Gangsta AI
Meta Shipped Its Third AI Model in Four Months. The Cheapest Tier Costs You Your Own Code.
The Pass

By Gordon Ramsay · 2026-08-07 · 4 min read

Right, listen up. Meta Superintelligence Labs has sent out its THIRD coding model in four months. Four months! Some kitchens change the specials board slower than this lot ships frontier AI. On Tuesday they plated up Muse Spark 1.2 and — for the first time — a coding agent called Muse Code, a terminal-based line cook meant to grind through real software engineering while you sit back and watch.
And you know what? Parts of it are actually cooked. Finally.
The dish
Muse Spark 1.2 lands at 57 on the Artificial Analysis Intelligence Index. For the money that is a genuinely tidy plate — well above average for its price band. Standard pricing is $1.25 per million tokens in, $4.25 out, with a context window north of a million tokens. Muse Code ships alongside it in beta, co-trained with the model so the two actually talk to each other in the pass instead of screaming across the kitchen.
Is it the best in the world? No. It is not touching Claude Opus 5 at the top of the leaderboard, and OpenAI's GPT-5.6 family is still setting the tables in the front of house. But at that price, for coding? It's not raw. It's not swimming in its own hallucinations. It's a solid mid-tier plate and I'll say so.
Then they ruined the sauce
Here's where I put my head in my hands.
Meta has a second menu. A contributor tier. Ten cents per million in, twenty cents out — a 12.5x discount on input, over 21x on output. Cheaper than a bag of chips. Sounds gorgeous, yeah?
Read the small print, you donkey. To get that price you hand Meta permission to train its future models on your prompts and your completions. Your code. Your half-finished ideas. The secret recipe you've been guarding for two years. You're not the customer at that table — you're the supplier, and you're paying for the privilege of being cooked.
“They're not selling you a discount. They're buying your kitchen, and charging you rent to watch.”
Oh, and they cap you at 60 requests a minute on that tier, so you can't even work fast while they harvest you. It's an idiot sandwich with extra mayonnaise.
The bigger service
Strip away the flambe and here's the real story: the frontier is now a price war fought on your data. Meta is racing OpenAI and Anthropic so hard it will trade compute for the one thing it can't scrape fast enough — real developers doing real work. The discount is the bait. Your repository is the catch.
Use the standard tier if the code matters. Use the contributor tier only on the throwaway stuff you'd happily nail to the front door. And never — NEVER — confuse a low price with a free lunch.
Taste before you trust
Here's the only rule that survives a kitchen this chaotic: never trust a single cook. One model tells you it's perfect; the next one finds the salmonella. Put Muse Spark up against ChatGPT, Claude and Gemini on the same question and watch them turn on each other — that's where the truth is, in the arguing. Run them side by side at the best AI comparison. Then, and only then, do you send the plate. Done.
Sources / Receipts
- Yahoo Finance — Meta debuts Muse Spark 1.2 and first coding agent as it ramps up competition with OpenAI, Anthropic (Aug 5, 2026)
- Artificial Analysis — Muse Spark 1.2 intelligence, performance & price analysis
- OrcaRouter — Muse Spark 1.2 & Muse Code: Meta's 3rd Model in 4 Months, explained
- Hero photo: Mark Zuckerberg, 2025 — The White House / AH, Wikimedia Commons (public domain)
More: Best AI models · Compare all AI · Frontier Models · All articles