Muse Spark 1.1: Meta's Agentic Coding Model That Costs 75% Less Than Rivals

Meta was written off in the AI race. After Llama 4 underwhelmed and OpenAI and Anthropic pulled ahead, the narrative was simple: Zuck missed. Then on July 9, 2026, Meta Superintelligence Labs dropped Muse Spark 1.1 — and the narrative cracked.

This isn't another open-weight Llama release. This is Meta's first real proprietary API play, led by Alexandr Wang (yes, the Scale AI founder), rebuilt from scratch over 9 months, and priced at 75% below rivals. It has 1M token context, multimodal reasoning, and it’s genuinely good at agentic coding. For the first time in a year, Meta feels dangerous again.

I dug through the launch, the benchmarks, and the fine print. Here’s what matters.

What is Muse Spark 1.1?

Muse Spark 1.1 is Meta's agentic foundation model — a coding-first, tool-using model designed to work autonomously for hours, not just autocomplete your next function.

It comes from Meta Superintelligence Labs (MSL), the new AGI skunkworks team formed in late 2025 after Meta poached talent from OpenAI, DeepMind, and Scale. The team is led by Alexandr Wang, who joined after Meta's $14B investment in Scale AI.

Key distinction: Spark is NOT Llama 5. It's a separate lineage, rebuilt "from the ground up" — new architecture, new training stack, new data curation. Meta is keeping the weights closed (for now) and selling it via API, marking a fundamental strategic pivot away from pure open-source. Llama isn't dead, but Meta finally admitted that to compete at the frontier, you need a business model that prints cash.

Spark 1.1 is currently in developer preview via Meta's new AI API platform. General availability is expected Q3 2026.

Key Specs Comparison

Let's cut through marketing:

*The 94M figure refers to Meta's claim around efficient long-horizon task throughput — essentially how many tokens it can reason over in a single agentic session without degradation. It's not a single-shot output limit. Marketing? Partly. But the underlying long-context retention is real.

The headline: 1M context at launch, multimodal reasoning that actually sees your Figma files and error screenshots, and native computer-use capabilities. It can click, scroll, type in a sandboxed browser. This is table stakes now for agentic models, but Meta's implementation is reportedly smooth.

Architecture & Training: Rebuilt From Ground Up Over 9 Months

Meta is vague on architecture details (as expected for a closed model), but here's what we know from their technical report and Wang's interviews:

1. From-scratch rebuild: MSL threw away much of Llama 4's stack. This is a 9-month sprint with a new mixture-of-experts (MoE) backbone optimized for inference efficiency — hence the cost advantage. Meta claims 3x better throughput per dollar than Llama 4 Behemoth.

2. Data engine from Scale: This is why they bought Wang's company. Spark's data pipeline uses Scale's synthetic data and eval infrastructure heavily. Focus on high-quality code commits, real GitHub issues, tool trajectories, and long-horizon agent traces — not just raw web scrape.

3. Long-context focus: Native 1M context with RoPE extensions and a new hierarchical attention mechanism that doesn't collapse at 500K+ like many competitors. For codebase-level tasks, this is a game changer — you can actually load a real repo.

4. Intelligence Index gains: Meta reports an 8-point jump on their internal Intelligence Index over the last 3 months of training alone, suggesting steep scaling curves late in the run. That's the kind of acceleration you want to see.

The takeaway: This is Meta operating like an AGI lab, not a research lab that publishes papers. Fast, secretive, product-driven.

Benchmarks vs GPT-5.6 and Claude Sonnet 5

Benchmarks are always cherry-picked, so read these with skepticism. But Spark 1.1 posts numbers that demand attention.

Meta claims it beats GPT-5.6 and Claude Sonnet 5 on:

• SWE-Bench Verified: 68.4% (vs 64.1% GPT-5.6, 65.8% Sonnet 5)
• Terminal-Bench 2.0: Top score with 48.2% agentic task completion
• OSWorld (Computer Use): 41% success on real desktop tasks
• LiveCodeBench: 72.1%

It also shows strong results on tool-use evals like BFCL v3 and Tau-Bench.

Reality check: Meta carefully avoided comparing to the absolute top tier — OpenAI's rumored GPT-6 preview and Anthropic's Claude Opus 4.8, which still lead on reasoning and hard math (AIME, GPQA). Spark 1.1 wins on agentic coding efficiency and mid-tier reasoning, not on raw IQ.

That positioning is smart. Most developers don't need a PhD-level mathematician — they need a reliable junior-to-mid-level engineer that doesn't charge $150/hour and can work overnight on a ticket queue. That's Spark's lane.

Pricing Revolution: 75% Cheaper and That Changes Everything

This is the real story.

Meta is pricing Muse Spark 1.1 at ~75% below GPT-5.6 and Claude Sonnet 5 for equivalent input/output tokens. Early developer preview pricing leaked at roughly $1.00 / $4.00 per million tokens (input/output) vs $5 / $20 for competitors.

Why can they do this?

1. Efficient MoE architecture = lower inference cost
2. Meta's insane infra advantage (their own data centers, custom silicon)
3. Strategy: Buy market share. They lost a year — now they're subsidizing.

If you're running agentic loops that burn 50K-200K tokens per task, a 75% cut isn't a discount — it's a different business model. Startups building coding agents (Cursor, Windsurf, Cline) can suddenly offer 4x more agentic depth for the same cost. That's how you win a platform shift.

This is classic Zuck: scale product meets model price war. He did it with Llama open weights to commoditize OpenAI's moat, and now he's doing it with cheap frontier inference.

Expect OpenAI and Anthropic to respond within 30 days.

Developer Experience & API

The API itself is… surprisingly good for a v1.

• OpenAI-compatible endpoint (drop-in replacement with base_url change)
• Native tool calling with parallel execution
• Built-in code interpreter and browser sandbox
• Streaming with function calls intact
• 1M context actually works in practice, with ~92% needle-in-haystack recall at full length (per independent tests)

The docs live at developers.meta.com/llama-api (yes, confusing name — it's still branded Llama API but serves Spark now). Auth is via Meta developer account, not yet self-serve credit card for all — you need to apply for preview, but approvals are reportedly fast (<24h).

Early developer feedback on X: fast, follows instructions better than Llama 4, less sycophantic than Claude, better at bash/tool errors than GPT-5. The coding agent demo where it builds a full-stack app from a prompt in 45 minutes is impressive, though curated.

Limitations: What Meta Isn't Telling You

I'm bullish, but let's be honest:

1. No comparison to true frontier models: They didn't bench against Opus 4.8 or GPT-6. That's intentional. Spark 1.1 likely still loses to those on hard reasoning.

2. Vision is weaker than GPT-5.6: Multimodal is there, but chart understanding and complex visual reasoning lags.

3. Closed weights = lock-in: After years of "open source AI for everyone," this is 180° turn. You can't self-host, you can't fine-tune deeply, and you're at Meta's mercy on pricing after the promo period.

4. Safety evals light: Meta's system card is thin compared to Anthropic's. For enterprise adoption, that's a red flag.

5. Ecosystem maturity: No equivalent to Claude Code or Codex yet. It's just an API, not a product.

My Take as an AI Engineer

We needed this.

The last 6 months became a duopoly — OpenAI and Anthropic trading blows at $10-20/M tokens while everyone else watched. That kills innovation. Meta entering with a legit agentic coder at 75% cheaper forces everyone to move faster and cheaper.

Is Spark 1.1 AGI? No. Is it the best coding model period? No, Opus 4.8 still has the crown for the hardest tasks. But is it the best value coding model for 90% of real engineering tasks — ticket triage, repo-wide refactors, test generation, CI fixes? Right now, yes.

If you're building AI-powered dev tools, you should be testing this today. If you're a founder deciding on your default model for agents, run the math. At 1M context and $1/M input, you can afford to give your agent the whole codebase — and that's a new paradigm.

Meta is back. Don't sleep on it.

---

FAQ

What is Muse Spark vs Llama? Llama is Meta's open-weight model family you can self-host. Spark (formerly "Muse") is Meta's new closed, proprietary agentic model sold via API. Spark is built by the new Superintelligence Labs team led by Alexandr Wang, trained from scratch, and optimized for coding and tool use. Think of Llama as Android and Spark as Pixel — same company, different strategy.

How much cheaper is Muse Spark 1.1 really? Meta claims 75% cheaper than rivals. In developer preview, it's ~$1 input / $4 output per million tokens vs ~$5/$20 for GPT-5.6 and Claude Sonnet 5. For agentic workloads that burn 100K+ tokens, that saving is massive. Final GA pricing may shift.

What is the context window of Muse Spark 1.1? 1 million tokens at launch, with efficient handling of long-horizon tasks (Meta claims 94M tokens effective throughput in a session). It supports text, code, and vision within that context.

How to access the Muse Spark 1.1 API? Apply for developer preview at developers.meta.com. Auth uses Meta developer accounts with OpenAI-compatible endpoints. Self-serve GA with credit card billing is expected Q3 2026. Approval is currently fast (<24 hours).

Published on EssaMamdani.com - AI Engineering Notes from the Field