Today's five
Meta shipped Muse Wonder. Its first frontier-class model, at $4 input and $20 output per million tokens. No parameter count, no benchmark numbers published.
Microsoft unveiled Maia 300. 288 GB of HBM4, 18 TB/s, and a claim of 40% more tokens per megawatt than leading GPU systems.
Apple's Welcome Home event is today. A smart home hub, a new Apple TV and HomePod mini 2 are expected. Nothing is confirmed yet.
A cryptomining botnet is hunting exposed AI servers. LiteLLM and Ollama instances on the open internet. Check yours.
Anthropic's investor day is this week, with reporting pointing at a $2 trillion IPO. None of it is confirmed by Anthropic.
Meta released its first frontier model
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on Monday 12 October. Internally the project was code-named Watermelon.
Alexandr Wang, who heads the lab, called them "true frontier models, competitive with the best models in the world." Meta's own blog went further: "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence."
Pricing is the hard fact in the announcement. Muse Wonder costs $4.00 per million input tokens and $20.00 per million output through the Meta Model API. Muse Wonder Full is not on the API at all — it is restricted to Muse Code subscribers.
That price is double the incumbents. GPT-6.1 Sol and Claude Sonnet 5.5 both run at $2 input and $10 output. Meta, arriving late to the frontier, has priced its first entry at twice the rate of the models it says it matches — while publishing no parameter count, no context window and no numeric benchmark scores.
Two capabilities are named. Active Memory retains information across conversations and retrieves it when relevant. Active Focus keeps an agent working toward a goal, trying different approaches when it hits an obstacle.
Meta's model card says Wonder scores comparably to Fable 5.1 and GPT-6 Astra, and higher than Fable, Opus and Astra on MemoryBench, which tests continual learning. No figures accompany those claims. The models were trained at Meta's Prometheus data centre in central Ohio.
Full breakdown, with the price comparison in a table: Meta Muse Wonder — $4/$20 and no benchmark numbers.
Microsoft's Maia 300 goes after tokens per watt
Microsoft unveiled its third-generation custom AI accelerator on Monday. The Maia 300 is built on TSMC's N2P node with 288 GB of HBM4 and 18 TB/s of bandwidth.
The claim that matters is efficiency rather than raw speed: Microsoft projects 40% more tokens per megawatt than leading GPU systems. A rack holds 96 accelerators in roughly a 167 kW envelope, with multi-gigawatt deployments starting in 2027.
Microsoft expects capital spending above $175 billion this year. Read the efficiency framing in that context — when you are spending at that scale, a 40% improvement in tokens per megawatt is a cost argument, not a performance one. Electricity, not silicon, is becoming the binding constraint.
Apple's Welcome Home event is today
Apple holds its "Welcome Home" product launch today, 13 October. Rather than a keynote, Apple is running hands-on sessions for media in New York.
Expected: a smart home hub reported as HomeView, a new Apple TV 4K, and a second-generation HomePod mini, which is reported to reach customers on 30 October.
None of this is confirmed. Every detail above comes from pre-event reporting, and Apple has announced only the event itself. The question worth watching is how much of Siri's rebuilt AI stack appears in the home — that is the part with consequences beyond Apple's hardware line.
A botnet is hunting exposed AI servers
Lumen's Black Lotus Labs published research on a cryptomining campaign it tracks as Canto Incognito, spreading malware named PoeLLM. Most victims were running LiteLLM or Ollama exposed to the open internet.
The operator scans for services on port 4000 (LiteLLM) and port 3000 (Gotenberg), then sends a crafted POST request that pulls down a payload. The likely LiteLLM path is a command injection flaw, CVE-2026-42271.
Why AI servers specifically. They sit on GPU hardware, which is exactly what a cryptominer wants, and they often hold API keys and user data. A self-hosted model server on a public IP is a high-value target with a published attack surface.
The command-and-control address is hidden in a poem hosted on GitHub. The malware pulls four words from fixed positions, maps them through a dictionary to four numbers, and assembles an IP address. The operator changed the poem eleven times between April and September — moving every victim to a new server without touching the malware.
If you self-host anything: PoeLLM — check your Ollama and LiteLLM now.
Anthropic's investor day, and a number nobody has confirmed
Anthropic holds an investor day this week, against a backdrop of reporting that it may seek a valuation near $2 trillion in what would be the largest IPO ever. Fortune reported an October timeline back in August.
Treat all of it as reporting rather than fact. Anthropic has confirmed no valuation, no date and no filing. And the number has sceptics: an analyst quoted by 24/7 Wall St on 8 October argued the company could be overvalued by as much as 93% at that price, calling it an "unprecedented test of investor gullibility."
Separately and more concretely, HCLTech announced a Global Premier Partnership with Anthropic covering Claude, Cowork and Claude Code, including a dedicated business unit and a centre of excellence for cybersecurity, governance and software engineering. Financial terms were not disclosed.
The US cut Britain out of frontier model testing for 17 days
Britain's AI Security Institute announced on 12 October that it has regained access to frontier AI models, ending a 17-day period in which that access was cut off.
The freeze began around 24 September and lifted on 11 October. Reporting attributes it to a request from the US Office of the National Cyber Director, which wanted domestic American safety reviews to come first.
Why it happened so easily. AISI's access runs on voluntary agreements dating from 2023. No developer is obliged to share anything. When Washington asked for American reviews to go first, there was no mechanism to appeal to — the arrangement simply bent. That is worth sitting with: the body testing frontier models for the UK can be switched off by a request, not a ruling.
During the freeze, Anthropic withheld Claude Mythos 5.1 from non-US entities. Google's Gemini 4 Argon is now under AISI review following the restoration. AISI had previously evaluated OpenAI's GPT-6 Astra.
UK AI Minister Kanishka Narayan is credited with securing the renewed access. He is separately expected to set out plans this week to strengthen British cyber defences and reduce dependence on foreign-controlled infrastructure — a programme that reads differently now that the last three weeks have demonstrated what that dependence looks like in practice.
Also today
- The IMF put a number on AI and jobs. Kristalina Georgieva said AI could affect around 40% of jobs worldwide — roughly 60% in advanced economies and 26% in low-income countries. Read those as exposure, not predicted job losses. They measure how much of a role AI could touch, not how many roles disappear.
- IonQ is acquiring Elisity in an all-stock deal worth up to $1.3 billion, including up to $150 million in performance-based payments, per the Wall Street Journal.
- Amazon and Decart are still only in talks. No deal has been signed since the WSJ report on 11 October. Our breakdown covers why Anthropic walked away in September.
- Google is defending a $15 billion data centre in Visakhapatnam, India against local concerns about power, water and land. It says the site will use air cooling and should be operational in 2028.
What to do with today
- If you self-host models, this is the only item here that needs action today. Check whether anything is listening on ports 3000 or 4000 from outside your network.
- If you are costing out an API, Muse Wonder at $4/$20 is twice Sol and Sonnet. Wait for independent benchmarks before paying the premium.
- If you are watching infrastructure, Maia 300 and the Google India fight are the same story: AI's constraint is moving from chips to power.
- If you read anything about a $2 trillion Anthropic, check whether the source is quoting Anthropic or quoting another outlet. So far it is all the latter.