Meta's Muse Models Are Quietly Frontier-Level

Muse Spark 1.1: An Agent, Not Just a Chatbot

Meta Muse Spark 1.1 is built for multimodal agentic work rather than pure reasoning contests. On benchmarks that measure real-world job tasks and MCP (Model Context Protocol) usage, it reportedly beats Opus 4.8 and GPT-5.5, and even tops an updated MCP Atlas chart above Claude Opus 4.8 and GPT-5.6. It performs worse on deep search QA, showing its strengths are task execution, not open-ended research.

The standout capability is computer use: navigating unfamiliar interfaces, maintaining context across long sessions, and deciding when to script an action versus click through a UI manually.

• Use it for workflow automation — tasks like contacting multiple vendors, comparing responses, and confirming orders across a web UI.
• Expect weaker performance on deep search-style QA — it's not built for that.
• Check the OSWorld 2.0 cost curve — Muse Spark 1.1 delivers comparable computer-use performance to Opus 4.8 at meaningfully lower cost.

Coding: Good Enough, Not the Best

Meta is explicit that Muse Spark 1.1 isn't meant to write the best possible code, but it's a large step up from the prior Muse Spark model — roughly three to four times better on the VIVE coding benchmark and nearly double on SWE Atlas. The key differentiator is a visual debugging loop: the model writes code, launches it in a browser, takes screenshots, inspects the output, and iterates until it works.

Meta's internal benchmarks (which the reviewer flags as self-reported and worth some skepticism) place it near Opus 4.8's level and above GPT-5.5 high. Meta is also using Muse Spark 1.1 internally to evaluate its own coding performance on Deep SWE tasks.

• Rely on the screenshot-and-fix loop for front-end or visual app generation rather than complex backend logic.
• Treat Meta's internal benchmark claims with some caution since the company has an incentive to flatter its own model.

Multimodal Strength and Pricing

Muse Spark 1.1 is described as unusually strong in perception, tool use, and grounded multimodal outputs — captioning, visual-to-code conversion, and executing agentic workflows that combine perception with action. On the BabyBench-style cost-efficiency comparison, it's reportedly the only model to score above 70 while costing under 20 cents, well below Opus 4.8 and GPT-5.5.

Pricing sits at $1.25 per million input tokens and $4.25 per million output tokens. On the Valves index (a benchmark weighting finance and coding tasks for economic impact), it ranked fourth, and it reportedly led on the Harvey legal agent benchmark, TaxEval, and MedScribe at the time of testing.

• Consider it for cost-sensitive multimodal pipelines — the price-to-performance ratio is the headline draw.
• A Facebook Marketplace demo — extracting product photos from smartphone video and auto-listing them — shows practical use for less tech-savvy users.

Muse Image and Muse Video vs. the Competition

Muse Image works as an agent rather than a one-shot text-to-image tool: it invokes tools, self-refines its own prompts, and can use test-time compute and web search to ground images in real information. In text-to-image arena rankings, the reviewer says it edges out Nano Banana 2, though Nano Banana 2 remains easier to access and cheaper for simple jobs. Against image editing benchmarks, Muse Image sits just behind the top model in both single- and multi-image edit categories. Generated images carry a watermark.

Muse Video sits below Seedance 2.0 but above Veo 3.1 in the reviewer's comparisons, and in his own side-by-side tests he found it edged out Google's Omni Flash in several cases — though he cautions results varied and examples may have been cherry-picked.

• Pick Muse Image for complex, multi-element compositions needing web-grounded accuracy; pick Nano Banana 2 for quick, simple edits.
• Test Muse Video against Veo and Seedance yourself — the reviewer found no dominant winner and stresses personal workflow testing over benchmark rankings.

The Payoff

For anyone building agentic workflows or multimodal pipelines, Meta's Muse lineup offers Opus-and-GPT-5.5-level task performance at a fraction of the cost — meaning the same automation work for significantly less spend, without waiting on the bigger-name labs.