In partnership with
Meta launches Muse Code with a 21x discount tier that trains on your code 🧑💻💰
TL;DR
Meta entered the AI coding agent market with Muse Code, a terminal-based agent powered by the new Muse Spark 1.2 model carrying a 1 million token context window. The headline: a contributor tier priced 21 times cheaper than standard, with the tradeoff that Meta trains on your code.
Key Takeaways
Muse Code runs on macOS and Linux, with the model and agent co-trained as a single unit rather than a model bolted onto an existing interface.
Muse Spark 1.2 carries a 1 million token context window and finishes second to Anthropic's Claude on benchmarks.
The contributor tier is 21x cheaper than standard pricing, but Meta gets training rights on your code in exchange.
Mark Zuckerberg and Chief AI Officer Alexandr Wang are both attached to the launch, signaling strategic priority.
For anyone working on proprietary code or client projects, the contributor tier's data licensing terms significantly affect the economics.
Why It Matters
The 21x pricing gap is designed to attract budget-conscious solo developers and early-stage teams into Meta's ecosystem while feeding its model-training pipeline. For solopreneurs building internal tools or open-source projects, the economics are compelling. For anyone shipping client work or proprietary products, the tradeoff deserves a careful read before committing.
This move also signals that Meta views coding agents as a strategic battleground worth subsidizing aggressively. With Anthropic, OpenAI, and now Meta all competing for developer workflows, the cost of AI-assisted development is dropping fast, but the terms attached to those discounts vary wildly.
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
📰 In the News
Headlines & Launches 📣
In June 2025, Commonwealth Bank of Australia deployed a voice bot and cut 45 support roles one month later. By August, every redundancy had been reversed and the bank issued a written apology. The bot's containment rate looked good on paper, but it missed the residual call volume that still required humans. A cautionary case study for any founder planning to cut headcount based on a single AI metric.
OpenAI is rolling out multi-product carousel ads in ChatGPT, pulling product data directly from retailer feeds similar to Google Shopping. Meanwhile, LinkedIn is actively suppressing AI-generated posts, penalizing content that reads like it came straight from a writing tool. If you run ads or post on LinkedIn, both shifts demand attention this week.
University of Cambridge researchers found that human reviewers failed to catch roughly one-third of AI agent requests that could lead to serious security problems. The study mimicked real sprint conditions with time pressure and familiar interfaces. Persuasive agent explanations and routine-looking actions made reviewers less vigilant. If you rely on human-in-the-loop as your safety net for AI coding agents, this research suggests that net has a significant hole.
SOCi audited over 350,000 business locations and found ChatGPT recommends about 1.2% of them. Consumer use of AI to find local services jumped from 6% to 45% in a single year. If your business isn't in that 1.2%, there's no dashboard showing you the customers who never learned you existed. The panel at 727 Tech Day broke down the four asset categories AI reads to decide who's real.
Hot New Tools 🧰
Agent Skills is an open-source framework that guides AI coding agents through a consistent define-plan-build-verify sequence. Most AI coding integrations lack a formal workflow layer, letting agents guess at action order. For solo developers and small teams, this structured loop is the difference between an agent that speeds up delivery and one that creates cleanup work. Written up by Dave R, Microsoft Azure and AI MVP.
Prime Intellect released Prime Agent under the MIT License on August 6. Instead of handing the model tool schemas and compacting context, it gives the model a single persistent IPython kernel where tools, skills, and sub-agents all operate as Python code. The self-refining loop lets the agent revise outputs within the same reasoning session. MIT licensed, so you can fork it and ship it commercially without friction.
Tencent Cloud open-sourced TencentDB Agent Memory v2.0, a self-hosted memory layer for teams running multiple AI coding agents. It organizes context into four asset types: Chat Memory, Skill, LLM-Wiki, and Code-Graph, all versioned and permissioned at the asset level. ACL-based visibility means different agents on the same team can operate with different memory views. MIT licensed, self-hosted via Docker, integrates with Claude Code, OpenClaw, Hermes, and CodeBuddy.
A developer called GOROman built Vibe Watch, an M5Stack-based wrist controller for managing multiple AI agents hands-free. The round display shows six live agent indicators at once, with color and animation reflecting each agent's state. Hardware buttons approve or reject actions with distinct audio feedback. MIT licensed, open source, roughly three hours to build.
Miscellaneous 🎁
Amjad Masad discusses the operational reality of running Replit as AI handles an increasing share of output. The interview focuses on how AI changes team structure, what job losses look like up close, and where software is headed. If you're thinking about how AI reshapes your own team, this 23-minute interview is grounded in a product that has already gone through that transition.
CapCut PC now integrates Seedance 2.5 directly into its editing environment, covering the full pipeline from text prompt to exported video. Three AI editing tools handle scene modification, element removal, and clip extension. The interesting part for operators: keeping generation and editing inside one multi-track environment cuts context-switching and speeds up the review-to-export loop.
Thanks for reading,
— Cagri Sarigoz
P.S. Don't Miss Out on These Resources:
🤝 Follow me on X and LinkedIn for regular updates and insights.
