Presented by
Meta announced something quieter than a new phone but bigger than a new filter on July 24. Meta AI can now make a plan, touch your calendar and inbox, and carry a task through to the end without you checking in at every step. The model behind it, Muse Spark 1.1, shipped on July 9, and it was built for exactly this kind of multi step work.
Tell it you're renovating a kitchen. It scouts Marketplace for furniture inside your budget and sends back a mood board. Ask for a half marathon training plan and it adjusts week by week around your actual calendar, then checks in every Monday. That's the pitch, and the benchmark numbers behind it back up more of it than most agent launches manage.
Completed Timesheets That Are Actually True
There are two ways to get a timesheet. Someone reconstructs the week from a calendar and rounds toward whatever feels safe. Or the record is read from the work as it happens - the right client, the right project, the reason attached. Only one of those survives a client dispute. Timeglass does the second one. It runs quietly on macOS or Windows, drafts your week for you, and you review and release it. The chasing stops. The data holds up. Start free.
WHAT IT ACTUALLY DOES WELL
The benchmarks hold up. On JobBench, the test built to measure exactly this kind of professional tool use, Muse Spark 1.1 scored 88.1, ahead of Claude Opus 4.8 at 82.2 and GPT 5.5 at 75.3, according to results reported by DataCamp. Meta also says the new version resists jailbreaks and prompt injection better than the first Muse Spark, with lower hallucination rates and less sycophancy, based on its own safety evaluation.
For a working professional, the useful part isn't the mood board. It's the daily briefing that catches a double booking before it becomes a problem, or the research pass that pulls from papers and creator posts so you skip an afternoon of tab switching. You can also interrupt it mid task. Tell it to change direction or cut a section, and it adjusts without starting over.
WHAT THE ANNOUNCEMENT LEFT OUT
Since December 16, 2025, conversations with Meta AI feed into ad targeting across Facebook, Instagram, and WhatsApp. There's no opt out unless you're in the EU, UK, or South Korea. An assistant that now reads your calendar and drafts things on your behalf sits on top of that same policy.
Muse Spark also has no open weight release, unlike the Llama models Meta built its early reputation on. Nobody outside Meta can inspect how it actually behaves. The safety documentation Meta published runs thinner than the system cards labs like Anthropic and OpenAI ship at launch. One commenter on Hacker News, quoted by the newsletter Handy AI, put it plainly: they couldn't think of a company they'd trust less with more personal data. That's a hard line, but the complaint underneath it, that convenience and data collection are shipping in the same release, holds up.
WHERE THIS LEAVES YOU
None of this makes the feature bad. It makes it a trade you should notice before you accept it. An assistant this useful needs access to the parts of your life that make it useful, your calendar, your inbox, your shopping habits. Meta's bet is that the same access is what makes it better than a chatbot with no memory of you.
If you're testing Meta AI for real tasks, decide now what you're comfortable connecting, and go check your ad settings before you forget they're there.
The real question by next year won't be whether an AI can act for you. It already can. It'll be which company earned enough trust to let it.
7 Strategies. Zero Guesswork.
The brands winning on Amazon stopped guessing. They know which external channels pull weight — and which ones don't. Levanta's free playbook breaks down 7 strategies driving millions in off-Amazon revenue.
Meta AI features and updates announced July 24, 2026
Task planning and follow through, not just single answers
Calendar and email connection so it can plan around your actual schedule
Daily briefings, pulls what's on your calendar, flags double bookings or changed plans, delivered at a set time
Standing tasks, set up once (weekly meal plan, sneaker drop alerts, trend digest) and it keeps running them without a reminder
Research mode, pulls from the web, research papers, and creator or community posts across Meta's apps
Slide generation from that research, turns findings into a deck
Real time steering, you can redirect it mid task, change the tone, cut a section, and it adjusts without starting over
One place for everything it makes, training plans, decks, mood boards, so you can revisit, build on, or share them
Rolling out today in select markets on the Meta AI app and meta.ai, with WhatsApp and more countries coming in the following weeks
Incognito chats still available for fully private conversations
Shopping with inspiration pulled from posts, Reels, and products shared across Meta's apps
Muse Spark 1.1, the model behind it, released July 9, 2026
Meta's second Superintelligence Labs model, successor to the original Muse Spark from April 2026
Multimodal reasoning model with a 1 million token context window it manages actively, pulling up earlier work and compacting what matters
Built for agentic work, can act as the main planner or hand off steps to subagents
Computer use, navigates real interfaces across multiple apps, adapts as information changes on screen mid task
Coding agent capability through the new Meta Model API, which is OpenAI compatible
On JobBench, scored 88.1 against Claude Opus 4.8 at 82.2 and GPT 5.5 at 75.3, per DataCamp's reporting on Meta's benchmark data
Meta says it resists jailbreaks and prompt injection better than Muse Spark 1.0, with lower hallucination rates and less sycophancy, per its own safety evaluation report
Meta's Advanced AI Scaling Framework evaluation found it within safe margins on chemical and biological, cybersecurity, and loss of control risk categories, according to Meta
API pricing around 1.25 dollars per million input tokens and 4.25 dollars per million output tokens, with 20 dollars in free credits to start, per DataCamp
No open weight release, a departure from the open Llama models Meta built its earlier reputation on
The first way to trade directly inside Claude and ChatGPT
Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else.
Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT.
The institutions built the game, Co-Invest gives you a way to beat them.
Context worth knowing
Since December 16, 2025, conversations with Meta AI feed into ad targeting across Facebook, Instagram, and WhatsApp, with no opt out outside the EU, UK, and South Korea
Meta AI launched as a standalone app in April 2025, originally powered by Llama 4
Meta AI passed 500 million users about a year after launch, according to Zuckerberg's earnings call comments
Zuckerberg has also teased agentic shopping tools for 2026, aimed at completing transactions inside Meta's apps using its access to personal and social data.
One last thing before you go. This isn't a newsletter you read once and consider handled. It works more like a standing check in. It flags what changed before you'd notice on your own, and points you toward what's worth doing next. The actual testing, the deciding what to connect and what to keep separate, still happens in your own week. Coming back to this regularly, issue after issue, is what tends to move that work forward.
