🍭 Your $30 AI bundle has a catch 🧾

The bundle is back

Good morning. Your AI bill is the new battleground. SpaceXAI folded Grok, Cursor and X into one $30 plan (previously ~$90), while Meta pushed into the frontier with an agent that remembers the brief. Before switching, check two things: what costs extra, and what the new model/agent can actually do today.

Let’s Dive In 👇️ 

Today’s Menu 🥠

• Your AI bill: the catch in X’s $30 bundle

• Meta’s New Frontier Model w/ memory: what Muse Wonder can do now

• Fast Snacks: Character.AI’s bankruptcy, new AI chips, model copying, replication research and smarter decisions

• Toolbox: voice commands, agent hosting and local models

• Test your eye: which robot photo is real?

Tuesday, October 13 · Read Time: 6 minutes

The Big Bite · Your AI Bill
💰️ X’s $30 Bundle Comes With a Billing Catch

The Bite: Last Wednesday, Elon Musk said SpaceX will use “the best back end model for any given task, including Claude Opus 5.5.” On Monday, SpaceXAI (its website now says SpaceXSI in places) announced The X Subscription, folding Grok, Grok Bot, Cursor and X Premium into one plan. Premium costs $30 a month before taxes and fees. The company says those benefits were “previously $90/month combined.” That comparison is the company’s, not a promise about your bill.

Why it Bites: Grok usage comes from one pool shared across Grok and Cursor, and it resets weekly. Other companies’ models in Cursor sit outside that pool: they are “billed separately at API list prices, through top-ups.” You can set spending caps. Even the $200 Ultra tier includes no allowance for third-party models. Grok Bot works differently: it picks its own model, sometimes Claude Opus 5.5, and its usage draws from the shared pool at no higher per-token cost. If Claude or GPT does most of your coding, the $30 sticker leaves out your biggest line item.

Our take: Price the models you actually use before you switch. Already subscribed? Keep your plan, or switch for a prorated refund. If you love Grok Bot, consider comparing the quality and output for the level of tasks that you perform. We love Grok Bot for some use cases so far, so highly depends the level of work you are carrying out.

Where Grok stands (our analysis): Grok trails the top tier on both major leaderboards (see the Frontier scoreboard below). So the question: is $30 for Grok, Cursor and X a bargain, or a sign SpaceXAI is competing on distribution rather than on the top model? When its own agent reaches for Claude, and Claude in Cursor costs extra, the bundle looks priced for reach, not for a model lead. That’s our read, not the company’s explanation. However, we do love grok bot and its capabilities.

Second Bite · Meta
🧠 Meta Wants AI to Remember the Brief

The Bite: Six weeks after Muse Spark 1.3, Meta announced Muse Wonder on Monday, calling it “a frontier model that leads on personal and professional agents.” Independent scores back part of that: Artificial Analysis puts Muse Wonder Full at 56, tied for second behind Claude Opus 5.5 at 58. Active Memory pulls relevant context from earlier conversations. Active Focus keeps it working, asking for clarification and checking in “before taking important actions.”

The catch: Wonder Full, the model behind those scores, is only for Muse Code subscribers. The API gets Muse Wonder, which scores 52 at Claude Opus 5.5’s price of $4 per million input tokens and $20 output. Consumer Muse access is coming soon, subscribers first.

Frontier scoreboard (Artificial Analysis Intelligence Index, best score per model; LMArena text rank in brackets)

1. Claude Opus 5.5 (Anthropic): 58 [LMArena #2]

2. Claude Sonnet 5.5 (Anthropic): 56

3. Muse Wonder Full (Meta): 56 [not yet rated]

4. GPT-6 Astra (OpenAI): 53

5. Gemini 4 Argon (Google): 53 [LMArena #1]

   . . .

   Best Grok: Grok 4.7 (SpaceXAI): 46 [best Grok on LMArena: grok-4.20-beta1, #37]

Source: Artificial Analysis, checked Oct 13; LMArena text leaderboard, Oct 8 update.

