The announcement of Grok 4.5 is less than a day old, but another AI model has already been released that operates at the same level as Claude Opus 4.8 and OpenAI’s GPT-5.5. Muse Spark 1.1 is Meta’s latest LLM, designed primarily for coding.
Meta’s AI vision revolves around “personal superintelligence,” meaning AI that helps you “pursue your goals, create what you imagine, deepen your relationships, and take action on what you value most.” That sounds quite ambitious, but it’s hard to quantify with hard numbers. The benchmarks, however, even though they come from Meta Superintelligence Labs itself, are quantifiable.
It’s immediately apparent that Muse Spark 1.1 fits into the now-familiar lineup of LLMs that fall just short of Claude Fable 5 but are much more affordable.
Reaching the (old) ceiling
Anyone who remembers 2023 well may recall that OpenAI’s GPT-4 dominated the scene. ChatGPT’s first upgrade was released using this model in March of that year. It took nearly a year for the competition to surpass that standard, despite dozens of prominent AI models coming and going. They all clustered just below GPT-4, even in their own benchmark results.
The dynamics are somewhat different in 2026, but a striking clustering does bring to mind that bygone era. While Fable 5 is ahead of the rest, the models Muse Spark 1.1, Grok 4.5, Opus 4.8, GPT-5.5, and GLM-5.2 can all be found in the same cluster. Take a look at Humanity’s Last Exam, Terminal-Bench 2.1, DeepSWE 1.1, and so on: they all deliver results that show, at most, a single, ever-changing outlier. It seems that no matter what the model developers are doing behind the scenes, they’re all using similar techniques to achieve virtually the same level of performance. That might be an illusion, but it’s still a remarkable coincidence.
For Muse Spark 1.1, those outliers are MCP Atlas and JobBench for tool usage, Humanity’s Last Exam as a reasoning test, and, somewhat out of the blue, financial analysis via Finance Agent v2. The original Muse Spark, which went largely unnoticed online due to its rather unremarkable results, is now far surpassed. After being absent from the top of AI benchmarks for over a year, Meta is clearly back.
Llama’s demise
Developers can gain access via the new Meta Model API. This API is still in public preview, however. Muse Spark 1.1 is not an open-weight model. This will be disappointing for users who were accustomed to Meta’s Llama series being freely available. After the first Llama models performed remarkably well as open-weight options compared to earlier LLMs, the momentum slowed after Llama 4. While the performance on paper was excellent, adoption was disappointing because real-world performance was nowhere near as competitive.
Earlier this year, Muse Spark replaced the Llama-based models within the Meta AI app and on meta.ai. The new model is now available in “Thinking” mode. It’s the missing piece of the puzzle in Meta’s strategy. CEO Mark Zuckerberg has already stated that his company will offer its own models and computing power as a cloud service.
Agentic tasks and computer use
Muse Spark 1.1 is thus built for agentic work, as clearly evident from its scores for tool usage. As the main agent, the model can create a plan and delegate tasks to parallel subagents to reduce latency. It also manages a context window of 1 million tokens, which is now the standard among LLMs. This is more than enough for most codebases, where AI models are increasingly being prioritized.
Meta also aims to make an impression when it comes to computer use. The model determines for itself when to write scripts and when to click directly. Computer use is inherently inefficient, so it would be beneficial if Meta could smooth this out. The alternative is, and remains, relying on integrations via, for example, MCP.
Competition remains fierce
The timing makes this week an LLM week par excellence. Following the return of Fable 5, three LLMs, all of which score quite strongly, have now been released in the space of half a week (Grok, Muse, GPT-5.6). Grok 4.5 is the first model from the rebranded SpaceXAI, which impresses particularly as a coding agent and was built through a collaboration with Cursor, which is now part of SpaceX following an acquisition by Elon Musk.