AIOctober 13, 2026· 3 min read
This week in AI: Meta's first frontier model, Character.AI's bankruptcy and a warning from the IMF
Our round-up of the AI stories that mattered this week, from Meta's Muse Wonder launch to Microsoft's Maia 300 chip and growing doubts about the AI spending boom.
By Frogion Team
Every week the Frogion team reads through the AI news so you don't have to. Here is what stood out in the week of October 13, 2026.
Meta ships its first frontier model
Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12. Wonder is priced at $4 per million input tokens and $20 per million output tokens, while Wonder Full is reserved for Muse Code subscribers. On the Artificial Analysis Intelligence Index the model landed in fourth place with a score of 56. A solid debut, but not the leap some people expected from Meta's first frontier release.
Character.AI files for Chapter 11
Character Technologies, the company behind Character.AI, filed for bankruptcy protection in Delaware on October 12. Disney Worldwide Services is the stalking-horse bidder at $20 million, but for a non-exclusive license to certain IP rather than the app itself. Google, which paid $2.7 billion to license the company's technology in 2024, is the largest unsecured creditor.
For app studios it is a sobering reminder: a huge, engaged user base does not automatically turn into a business.
Microsoft unveils Maia 300
Microsoft's new Maia 300 accelerator carries 288 GB of HBM4 memory at 18 TB/s and is rated at 950 watts. Microsoft claims 40% more tokens per megawatt than leading GPU systems and plans multi-gigawatt deployments from 2027. Its predecessor, Maia 200, already runs Copilot.
Small models keep getting cheaper and more open
- Microsoft Decision-1, a classifier built on Qwen3.5-9B for routing, verification and workflow control, costs $0.042 per million input tokens with free output.
- Perplexity open-sourced pplx-embed-v2-late under the MIT license: a 0.6B edge model and a 9B model that share one embedding space, so an index built with the big model can be served by the small one.
- Liquid AI released the open-weight decision models d1-3B and d1-omni-600M, with d1-3B running in 8 ms on an RTX 4090.
The money question
A semiconductor selloff spread through Asia and Europe, with SK Hynix down more than 10% and Samsung down 7.4% in a single session. The IMF's October financial stability report then named a collapse in AI-related investment as a near-term risk, estimating that a steep drop could trim global GDP by 1.1% by 2028. Meanwhile, IDC data reported by Reuters put Nvidia's share of China's data-center AI chip market at 34% for the first half of 2026, with Huawei at 28%.
Also worth noting
- Google's Gemini coworker agent now gets its own Workspace identity, with an email address, calendar, Drive storage and a seat in the company directory. Private preview started October 8 for Gemini Enterprise customers.
- Butterfly Effect, the company behind the Manus agent, raised more than $500 million in a round led by Boyu Capital and IDG Capital.
- A UC Berkeley study of 1,222 participants found that ten minutes of ChatGPT use reduced people's persistence on a hard task once the AI was taken away.
- Anthropic's usage policy update, effective November 12, bars sustained abusive behaviour toward Claude and adds a section on election interference.
Compiled from public reporting. Figures are as reported by the sources below; check the originals before relying on them.
Sources
Cover photo by @steve_j on Unsplash.