Meta now sits level with Anthropic’s second model, two points off Opus. On this index, OpenAI and Google trail Opus by five, and Grok sits 12 back. That gap is the backdrop for X’s bundle leaning on Claude.

Why it Bites: Meta is now in the frontier race on independent scores, not just its own charts. Memory is the pitch: an agent that remembers your audience, constraints and preferences saves you repeating the brief. The test is whether it remembers the right details and notices when they change.

Our take: Try one recurring task with clear limits on what the agent may change. Memory should cut repetition, not widen permissions. Meta’s safety claims are still its own; a full report with additional safety evaluations is promised soon.

Fast Snacks 🥡

🎭 Character.AI Files for Chapter 11
Last month, Disney hired Character.AI’s CEO, Karandeep Anand. On Monday, the chatbot company filed for Chapter 11 in Delaware to sell all or substantially all of its assets, with Disney as the $20 million stalking horse bidder for a non-exclusive perpetual license to certain IP and claims. Rival bids can top it, with a sale expected to close in late November or early December. Why care? After lawsuits and safety scrutiny, the company says that without a workable bid, “the platform will be wound down.”

⚡ Maia 300 Is a 2027 Story
Maia 200 already serves Microsoft production workloads, including Copilot. Monday’s announcement puts its successor, Maia 300, in production deployments in 2027. Microsoft expects “40% more tokens per megawatt compared to leading GPU-based systems,” without naming the GPU. That’s a projection, not independent testing. Why care? More AI output per megawatt could expand cloud capacity, but no customer price is announced.

🔒 Anthropic Makes Model Copying Harder
Anthropic has reported on labs harvesting Claude’s outputs to train copycats since at least February. On Monday it said new defenses have “reduced the amount of data harvested by distillers by around 90 to 99%.” Distillation uses one model’s outputs to train another. Legitimate developers pay part of the cost: preserved thinking forced some integration changes, and since September 24 Anthropic charges for requests blocked before Claude answers in areas with false positives below 0.1%, distillation included. Check your integration and billing logs. Anthropic’s own caveat: “We don’t yet know how long these gains will last.”

🧪 AI Replication Is Still Missing Steps
Last week Anthropic disclosed Claude taking unintended actions on real websites during testing. Monday’s research asks a different question: could small models copy themselves onto cheap devices? In its tests, a 9B model got inference running on 8 of 9 hardware types, and GLM-5.2 replicated end to end across a test fleet, with the hacking step deliberately made trivial. Anthropic: “We don’t yet have a full rogue replicator.” Its year-end timeline is a trend-line forecast, not a finding.

🧭 SubQ Snap Gives Small Decisions More Context
Friday brought a wave of decision models, which pick from fixed answers instead of writing text. Monday’s SubQ Snap makes those choices with up to 384K tokens of text and images; SubQ says today’s decision models “stay at 32K tokens.” Price: $0.05 per million input tokens, output free. It’s a limited public preview, so you may hit a waitlist.

Toolbox
🧰 Three Tools Worth a Look

🎙️ Atticus needs no API key and no extra bill: it runs on the Claude Pro or Max plan you already pay for, through Claude Code. Hold a shortcut, talk, and it moves windows, finds files and reads your screen. Free and open source (MIT). It asks before anything deletes, sends or buys.

🛠️ Agent37 began in February as an OpenClaw wrapper, its founder says. Now it hosts sandboxes for personal agents at what he calls “under 10% of the cost of most sandbox providers.” The free plan includes about 10 awake hours a week, and memory pressure can restart heavier sandboxes.

💻 Janas targets the usual pain of running a model bigger than your RAM. On the developer’s 32GB laptop, llama.cpp managed 2.8 tokens a second with an 80B, 48GB model; Janas reports 37 (24 without drafts). Not independently verified. GPL3, Linux and Windows x86-64.

Test Your Skill
Snack Quiz: Choose The Real Image 🔍

One is an archival photograph. The other is an independent AI recreation. Which handshake happened in a real laboratory?

Image A

Image B

Original photo credit: NASA. Reveal and source, photographed June 17, 2011 at Johnson Space Center. The companion scene was generated with AI.