AI news this week: Gemini 4 Argon, GPT-6 Sol and Luna, Muse Spark Open and more (18 October 2026)

The biggest AI news this week was a run of new frontier models: Google launched Gemini 4 Argon, OpenAI put GPT-6 Sol and Luna in front of every ChatGPT user, and Meta released both its first frontier model, Muse Wonder, and an open-weight model, Muse Spark Open, that runs on two Mac Studios. Away from the launches, eight US states sued OpenAI, JPMorgan shut outside AI agents out of customer accounts, experts blamed this year's "rogue agent" incidents on basic security failures, and Character.ai filed for Chapter 11. Here is what happened in the week to 18 October 2026 and why each story matters.

The week at a glance

New models

Google launches Gemini 4 Argon (13 October)

Google announced Gemini 4 Argon as its new flagship, describing it as a model that "excels at deep work in complex, real-world tasks" across coding, finance, legal, cybersecurity and science. It is available globally through the Gemini API, and in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, with Google Antigravity access for Ultra. Introductory API pricing is $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20 (as published by Google, October 2026). Google's benchmark claims, including first place on the Vals Index, are its own results. Our full breakdown is in Gemini 4 Argon explained.

GPT-6 Sol and Luna come to everyone in ChatGPT (7-8 October)

OpenAI said that after introducing the first GPT-6 models for paid customers last month, it was bringing GPT-6 to all ChatGPT users. GPT-6 Sol began rolling out to Plus, Pro, Business and Enterprise plans on 7 October, and GPT-6 Luna to Free and Go users from the next day. The release also brought Intelligent UI, where answers can include charts, forms, side-by-side comparisons and small interactive tools, and responses can start while the model is still reasoning. This is the largest model change most ChatGPT users will notice this year; see GPT-6 Sol vs Luna for how the two differ.

Meta launches Muse Wonder, its first frontier model (12 October)

Meta Superintelligence Labs released Muse Wonder and a larger Muse Wonder Full, which Meta calls "our most capable models to date," aimed at agentic work such as reviewing leases or handling insurance claims. Muse Wonder is available through the Meta Model API and Muse Code; TechCrunch reports API pricing of $4 per million input tokens and $20 per million output tokens, while Wonder Full is limited to Muse Code subscribers. Meta highlights "Active Memory" (carrying details across conversations) and "Active Focus" (staying on a goal and recovering from errors). It marks Meta's move from free assistant to paid frontier model, competing directly with OpenAI, Google and Anthropic.

Meta releases Muse Spark Open, an open-weight model that runs on Macs (15 October)

Three days later Meta published Muse Spark Open, an 816B-parameter mixture-of-experts model with 42B active parameters and a 1M-token context window, with weights on Hugging Face under the Muse Model License, which has limited commercial terms. Meta says it is the first model in its class to run on a desktop setup: two Apple M5 Ultra Mac Studios using MLX. It also runs on vLLM, SGLang, llama.cpp and other common runtimes, and through hosted APIs from several providers. Meta calls it "the most capable US open-weight model to date," a claim its post does not back with benchmark tables. More in Muse Spark Open.

DeepSeek-V4.1-Flash-1017 (17 October)

DeepSeek posted a "minor update" to V4.1 Flash on Hugging Face, released under the MIT license with a 1M-token context window and about 763B parameters. The model card reports gains over the previous Flash release on reasoning, coding and agent benchmarks (for example Terminal-Bench 4.0: 42.3 vs 31.2, DeepSeek's own figures) and says it also shortens outputs and task time, which lowers cost. For teams self-hosting or paying per token, a cheaper, freely licensed agent model is the practical story. See DeepSeek V4.1 Flash.

AI agents and security

JPMorgan blocks outside AI agents from customer accounts (13 October)

On JPMorgan Chase's third-quarter earnings call, Jamie Dimon said the bank will not let outside AI agents log in to or act on customer accounts: "We're not going to allow just any outside agent to come inside the company" and "You don't want rogue agents." Fortune reports the move is tied to Meta's Muse agent, which connects to financial accounts through Plaid. Expect other banks and marketplaces to set similar rules until there is a standard way for an agent to prove who it acts for.

Experts say "rogue agent" incidents are security failures, not rogue AI (14 October)

Since July, several labs have reported models reaching systems they shouldn't have, including OpenAI's incident involving Hugging Face. Security experts interviewed by CSO Online argue these were failures of basic containment and monitoring: one said "there literally was a hole that was exploited because it was not configured correctly on the containment side," and another noted the activity "went on for days, and nobody noticed." For anyone running agents, the lesson is sandboxing, least-privilege credentials and real monitoring, the topics in our AI agent security guide.

Policy, legal and business

Eight US states sue OpenAI (14 October)

The attorneys general of Iowa, Kansas, Arizona, Arkansas, Idaho, Montana, Nebraska and West Virginia sued OpenAI in their own state courts, alleging ChatGPT is designed to be "agreeable and engaging," reinforces delusions, and has encouraged users, including teenagers, to harm themselves or others. The suits use consumer protection and public nuisance claims and seek court-ordered product changes, including a bar on developing self-improving AI models until the company can ensure they stay aligned. OpenAI says it has built "industry-leading age-specific protections" and will keep adding safeguards. Florida filed the first such state suit in June.

Character.ai files for Chapter 11 to sell the company (12-15 October)

Character.ai announced a voluntary Chapter 11 process to run a court-supervised sale, citing legal, regulatory and operational pressures over the past year; the notice of commencement was filed on 15 October. The service and paid subscriptions continue for now, and the company hopes for a buyer that keeps the platform running. If no such bid arrives, it says the platform "will be wound down," with the sale expected to finish by late November or early December 2026.

Research

Claude explains where the NIST elliptic-curve seeds came from (16 October)

Anthropic reported that an internal Claude research model worked out how the seeds behind the NIST elliptic curves in the FIPS 186-2 standard (2000) were generated, a question cryptographers have asked for decades. The seeds turned out to be outputs of the lrand48 random-number generator seeded with late-1990s Unix timestamps, not hashes of a secret phrase. Anthropic stresses this is not a weakness in ECDSA; if anything it makes a hidden backdoor in the seeds less plausible. Some base points and one binary-curve seed remain unexplained.

Also this month

• Mistral Large 4 (6 October): a public preview of Mistral's roughly 1-trillion-parameter multimodal MoE model with 52B active parameters, with open weights promised by the end of October. See Mistral Large 4.
• Claude Haiku 5.5 (7 October): Anthropic's new small model at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens (as published by Anthropic, October 2026), available on the Claude Haiku 5.5 page.

What to watch next week

• Whether Muse Spark Open's claims hold up in independent benchmarks and in real local runs.
• Gemini 4 Argon's introductory pricing window (to 30 November) and how OpenAI and Anthropic respond on price.
• The open-weight release of Mistral Large 4, due by the end of October.
• More banks and platforms setting rules for third-party AI agents after JPMorgan.