Product launch · XiaoHu explainer

Meta launches Muse Spark 1.1: competitive with GPT-5.5 and Opus 4.8 on multiple agent evaluations

Meta CAIO Alexandr Wang: an industry-competitive agentic and coding model. API pricing: $1.25 input / $4.25 output per 1M tokens.
At a glance
  • Meta Superintelligence Labs ships Muse Spark 1.1 for agentic work; Meta Model API enters public preview; Thinking mode lands in Meta AI / meta.ai.
  • Feature spine: main agent + parallel subagents, 1M-token active context, desktop/browser/mobile computer use, coding, multimodal action.
  • Wang's scorecard: JobBench 54.7, MCP Atlas 88.1, HLE w/ tools 62.1 and more sit with or above GPT-5.5 / Opus 4.8 on several agent lines; coding is near the frontier, not a sweep.
  • Pay-as-you-go: input $1.25, cached input $0.15, output $4.25 per 1M tokens; web search grounding $2.50 per 1,000 queries; no long-context premium.
  • Compatible with OpenAI SDK, Anthropic SDK, OpenCode / Claude Code; three API shapes, one model, one price.
Based on Meta AI's blog, Model API docs and pricing page, the 112-page eval report, and Alexandr Wang's Jul 9, 2026 thread. Benchmarks are official / Meta-published.
1Release

Say what happened first

On July 9, 2026, Meta Superintelligence Labs released Muse Spark 1.1 and, for the first time, put it on Meta Model API for external developers. The same day, Meta CAIO Alexandr Wang wrote:

muse spark 1.1 is an industry-competitive agentic and coding model. across many agentic evals it rivals gpt-5.5 and opus-4.8. available now through the new meta model api and in meta ai. - Alexandr Wang, @alexandr_wang, 2026-07-09

In plain terms: this is not a "chat feels slightly better" patch. It is aimed at agent work and coding. On many agent evals, Meta claims parity with GPT-5.5 and Opus 4.8. Below: features first, then each layer, then scores, then money.

2Features

What it can do: one map before the deep dive

Read this table once. Everything later only expands a cell.

Positioning
Multimodal reasoning model for agentic tasks: tools, computer use, coding, multimodal understanding.
Multi-agent
Main agent gathers context and plans; parallel subagents execute; subagents know tools and escalate when stuck.
Context
1,048,576 tokens; active management: remember actions, retrieve early facts, compact while keeping critical steps.
Computer use
Desktop, browser, and mobile. Script when faster; click when simpler; batch actions per step.
Coding
Works with common agentic coding harnesses; large repos, multi-turn, screenshot UI checks.
Multimodal
Images, video, documents; details persist across long workflows and feed into browser/desktop actions.
API surface
Parallel tools, streamed tool args, cross-turn reasoning, search with citations, structured JSON, Files API, prompt caching, tunable reasoning_effort.
Compatibility
OpenAI SDK / Anthropic SDK / OpenCode / Claude Code; Responses, Chat Completions, Messages share one model and one price.
Where
Consumer: Thinking mode in Meta AI app and meta.ai. Developers: Meta Model API preview (api.meta.ai/v1, muse-spark-1.1).
Early partners
Replit, Box, Cline, and others.

One-line spine: learn to assign parallel work like a project manager, bind desktop/browser/mobile and coding into one agent pipeline, then ship it via a high-score, relatively low-cost API.

3Layer 1

How agents work: main assigns, subs run in parallel

  1. Main agent plansGather context, make a plan, split work instead of clicking every UI step alone.
  2. Subagents execute in parallelStay on task, know available tools (native, MCP, custom skills), escalate when needed.
  3. Active context management1M tokens are managed: remember actions, retrieve early work, compact chatter while keeping decisions and open steps.
Main agent
context + plan
Sub A
code / tools
Sub B
browser / desktop
Sub C
search / verify
Deliver

Zero-shot generalization to new tools/MCP/skills is claimed. WideSearch shows multi-agent above single-agent at the same latency proxy.

WideSearch multi-agent vs single-agent
Official chart · WideSearch: multi-agent (solid) above single-agent (dashed)
4Layer 2

Computer use: desktop, browser, and mobile

Wang is explicit: desktop, browser, and mobile. Training target:

  • Write a script when automation is faster
  • Click the UI when direct interaction is simpler
  • Emit batches of actions per step, not one click at a time
Think of it this way

Give an intern a remote device and only a goal. They choose whether to open an app and click a few times, or script bulk checks of menus and inventory.

