Meta releases Muse Wonder, its first frontier model

On 12 October, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. Meta calls them its first frontier models. They're bigger than the low-latency Muse Spark. Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is only available to Muse Code subscribers. Meta's model card puts both models level with Fable 5.1 and GPT-6 Astra, and says Muse Wonder beats them on MemoryBench, a benchmark for memory and continual learning. Meta's pitch centres on "active memory", meaning the model keeps and combines what it learns about a user over time. Meta trained the models at its Prometheus data center in Ohio.

For builders: The price is in the same range as Opus 5.5, so it's a realistic option for comparison testing. Don't plan on self-hosting, though, because Meta has only described access through the API and subscriptions.

Gemini 4 Argon reaches AI Pro subscribers

On 13 October, Google began rolling Gemini 4 Argon out to Google AI Pro and AI Ultra subscribers in the Gemini app and in Search's AI Mode. Google announced Argon on 30 September and first offered it to cyber defenders in its Fairwind Program. Google says Argon scores 77.9% on DeepSWE v1.1 and raises the output limit from 64K tokens to 1M. The Antigravity app and CLI, AI Studio and Android Studio Canary still require Ultra. API pricing is $2 input and $10 output per million tokens until 30 November 2026, then rises to $4 and $20.

For builders: If you want to benchmark Argon on long coding tasks, run those tests before 30 November, while the API costs half the standard rate.

Anthropic says its distillation defenses cut harvesting by up to 99%

On 12 October, Anthropic described new protections against distillation, where other labs harvest Claude's outputs and reasoning to train their own models. With Fable 5.1 and Opus 5.5, API accounts created on or after 31 August can't edit Claude's earlier context. A two-stage classifier with a false positive rate under 0.1% also screens requests. Anthropic's early data shows harvested data down by about 90 to 99%. It names Zhipu and Moonshot as cases it has seen, and it expects attackers to adapt.

For builders: Preserved thinking may break integrations that rewrite earlier turns of a conversation. Since 24 September, blocked requests have been billed, so check your error handling.

Anthropic red team tests self-replication by small models

Anthropic's Frontier Red Team reported on 12 October that small open models are getting close to copying themselves onto cheap hardware. A 9B Qwen model got inference running on 8 of 9 hardware types. In a simplified 8-node test, only GLM-5.2 replicated end to end. It reached all 8 nodes in 4 of 5 trials. The exploit used in the test was deliberately trivial, so the authors treat the results as upper bounds. Their suggested defenses are traffic monitoring, honeypots, quarantining infected subscribers and secure default credentials.

For builders: If you ship edge or IoT devices that can run local models, close unauthenticated ADB and remove default passwords now.

Character.AI files for Chapter 11

Character.AI filed for bankruptcy on 13 October. Its filing puts assets and liabilities each between $50 million and $100 million. Disney has agreed to buy some assets for $20 million and gets a non-exclusive, perpetual licence to the company's technology. CEO Colin Adams said that if no bid keeps the service running, "the platform will be wound down". The sale should close in late November or early December. The company has faced lawsuits linked to a teenager's death and an investigation in Texas.

For builders: If you build on a third-party companion platform, plan for it shutting down. Keep your data exportable.

SpaceXSI bundles Grok, Cursor and X Premium

On 13 October, SpaceX Super Intelligence launched "The X Subscription". It has five tiers: Free, Plus ($8 a month), Premium ($30), Super ($100) and Ultra ($200). The latest Grok models need Premium or higher, and usage caps go up with each tier. The upgrade option is rolling out gradually.

For builders: Teams already paying for Cursor should compare their current costs against the bundle tiers.

Amex proposes an intent standard for agent purchases

American Express announced its Intent Intelligence Framework on 13 October as part of its ACE Developer Toolkit. A Create Intent API records what the customer allows, such as the merchant and a maximum budget. The customer approves it, and the system issues an Intent ID and a Proof of Intent Token. Amex can then check each purchase against what the customer approved.

For builders: If you're building shopping agents, capture what the user has authorised as structured constraints and keep audit logs of each approval.

IMF credits AI investment as it lifts 2026 growth forecast

The IMF's October World Economic Outlook, released on 13 October, forecasts global growth of 3.1% for 2026. That's up 0.1 point from July. The 2027 forecast is 3.4%. Managing Director Kristalina Georgieva said spending on AI infrastructure is a key reason growth is holding near 3%.

For builders: Compute budgets and funding currently depend on continued AI capital spending, so watch for signs that it's slowing.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS