Meta Muse Spark 1.1: Cheap Long-Context AI Agents

On July 9, 2026, Meta Superintelligence Labs shipped Muse Spark 1.1 alongside the Meta Model API — the company's first paid, self-serve developer API. Mark Zuckerberg broke a three-year silence on X to announce it, which tells you how Meta is positioning this: not an experiment, a product.

The headline numbers are aggressive. A 1M-token context window. Priced at $1.25 per million input tokens and $4.25 per million output — roughly a quarter of what Claude Opus 4.8 ($5/$25) or GPT-5.5 (about $5/$30) charge. And, notably, no open weights. After years of the Llama open-release playbook, Meta is keeping this one behind an API.

That combination makes Muse Spark interesting. But the benchmark picture is more nuanced than the launch coverage suggests, and getting the routing decision right matters more than the price tag.

What Muse Spark 1.1 actually is

Meta describes it as "a multimodal reasoning model built for agentic tasks." That framing is precise, and worth taking literally. This is not a general-purpose coding model that happens to do agents. It is an orchestration model — built to run long, multi-step workflows across tools.

Four capabilities define it:

Self-managed context. The model actively manages its 1M-token window rather than passively filling it. It maintains an action history, retrieves details from earlier steps when needed, and compacts less critical material as the session grows. For agent loops that run for hours, this is the difference between a coherent run and a session that degrades into confusion.

Parallel subagent delegation. As the primary agent, Muse Spark gathers context, forms a plan, and delegates execution across parallel subagents. The subagents understand which tools are available and know when to escalate back up. This is native orchestration, not a framework bolted on top.

Zero-shot tool generalization. It generalizes to new native tools, MCP servers, and custom skills without having been trained on them. If you have already invested in MCP servers, they work out of the box.

Computer use across surfaces. It handles workflows spanning desktop, browser, and mobile. Meta highlights a specific behavior: rather than clicking through an interface step by step, the model decides when to write an automation script instead. That judgment call — script versus direct interaction — is where most computer-use agents waste enormous numbers of tokens.

The benchmarks tell two different stories

Here is where the marketing and the data diverge. Muse Spark is genuinely at the top for tool use and professional workflows, and genuinely mid-pack for raw coding.

Read that table carefully. Muse Spark leads where the task is orchestrating tools over a long horizon, and trails by 8 to 14 points where the task is writing and fixing code in a repository. Independent trackers place it in the same cluster as Grok 4.5, Opus 4.8, GPT-5.5 and GLM-5.2, with Claude Fable 5 ahead of that group overall.

The practical translation: do not swap out your coding agent for this. Do consider it for the workflow layer that sits above your coding agent — the part that reads tickets, queries systems, coordinates steps, and calls other tools.

Migration is a two-line change

Meta made the adoption path deliberately frictionless. The Meta Model API speaks both the OpenAI SDK format (Chat Completions and Responses) and the Anthropic Messages format. Pointing an existing agent at Muse Spark is a base-URL and key swap, not a rewrite.

import OpenAI from "openai";
 
const client = new OpenAI({
  baseURL: "https://api.meta.ai/v1",
  apiKey: process.env.META_API_KEY,
});
 
const response = await client.chat.completions.create({
  model: "muse-spark-1.1",
  messages: [
    {
      role: "user",
      content: "Audit our invoice pipeline for TTN compliance gaps and summarize findings.",
    },
  ],
});

Because it accepts the Anthropic Messages format too, a multi-provider setup behind an LLM gateway needs almost no adapter work. New sign-ups get $20 in free credits before pay-as-you-go billing starts. Public preview is US developers only as of launch — a real constraint for teams in Tunisia, Saudi Arabia, and the broader MENA region, where access will require waiting for regional availability or routing through a US-based deployment.

Where the cost math actually lands

The price gap is the strongest argument for Muse Spark, and it compounds in exactly the workloads it is built for.

Consider a long-horizon agent that consumes 400K input tokens and produces 60K output tokens per run — realistic for a research or migration task that reads a lot and writes a focused result.

• Muse Spark 1.1: (0.4 × $1.25) + (0.06 × $4.25) = $0.755 per run
• Claude Opus 4.8: (0.4 × $5) + (0.06 × $25) = $3.50 per run

At 2,000 runs a month, that is roughly $1,510 versus $7,000 — about a 78% reduction. For an agent fleet doing document processing, compliance sweeps, or internal-tool orchestration, that difference decides whether the project ships.

But the math only holds if quality holds. Two cautions from early users: throughput sits around 100 tokens/second, and the model has a reported tendency to overthink coding problems — burning output tokens on deliberation for tasks where a cheaper, more decisive model would simply answer. Output tokens are the expensive side of the ledger, so overthinking erodes the very advantage you bought it for.

A practical routing strategy

The right posture is not "switch to Muse Spark." It is route by task shape:

1. Long-context orchestration, tool-heavy workflows, computer use → Muse Spark 1.1. This is its home turf, and the price advantage is largest exactly where token volume is highest.
2. Repository-level coding, bug fixing, refactors → Claude Opus 4.8, Fable 5, or GPT-5.6 Sol. The 8-to-14-point benchmark gap is real work quality, not noise.
3. High-volume simple classification and extraction → Claude Haiku 4.5 or a small open model. Muse Spark's reasoning depth is wasted here, and its overthinking tendency actively hurts.

If you are already running a multi-model fallback strategy, Muse Spark slots in cleanly as the long-context tier. If you are not, its arrival is a good reason to build one — a fourth serious contender in the agentic model market means pricing pressure will keep moving, and hard-coding a single provider leaves money on the table.

The bigger signal: Meta's open-only era is over

The strategic story matters as much as the model. Meta built its AI reputation on open weights — Llama was the default self-hosted model for years, until Qwen overtook it. Muse Spark 1.1 is closed, paid, and API-only.

That is a deliberate pivot toward monetizing inference rather than seeding an ecosystem. Early partners — Replit, Cline, and Box — are all tool builders, not researchers. For teams that chose Llama specifically for sovereignty and self-hosting, the signal is worth reading: Meta's frontier work may not be coming to your own infrastructure anymore.

The verdict

Muse Spark 1.1 is the best-value long-context agentic model available today, provided you use it for what it is. It leads on tool orchestration, ships genuinely useful context management and subagent delegation, and costs roughly a quarter of the flagships.

It is not the best coding model, and treating it as one will produce mediocre results at a discount — the worst trade in engineering. Benchmark it against your actual workload before committing, watch your output-token spend for signs of overthinking, and if you are outside the US, plan around the regional availability gap.

Need help designing a multi-model routing layer or evaluating agent models against your own workloads? Talk to the Noqta team — we build production agent systems for businesses across Tunisia and the MENA region.