program interface, identify technical problems, and locate erroneous code snippets for repair. In terms of multi-modality, it can extract product information based on videos taken by users, complete product releases on Facebook's second-hand market on behalf of users, and can also complete practical tasks such as ordering food while generating code. Feedback from community testing is generally positive. Open source developer Simon Willison did the classic "Pelican Riding a Bicycle" SVG test after getting preview access for a few days, and was deeply impressed by the clip in the model evaluation report where two Muse Sparks talked to each other and talked about their existence. The founder of Julius AI lets it write and run the entire Minecraft game project directly in the code workbench. There are also users testing image recognition to let the model determine whether the food in the picture is edible. Muse Spark 1.1 passed the test, while the control Claude Fable 5 failed.
Pricing is the sharpest weapon in this release. The public preview version of the Meta Model API is open to US developers. It charges US$1.25 per million input tokens and US$4.25 per million output tokens, which is about a quarter of the same level of products as GPT-5.5 and Claude Opus 4.8. New users also receive a free credit of US$20. According to Sina Finance, its input and output costs are more than 50% lower than Anthropic and OpenAI products. The confidence comes from the profit support of the advertising business-Meta's operating income will grow by 20% in 2025 and accelerate to 30% in 2026. Community discussions generally believe that this is first and foremost a price war rather than a pure capability breakthrough. Developers on Hacker News have repeatedly emphasized the importance of cached input price, because multiple rounds of programming and Agent workflows will reuse a large number of contexts, and the cached price directly determines the actual cost. For teams running hundreds of thousands of Agent calls per day, the $4.25 per output price makes large-scale Agent batch processing economically feasible. On the business path, Meta uses low-priced APIs to attract developers, and puts models into its own consumer products (Meta AI applications have added a "Thinking" mode), forming a dual-line monetization of B-side and C-side.
Positive reviews focus on tool calls and long context. Many developers have reported that its tool calling success rate is higher than the first round of GPT-5.5 Agent links. The most promising scenarios are debugging, diagnosis and incident response: first grep logs, run performance analysis, organize reports, and then hand the results to stronger models for repair. The million-token window has withstood stress testing in real-world testing, with stable performance across large code bases and multi-file searches without the need for tiles or summarization. There was also a lot of negative feedback. Some users complained that the Meta AI iPad client equipped with the new model has extremely poor stability, with frequent crashes and garbled text output. Others complained that Meta AI could not even create an ordinary Excel table. Some developers bluntly stated that its overall quality is not as good as the Claude Sonnet series. "Strong tool usage does not mean that the overall quality is better." Some people also think that for a new model that has not yet accumulated a reputation, this price is not cheap, unless the quality can really be stable and close to the top level.
Third-party benchmark tests outline its capabilities: it is stronger in Agent, not weaker in programming, but it is not the best in all aspects. In the internal tests announced by Meta, it ranked first with a score of 88.1 in the MCP Atlas tool call test, higher than Claude Opus 4.8's 82.2 and GPT-5.5's 75.3; it also ranked first in the JobBench professional scenario tool usage test with a score of 54.7. In terms of computer operation, OSWorld-Verified scored 80.8, second only to Opus 4.8's 83.4; Terminal-Bench 2.1 scored 80 points, and SWE-Bench Pro scored 61.5 points. The Vals AI list shows that it ranks 4th in the comprehensive index (accuracy rate 68.41%), 6th in the multi-modal index; 2nd in the financial agent and corporate financial analysis tests, 1st in the medical record processing, and all three vertical industry tests of law, taxation, and medical documents. It ranks first, and the legal agent task has a clear gap. The Artificial Analysis Intelligence Index has increased by 8 points in three months, joining four cutting-edge models with scores above 50 (the others being GPT-5.5, Opus 4.8, and Kimi K3). The focus of media interpretation is on the strategic level. Zhidongxi, TechCrunch and other media regard it as a watershed for Meta from "open source for ecology" to "closed source charging", and also a new variable in the AI programming tool market. Industry analysts generally believe that Meta’s aggressive token pricing in exchange for users and exposure will put real pressure on Anthropic and OpenAI’s enterprise customer quotations in the short term.
The biggest technical controversy is the Terminal-Bench ranking scandal. Some community members discovered that Meta adjusted the resource configuration when running the benchmark, which triggered a debate about whether the score can still represent the same benchmark after changing the resource configuration. Some developers discounted the official score based on this. Secondly, there are questions about reliability: Meta plans to replace 90% of content reviewers with AI by the end of 2026, and last month it was reported that Meta’s AI customer service Agent had mistakenly given access to about 20,000 Instagram accounts to hackers. As the company pushes for agents that can act on behalf of users, this history has made outsiders wary of its security control capabilities. Route diversions themselves come at a cost. Abandoning open weights leaves communities that rely on Llama for self-deployment feeling abandoned, and teams that need private deployment or open source weights can only fall back on the Llama series or switch to open source alternatives such as Kimi K3. In addition, the investment circle still has doubts about Meta's "multi-pronged attack and lack of focus" - self-developed chips, super-intelligent laboratories, and the Metaverse are all advancing simultaneously, which will push up capital expenditures in the short term, and the lessons learned from the Metaverse are still there.
This model is most suitable for development teams building Agent workflows: MCP multi-tool orchestration, cross-application automation, long-context code base Q&A, large-volume Agent invocation and other scenarios. Its success rate and unit price combination have almost no rivals in the current market. Startups and independent developers with sensitive budgets and high call volumes are also worth a try. The $20 free credit is enough to complete a complete evaluation. Conversely, teams pursuing extreme code capabilities may still stay on Claude Opus 4.8; enterprises that require open source weight, local deployment, or data not leaving the domain are not suitable. It is a pure API closed source product with no privatization path. Ordinary users can experience it for free in the "Thinking" mode of the Meta AI application, but the current client stability problem has not been completely resolved, and it is not recommended to rely on it for important work.
Muse Spark 1.1 is Meta's frontal attack on the enterprise AI market with "quarter price + Agent's long board". Whether it can maintain its reputation depends on the clarification of the controversy over the fake list and its reliability performance in the actual production environment. In the short term, it has at least lowered the price anchor of cutting-edge models by one notch; in the long term, the combination of Meta's self-developed chips and advertising cash flow makes it possible for this price war to last longer than opponents expect.

