{"slug": "meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week", "title": "Meta's New Personal Agent is Free Up to 100M Tokens per Week", "summary": "Meta has launched Muse, a personal AI agent available through its app, web, and WhatsApp, offering a free tier of up to 100 million tokens per week. Muse operates with a dedicated cloud computer and browser, can perform tasks like booking travel and negotiating bills, and includes a Sentinel agent for safety oversight. Separately, Anthropic researcher Jacob Coxon resigned, warning that frontier labs are gambling with lives by racing to develop AI despite fears of existential risk.", "body_md": "- [unwind ai](../)\n- Posts\n- Meta's New Personal Agent is Free Up to 100M Tokens per Week\n\n# Meta's New Personal Agent is Free Up to 100M Tokens per Week\n\n## + Anthropic researcher says the labs are gambling with our lives\n\n### **Start here ↓**\n\n**Meta brings personal AI agents to your WhatsApp.**\n\nGrok Bot made “AI employees” a product category last month. Now, Meta brings its own version of persistent agents with Muse, available through its own app, web, and WhatsApp.\n\nMuse gets a dedicated cloud computer and browser, then works across email, calendars, Instagram, and ordinary websites. It can book travel, negotiate bills, sell a car, build missing tools, and keep working after the user closes the app.\n\nMuse is a serious attempt to make this level of access safe for consumers: a separate Sentinel agent inspects everything leaving the VM, passwords stay hidden from Muse, sensitive actions require approval, and every action enters an audit trail.\n\nThere is a free usage tier with a generous limit of 100 million tokens per week, paid subscriptions for heavier use, and support for Meta’s AI glasses is coming.\n\n## 🚀 **Shipped**\n\n**Inception Labs released Mercury 2.5 that runs at 1,107 tokens per second.** It has a 260K context window, parallel tool calls, adjustable reasoning, and schema-aligned JSON, currently discounted by 80% at launch. Their “40% more intelligent” claim is vague, but Augment says Mercury cut context-compaction latency from roughly 150 seconds to 27, making it compelling for usecases like search, voice, and agents.[Inception Labs](https://www.inceptionlabs.ai/blog/introducing-mercury-2-5?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**A 17 GB Qwen quant matched the 55 GB original.** Quesma tested Qwen33.8 27B quantizations across GPQA, IFBench, and Terminal-Bench 2.1. The 17 GB Q4_K_M version model showed no measurable loss from BF16 and fits on a 24 GB card with room for roughly 64K tokens of context. Two-bit remained usable; one-bit fell to random chance and got worse when allowed to reason longer.[Quesma](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**OpenAI’s new image model is built to survive your edits.** ChatGPT Images 2.5 brings sharper details, stronger reference-photo fidelity, and up to 50% lower latency than Images 2.0. Its most useful improvement is restraint: change one product, face, background, or line of copy, and the model is less likely to wreck everything else. ChatGPT also gets Sketch, comments, templates, and many other features![OpenAI](https://openai.com/index/introducing-chatgpt-images-2-5/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**A 3B speech model handles 22 languages, laughter, and sighs.** Rumik OSS 1 supports code-switching, four multilingual voices, delivery controls, and precisely placed vocalizations. It beat Cartesia and ElevenLabs on Rumik’s own emotion benchmark but remained far behind Gemini. Released under a non-commercial license.[Rumik OSS 1](https://huggingface.co/rumik-ai/rumik-oss-1?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n## 🧠 **Worth Knowing**\n\n**The frontier labs want applause and an emergency brake at the same time.** AI researcher Jacob Coxon resigned from Anthropic and said both OpenAI and Anthropic privately fear their systems could kill humanity but continue racing because each believes it must arrive first. The thread is not compelling because one researcher predicts doom. But because it turns the industry’s governance premise against itself: if the people who believe the stakes are civilizational still cannot coordinate, “trust the responsible lab” is not a safety plan.[Jacob Coxon’s thread](https://x.com/hilbertspaess/status/2097476196791709843?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Coding agents know testing vocabulary, not testing judgment.** Dan Luu tested 26 instructions for implementing Zstandard in Rust, including TDD, fuzzing, Lean, TLA+, and “Make no mistakes,” with roughly 80 runs per condition and reasoning setting. Default prompting performed above average, while several popular skills underperformed. Agents often wrapped their usual weak tests in a fancier framework or proved properties irrelevant to correctness.[Dan Luu](https://danluu.com/agentic-testing/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**OpenAI says 10,000 agents solved Navier–Stokes.** OpenAI claims its agents exchanged 2.7 million messages and generated 130 billion output tokens to produce a finite-time blowup proof, followed by a Lean formalization. The repository gives mathematicians something concrete to inspect, but the result has not been independently reviewed or accepted by the Clay Mathematics Institute.[OpenAI](https://openai.com/index/navier-stokes-solution/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**A database may be the cleanest agent memory system.** Ben Dicken proposes storing messages, logs, and operational context in Postgres, SQLite, MySQL, or DuckDB, then restricting the agent with views, scoped credentials, result caps, and short timeouts. It is less fashionable than another retrieval layer, but much easier to inspect and enforce. Agents are already good at SQL; the missing piece is governance.[Ben Dicken on X](https://x.com/BenjDicken/status/2096997823812350381?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**A 16 GB Mac can run Qwen3.5-4B at 61 tokens per second.** The 4-bit model uses just 6 GB on a base M4 mini, leaving roughly 10 GB for the rest of the system. Rapid-MLX exposes local models through an OpenAI-compatible API, so the same setup can power Cursor, Claude Code, Aider, and other agent tools without sending prompts to the cloud. Its new recipe command checks your Mac and recommends the smartest model that fits alongside a faster alternative.[Rapid-MLX](https://rapidmlx.com/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**See which earlier tokens shaped every generated token.