Gemini 4 Argon arrives and Meta ships Muse Wonder frontier models

Google prices Gemini 4 Argon low until 30 November

Developers can start building with Argon at US$2 per million input tokens and US$10 per million output tokens. Cached input costs 95% less. On 1 December those rates double to US$4 and US$20 for the public model. Google released Argon on Tuesday 13 October and is making it available in every region through Google AI Studio and the Gemini API. One request can hold up to 1 million tokens of input and output combined. Pro and Ultra subscribers can use Argon in the Gemini app and in Search's AI Mode. Ultra subscribers also get it inside Google Antigravity.

Google's results put Argon top of the Vals Index, a test of work in tax, legal, coding and finance. On Gray Swan's indirect prompt injection test, attackers got through 0.7% of the time over 15 adaptive tries. Google is also adding hundreds of partners to its Fairwind Program. Security defenders who pass vetting can apply to use Argon with its cyber guardrails removed.

What it means for NZ organisations: Teams already on Google Cloud can trial Argon cheaply on their own documents before December. Build your cost model on the US$4 and US$20 rates, because that's what you'll pay in production.

Meta's Superintelligence Lab releases Muse Wonder

On Monday 12 October Meta released Muse Wonder and a bigger version called Muse Wonder Full. They're its first frontier-class models since Alexandr Wang reorganised Meta's AI work in August 2025, and they're more capable than the small Muse Spark. Muse Wonder costs US$4 per million input tokens and US$20 per million output tokens through the Meta Model API. You can only get Wonder Full with a Muse Code subscription. Both models will run Meta's Muse agent, which handles long tasks.

Meta's model card says the two models score about level with Fable 5.1 and GPT-6 Astra. It also says Muse Wonder leads on MemoryBench, a test of how well a model keeps and combines what users tell it. Meta ran all these tests itself. It trained both models at its Prometheus data centre in central Ohio.

What it means for NZ organisations: Meta now sells frontier AI by the token, so it's worth adding to vendor shortlists. Test the memory features on your own data, and check where that stored information sits and how long it's kept.

OpenAI engineer charged over a recorder found in an executive room

US prosecutors announced on 13 October that Abdulrahman Alfozan, 35, had been charged with using or trying to use a device to intercept oral communications. He's a Saudi national who's worked as an OpenAI engineer since July 2025. He was arrested on 9 October and made his first federal court appearance on 13 October. If he's convicted, he faces up to five years in prison, a fine of up to US$250,000 and three years of supervised release.

The complaint alleges an encrypted recorder was wrapped in toilet paper and plugged into a USB port on the back of a television in an executive meeting room in San Francisco on 28 September. Staff found it on 30 September. On 5 October, security staff opened a bag he'd left in another room and found a similar second device next to his wallet and ID. Prosecutors also point to chatbot prompts about building recorders, sent from an account linked to his spouse. OpenAI says it fired him, reported the case to the FBI and used its own models to work out where the device came from. It says it has no sign that he accessed or took model weights. None of this has been proven in court.

What it means for NZ organisations: Physical access to boardrooms is an insider risk too. Check that badge logs, CCTV and room sweeps match how sensitive your executive meetings are.

Character.AI seeks a court-supervised sale, with Disney bidding first

Character.AI's parent company announced a voluntary Chapter 11 filing in Delaware on 12 October so it can sell all or nearly all of its assets. Disney has offered US$20 million as the stalking-horse bidder, which sets the minimum other bids have to beat. The offer covers a perpetual, non-exclusive licence to some of the company's IP and legal claims. It also includes transition services from former Character.AI staff that Disney has hired. Users can keep using the service while the sale runs, and the company expects it to close in late November or early December.

What it means for NZ organisations: Even a well-known AI startup can end up in a fire sale. Contracts with smaller AI suppliers should say what happens to your data and service if the supplier is sold or wound up.

Jack Clark now expects AI that builds its successors by 2027

Jack Clark, an Anthropic co-founder, said in Import AI 476 on 12 October that he now thinks full recursive self-improvement is more likely in 2027. He'd previously expected 2028. He pointed to TasteVal, a P-Zero Research benchmark that measures how well models pick experiments and read their results. Claude Opus 5.5 topped it and beat a baseline of 24 human experts, at about a thirtieth of the experts' average cost per attempt. The same issue said Western open-weight models such as Reflection's Beam and Mistral Large 4 trail Chinese models on several coding tests.

What it means for NZ organisations: Faster progress means AI plans need reviewing more than once a year. If you're choosing open-weight models, test Chinese and Western options on your own tasks and weigh data sovereignty alongside the scores.