User Reviews

Janet_Green007

—

Meta plans to replace 90% of its content moderators with AI before the end of the year. Last month it was revealed that AI customer service mistakenly gave permissions to 20,000 Instagram accounts to hackers. Now it is introducing an agent that can operate on behalf of users. I have doubts about its security control.

Benjamin_RamirezSr

—

Under the stress test, the retrieval performance of the 1M context in the multi-file code base was quite stable. Active context compression supported the quality of long sessions, and the agent loop command did not crash even after running for an hour. It can also accept audio and video inputs. Neither GPT-5.5 nor Opus 4.8 at the same price range can fully keep up with this input range. Multi-modality is a quiet differentiator, and overall it is much better than I expected.

kc7fg

—

If you need open source weight or local deployment, you can leave. This is a pure API closed source product with no privatization path. If you want to self-host, you can only use Llama or Kimi K3. The Super Intelligence Laboratory and the Llama open source team that make this model operate separately. One specializes in paid cutting-edge performance, and the other continues to be open source. Meta walks on two legs, but it means that there is no choice for enterprises whose data cannot go out of the domain.

TokenTiger152_eth

—

To sum up, a production agent that handles 500 customer service conversations every day, with an average of 40,000 input and 3,000 output tokens per call, costs just over 31 US dollars a day to run Muse Spark 1.1, which is more than 900 US dollars a month; the same job costs 145 US dollars a day on GPT-5.6 Sol. For tool-intensive businesses, this price difference is not about saving money, but directly changing the business model.

MMorales_9990

—

Note that the official comparison is between GPT-5.5 and Opus 4.8. The new generation was released in the same week as the release. It is common practice to use the previous generation as the control group, but you should be aware of the scores when looking at them.

MetcPro

—

It was only three months between 1.0 and 1.1, and the internal test pass@1 jumped from 48.1% to 67%. This iteration speed is a bit scary.

ERuiz_7744

—

The 1M window is real, but the MRCR long context retrieval score is only 54.1. The long window does not mean the search is accurate. After millions of tokens are filled, the key information recall is not at the cutting-edge level. Important scenes still have to be retrieved by yourself.

AnthonyRamirez

—

The API is temporarily unavailable outside the United States, and the European Union does not even have a timetable. Overseas teams should not bet their architecture on it yet. New users are given a $20 credit, which can generate approximately 4.7 million output tokens, which is enough for a complete evaluation.

RBaileyJr

—

The hallucination rate is a flaw. Artificial Analysis measured 38%, which is almost the highest among this generation of models. I myself have come across it as a straight up fact sheet. It leads the agent list, but the pure programming SWE-Bench Pro 61.5 loses to Opus 4.8's 69.2. Moreover, it talks a lot, and it consumes more tokens than average to get the answer. The actual cost and response time are stretched, and important answers must be checked by yourself.

MarilynMyersX

—

The tool calling is really powerful, and I almost never miss a beat when running MCP!

PinjaWalli

—

It's fast and cheap, and it's easy to use, but it's not a king, fast but not the fastest, smart but not the smartest.

Amy_Hill369

—

The demonstration of screenshot-driven debugging is quite amazing. It builds its own chat application, automatically takes screenshots to check the interface, and traces bugs back to specific lines of code. After fixing them, screenshots are taken to verify. This closed loop of checking, modifying, and verifying is very practical for front-end development. There is also a sandbox demonstration, which only gives it two tools, run and write_file, plus a one-time Docker container. It selects all 48 shell commands to fix a SWE-bench bug. It also searches the git history to find the commit that introduced the problem. You only need to provide the container and the target, not the script.

DianeHern_andez

—

From Llama's open weight to closed source paid API, Meta's turn has been thorough enough, and the community that relied on Llama for self-deployment has been abandoned this time. Moreover, the new model will gradually replace the Llama-driven chatbots in WhatsApp, Instagram, Facebook and smart glasses, which is equivalent to a complete overhaul of its product matrix with billions of users. The open source route seems to have really changed.

crazypeacock280

—

Meta is not an ordinary model update this time. It is directly aimed at the AI programming tool market. It can fix code vulnerabilities, do large-scale code migration, and adjust tools across multiple applications. Those who make programming tools should be nervous.

AdrijanaKapetanović

—

The API is fully compatible with the OpenAI format. Just change the base_url to api.meta.ai/v1 and the model name can be changed. The existing code basically does not need to be touched. The migration cost is almost zero. Teams who want to do A/B testing can directly jump into it.

xMiljaPeura_dev

—

The Terminal-Bench score is controversial. The resource configuration was changed during the benchmark run. Can the score after changing the configuration still represent the same benchmark? I have reservations.

JaniceBrown

—

The most promising uses are debugging and incident response: first let it grep logs, run profiler, organize reports, and then pass the conclusions to stronger models for repair. Tool calls can be made hundreds or thousands of times, and even if the single failure rate is only a few percent, the cumulative token cost and timeout are considerable. Fortunately, common tools such as bash and HTTP can basically survive by relying on verification and retrying.

Web3Kine

—

In the image recognition test, it judged whether the food in the picture was edible. It answered correctly. Claude Fable 5 failed the same question. Multimodality is really something.

Megan_Brooks_X1

—

In actual use, the overall quality is not as good as Sonnet. Strong tool usage does not mean that the overall quality is better. Don’t let the list lead you to the rhythm.

冬雪_4

—

The key is the price of cached input. Multiple rounds of coding and agent workflows will reuse a large amount of context. Low cache prices are the real savings. Just looking at the price is meaningless.

LawrenceMorales_2021

—

Models and APIs are promoted in the same release framework, and the price war signal is too clear. Meta uses advertising cash flow to subsidize the model. The input and output costs are more than 50% lower than Anthropic and OpenAI. The operating income of the advertising business will still increase by 20% in 2025, and there is sufficient ammunition. At the same time, it was also officially announced that the self-developed chip Iris, Broadcom designed and manufactured by TSMC, made it clear that it wanted to reduce its dependence on Nvidia. This battle between OpenAI and Anthropic is really difficult.

heavydog556

—

After reading the pricing, I recalculated the bills for several agent projects in my hand. The input was 1.25 and the output was 4.25 US dollars per million tokens, which is only about a quarter of competing products at the same level. The cost structure of super individuals is really going to change.

SGarcia_88944

—

Someone was able to write and run an entire "Minecraft" game project directly in Julius's code workbench. I was dumbfounded.

史琪

—

Xiao Zha’s first post when he returned to X after three years was to officially announce this model. He was still on Musk’s territory, and he was very determined to spread the word.

EGutierrez_88

—

The stability of the iPad version of Meta AI is too poor. It crashes frequently and the output text is garbled and broken. No matter how strong the model is, the client is still in vain.

Diana.Sanchez_770

—

Simon Willison got a few days of preview access to do the classic Pelican Cycling SVG test. He said that the section in the evaluation report where the two Muse Sparks talked about their existence was particularly interesting. After reading it, I also read the original text. It was really fun.

Christian.Perry9

—

Let it be turned into an ordinary Excel spreadsheet and it will turn over, so what's the agent that we agreed on?

Muse Spark 1.1

In-depth Report

User Reviews

Similar Products