** This in-browser visualizer lets you hover over Qwen3-0.6B’s output and trace attention back through the prompt. It is especially effective at showing how models copy exact details or combine related phrases. The visualization compresses a lot of internal activity, so use it to build intuition rather than prove why a model made a decision.[LLM Attention Visualization](https://ishamf.dev/p/llm-attention-visualizer/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Vercel squeezed syntax highlighting into a 27.5 KB model.** `gpu-lexer` uses WebGPU to classify code tokens without being told the programming language. Its labels disagree with Shiki 12.57% of the time on held-out files, so it is not ready to replace grammar-based highlighting. It is still a great example of a tiny model replacing language-specific software rules.[gpu-lexer](https://gpu-lexer.vercel.app/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n## 🔧 **Clone and Run**\n\n**Clone & Run of the Day** **Mantis makes security agents reproduce vulnerabilities.** Google’s Mantis provides agent skills for threat modeling, vulnerability discovery, reproduction, patching, and verification. Run it only in an isolated environment and keep a security expert in the loop. [GitHub](https://github.com/google/mantis?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Thirty-one thousand people starred a prompt that tells agents to shut up.** `i-have-adhd` gives coding agents ten rules: lead with the action, number the steps, suppress tangents, and stop ending with “Hope this helps.” Its popularity is less a story about ADHD than a brutal review of default agent communication.[GitHub](https://github.com/ayghri/i-have-adhd?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Claude Style Patch attacks “Claudish” prose one habit at a time.** This drop-in `CLAUDE.md` section targets dense setups, fragments, label-and-colon constructions, and other recurring Claude mannerisms. Each ban includes a concrete repair, which models follow more reliably than a vague request to “write naturally.”[GitHub](https://github.com/andrewroxby/claude-style-patch?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Deltafin runs full Kimi K3 on a laptop at 0.29 tokens per second.** The Deltafin fork streams Kimi K3’s complete 2.8T expert bank from SSDs on Apple Silicon and publishes its manifests and benchmark package. It is far too slow for ordinary use, but that is what makes the artifact useful: it shows exactly where consumer weight-streaming breaks.[GitHub](https://github.com/argonautlabsai/deltafin?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Stop agents from maintaining paperwork instead of shipping.** `forward-implementation-first` targets agents that get trapped rebuilding hashes, receipts, locks, and progress metadata while the real output already works. It forces the agent to separate implementation and focused validation from administrative busywork.[GitHub](https://github.com/Vuk97/forward-implementation-first?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Reef turns feedback into new agent versions.** Reef connects live inference, recorded interactions, feedback, evaluation, training, and versioned deployment. It can update model weights or evolve prompts and skills without interrupting serving. [GitHub](https://github.com/Human-Agent-Society/reef?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n[Awesome LLM Apps](https://github.com/Shubhamsaboo/awesome-llm-apps?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week) **is a curated collection of 100+ AI Agents, Agent skills, and RAG apps.** It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. [(Now accepting GitHub sponsorships)](https://sponsorunwindai.com/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n## 📊 **By the Number**\n\n**Number of the Day****An agent burned 1.3 million tokens and still failed the screenshot check.** Ten model-and-harness combinations received the same Three.js prompt. Astra 6.0 Max on Codex took 37 minutes and 1,335,495 tokens, while GLM 5.3 Flash Max took nine minutes and 475,143 tokens; neither passed both completion checks. The table is too small to rank models, but large enough to show how dramatically the surrounding harness changes cost.[Hangar Harness / Model Tests](https://alvins82.github.io/hangar-harness-model-tests/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\n**Mistral raised €3 billion at a valuation above €21 billion.** Samsung led what Mistral calls Europe’s largest-ever technology equity round, joined by EQT’s Scaleup Europe Fund and PSG Equity. “Sovereign AI” remains marketing until the models compete, but Europe’s strongest open-weight lab now has enough capital to buy the compute needed to prove its case.[Mistral](https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week)\n\nThat's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.\n\nIf you found one thing to try, share the issue with someone who ships.\n\n### The New Rules of Online Visibility\n\nYour customers are searching in places your strategy doesn’t reach.\n\nSo before your business is buried and left behind, you need to understand the new rules of SEO.\n\n BELAY's [SEO in the Age of AI](https://resources.belaysolutions.com/partners/beehiiv-as-one?utm_campaign=22138128-Beehiiv&utm_source=beehiiv&utm_medium=Primary&utm_term=JHL0VVEUDT&utm_content=Page1&_bhiiv=opp_d84291d3-87d8-4978-9ecf-ba6068c61018_3057ce03&bhcl_id=50a5fb0a-1866-42d0-a234-f1a991d76544_SUBSCRIBER_ID_{{email_address_id}}) report explains how search is changing, what AI means for your visibility, and the practical steps small businesses like yours can take to stay visible. \n\nBELAY’s U.S.-based Marketing Assistants turn strategy into execution, helping your business stay visible, credible, and competitive in every search.", "url": "https://wpnews.pro/news/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week", "canonical_source": "https://www.theunwindai.com/p/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week", "published_at": "2026-09-09 12:30:00+00:00", "updated_at": "2026-09-09 12:47:55.920666+00:00", "lang": "en", "topics": ["ai-products", "ai-agents", "ai-safety"], "entities": ["Meta", "Muse", "WhatsApp", "Anthropic", "Jacob Coxon", "OpenAI"], "alternates": {"html": "https://wpnews.pro/news/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week", "markdown": "https://wpnews.pro/news/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week.md", "text": "https://wpnews.pro/news/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week.txt", "jsonld": "https://wpnews.pro/news/meta-s-new-personal-agent-is-free-up-to-100m-tokens-per-week.jsonld"}}