Official demo · Agentic dinner party organization · Source: Meta AI blog
Computer-use clip from Alexandr Wang's thread · Source: @alexandr_wang
OSWorld 2.0 cost curve
Official chart · OSWorld 2.0: USD per task vs score; blue = Muse Spark 1.1
5Layer 3

Coding: large repos, multi-turn, screenshot self-check

Official demo · Build, screenshot, fix in OpenCode · Source: Meta AI blog
Official demo · DeepSWE multi-effort self-eval + dashboard · Source: Meta AI blog
Vibe Code and SWE Atlas
Official chart · Vibe Code 72.2 (prior 19.7); SWE Atlas QnA 42.0 (prior 24.2)
Meta internal coding bench
Official chart · Meta Internal Coding Bench: 1.1 = 68.3 vs Opus 4.8 = 69.0
6Layer 4

Multimodal: see details, then act

Official demo · Phone video → extract photos → list on Facebook Marketplace · Source: Meta AI blog
Related on this site
Muse Image / Muse Video

Media generation line. Image mechanics live there; this piece covers Spark's agent / coding / API line.

7Scores

Multiple agent evals: where it leads, where it only matches

Full table from Wang's thread (same numbers as Meta's Muse Spark product page). Everything visible—no tab switching.

Wang benchmark table
@alexandr_wang thread chart · Muse Spark 1.1 vs Gemini 3.1 Pro / Opus 4.8 / GPT-5.5
ClassBenchmarkSpark 1.1PriorOpus 4.8GPT-5.5
AgentMCP Atlas88.182.282.275.3
AgentJobBench54.717.048.438.3
AgentToolathlon-Verified75.649.476.273.5
AgentOSWorld-Verified80.853.383.478.7
AgentHLE (w/ tools)62.150.457.952.2
AgentFinance Agent v257.2-53.951.8
CodingTerminal-Bench 2.180.067.382.783.4
CodingSWE-Bench Pro61.555.069.258.6
CodingDeepSWE 1.153.310.059.067.0
MultimodalCharXiv Reasoning88.488.989.984.8
MultimodalBabyVision76.339.981.283.6
How to read it

Where Meta claims lead: MCP Atlas, JobBench, HLE w/ tools, Finance Agent v2 often top or near-top.

Neck-and-neck: Toolathlon and OSWorld sit next to Opus; Terminal-Bench slightly behind GPT-5.5 / Opus.

Still behind: DeepSWE 53.3 vs GPT-5.5 67.0; BabyVision 76.3 vs GPT-5.5 83.6. Not a clean sweep.

Vs prior Muse: JobBench 17.0→54.7, DeepSWE 10.0→53.3, OSWorld 53.3→80.8.

JobBench
Official chart · JobBench: 54.7 vs Opus 48.4 vs GPT-5.5 38.3
MCP Atlas
Official chart · MCP Atlas: 88.1 leads Opus / prior at 82.2
8Price and API

What it costs, how you connect

Wang's line is "premium performance at low cost." Official pay-as-you-go pricing (USD per 1M tokens, no minimum):

$1.25
Input / 1M tokens
$0.15
Cached input / 1M tokens
$4.25
Output / 1M tokens
$2.50
Web search / 1,000 queries (+ tokens)
Meta Model API pricing
@alexandr_wang pricing chart · matches Meta Model API pricing page
  • No long-context premium: same rate empty or nearly full
  • Prompt caching bills matched prefixes at cached-input rate
  • Responses / Chat Completions / Messages: same model, same price; Responses is fullest for agents
  • Rate limits per team: Free 60 RPM / 2M TPM; Paid 3,000 RPM / 4M TPM
  • Small injected steering tokens are not billed
Connection
base_url = https://api.meta.ai/v1
model    = muse-spark-1.1
key      = MODEL_API_KEY (dev.meta.ai)
"What's most impressive about Muse Spark is how much it packs into one model: massive million-token context, full multimodal support, built-in search with citations, strong reasoning, top-tier coding abilities, structured output, and parallel tool calling - all in a clean OpenAI-compatible package." - Amjad Masad, Replit CEO
9Safety

Stronger capability, tighter risk evals

The 112-page eval report: without mitigations, chem/bio and cybersecurity cannot be ruled out at Framework "high risk"; with deployment mitigations, residual risk is "moderate or lower." Loss of control stays moderate or lower. Cyber capability jumps (Cybench pass@1 65.4→92.9) while jailbreak / injection ASR falls (StrongREJECT 25.2→0.5; AgentDojo 11.9→0.7; Agentic Misalignment 47.7→1.1).

10Who benefits

Who should try it

Agent products

Parallel tools + 1M context + main/sub orchestration; OpenAI/Anthropic-compatible harnesses can retarget endpoints.

Coding agents / frontend

Large repos and screenshot UI checks; internal coding nearly ties Opus; cached input $0.15 helps long runs.

Cross-app desktop flows

Desktop + browser + mobile computer use; mid-task goal changes are a featured demo pattern.

Multimodal errands

Video/screenshot in, then drive a browser to list, fill, submit.

Sources: Meta AI Blog (2026-07-09) · dev.meta.ai docs and pricing · Muse Spark product page · Muse Spark 1.1 Evaluation Report · Alexandr Wang thread 2075218936266998230. Demos on pic.xiaohu.ai.