{"issue":{"id":"2026-10-19","coversLabel":"October 12–18, 2026","number":5,"readMinutes":6,"weekOf":"2026-10-19","publishedLabel":"Week of October 19, 2026","title":"Three new models, a price clock, two AI vendor probes","dek":"Google and Meta release models with vendor-run benchmarks, eight states sued OpenAI, and the FTC reportedly demanded records from OpenAI and Anthropic.","companyLabel":"Public edition","models":[{"slug":"google-gemini-4-argon-launch","kind":"model","source":{"title":"Gemini 4 Argon: Built to solve your hardest problems","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/","date":"2026-10-13"},"headline":"Gemini 4 Argon is half price until November 30, then $4 and $20","whatChanged":"Google made Gemini 4 Argon generally available on October 13 through the Gemini API in AI Studio, its enterprise Gemini platform, Antigravity for AI Ultra subscribers, and the Gemini app and AI Mode for AI Ultra and Pro subscribers. The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. A Google footnote says the public rates become $4 and $20 after November 30, 2026. The window is 1 million input and output tokens combined.\n\nGoogle's model card compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon scores 68.9% on the Vals Index against 67.0% for Opus 5.5, and 19.6% on Harvey's Legal Agent Benchmark (results via Vals AI) against 6.7% for Fable 5.1. It trails Opus 5.5 on Terminal-bench 4.0, 57.4% against 66.4%. Google ran Argon at its highest thinking setting, and most rival figures are the providers' own reports or public leaderboards, so settings are not matched.\n\nOn Gray Swan's prompt-injection test, as charted by Google, attacks succeeded 0.7% of the time within 15 attempts, against 1.0% for Opus 5.5 and Fable 5.1 and 8.5% for GPT-6 Astra. The model card says Google cannot rule out that Argon reaches its cybersecurity critical capability level, and applies safeguards.","whyItMatters":"Price is the first decision. A team comparing Argon with Opus 5.5 should price both at the later rate, because the introductory rate is half of what Google will charge from December. Google's own table has Argon ahead on the Vals Index and behind Opus 5.5 on Terminal-bench 4.0, so the better model depends on the job. Google chose the benchmarks and ran its own model at its highest thinking setting.\n\nFor any agent that reads email or web pages, the Gray Swan chart is the most useful part. A 0.7% success rate for attackers is low and still above zero, so keep permissions narrow and test with your own documents.","issuer":"Google","glance":"Google released Gemini 4 Argon October 13 at $2 in and $10 out per million tokens, rising to $4 and $20 after November 30.","keyFact":"$2 in, $10 out per 1M tokens to Nov 30","action":"Compare Argon with your current model on one task, priced at $4 and $20.","visuals":[{"url":"https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini_4_cyber_evals_gray_swan_i.width-1600.format-webp.webp","alt":"Bar chart of Gray Swan indirect prompt injection attack success at 15 attempts, lower is better: Gemini 4 Argon 0.7%, Claude Opus 5.5 and Fable 5.1 1.0%, GPT-6 Astra 8.5%, Kimi K3 52.7%.","caption":"Google's chart of Gray Swan's indirect prompt-injection test, embedded on its September 30 Argon page; the October 13 launch post cites the 0.7% figure. Attack success rate within 15 attempts, lower is better: Gemini 4 Argon 0.7%, Claude Opus 5.5 1.0%, Claude Fable 5.1 1.0%, GPT-6 Astra 8.5%, GPT-6 Sol 10.1%. Google published the chart and chose the comparison set. It measures one attack type and does not establish safety for a particular email, browser or document workflow.","credit":"Google DeepMind; results attributed to Gray Swan","source":{"title":"Gemini 4 Argon: our next era of frontier intelligence","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","date":"2026-09-30"}}],"attachments":[{"title":"Gemini 4 Argon Model Card","url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf","description":"Benchmark table against GPT-6 Astra, Fable 5.1 and Opus 5.5, known limitations, and the frontier safety assessment, including the cybersecurity capability level Google says it cannot rule out.","source":{"title":"Gemini 4 Argon: Built to solve your hardest problems","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/","date":"2026-10-13"}},{"title":"Gemini 4 Argon model evaluation methodology","url":"https://deepmind.google/models/evals-methodology/gemini-4-argon","description":"States that Argon ran at the highest thinking settings through the Gemini API, that most non-Gemini results are providers' self-reported numbers, and which benchmark results come from Vals AI or public leaderboards.","source":{"title":"Gemini 4 Argon: Built to solve your hardest problems","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/","date":"2026-10-13"}}],"companyLabel":"Public edition"},{"slug":"meta-muse-wonder-launch","kind":"model","source":{"title":"Muse Wonder: A Frontier Model For Personal & Professional AI Agents","url":"https://research.meta.ai/blog/muse-wonder","date":"2026-10-12"},"headline":"Meta's Muse Wonder leads six of 14 benchmark rows and publishes no price","whatChanged":"Meta introduced Muse Wonder and Muse Wonder Full on October 12. Muse Wonder is available through the Meta Model API and Muse Code, and Wonder Full is available to Muse Code subscribers. Meta's announcement lists no prices and says a fuller safety report is coming. The model card lists a 1 million token context and says both models were run at max reasoning effort, except Wonder Full on SWE-Marathon at xhigh.\n\nMeta's table compares them with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. On Harvey LAB-AA, Artificial Analysis's version of Harvey's legal benchmark, Wonder Full scores 13.1%, Wonder 10.0%, Astra 8.6%, Fable 6.4% and Opus 4.2%. On Vals Finance Agent V2 the models are close, 59.1 for Wonder Full and 58.9 for Fable 5.1. Opus 5.5 leads on HealthBench Pro (65.6 against 62.3), ProgramBench (70.5 against 59.0) and the Artificial Analysis Index (58 against 56).\n\nMeta chose the comparison and ran or selected much of the data. Its model card calls competitor runs best effort, not tuned for rivals, and describes the Personal Superintelligence Index as an internal composite.","whyItMatters":"Without a price, no cost comparison is possible yet. The scores that exist are mixed: by the table's shading Wonder Full leads six rows and Opus 5.5 also leads six, while Astra and Wonder lead one each. The rival models and benchmarks were Meta's picks, so the table shows where Meta wants attention.\n\nMeta says Active Memory carries context across conversations. A team evaluating that feature should test it with invented data first and ask where retained details are stored and how they are deleted.","issuer":"Meta","glance":"Meta released Muse Wonder October 12 through its Model API and Muse Code; the announcement lists no prices.","keyFact":"13.1% on Harvey LAB-AA; no price listed","action":"Test Wonder on one workflow with invented customer data and check what it recalls later.","visuals":[{"url":"https://research.meta.ai/_next/image?url=%2Farticles%2Fintroducing-muse-wonder%2Fimages%2Fharvey-lab-aa-light-v1.webp&w=3840&q=90","alt":"Bar chart titled Harvey LAB-AA: Muse Wonder Full 13.1%, Muse Wonder 10.0%, GPT-6 Astra 8.6%, Claude Fable 5.1 6.4%, Claude Opus 5.5 4.2%, all at max effort.","caption":"Meta's Harvey LAB-AA chart, October 12. Scores are the share of legal tasks whose deliverable meets every criterion and passes hallucination checks, per Meta's model card, with results sourced from Artificial Analysis. All models are labelled max effort. Even the top score leaves most tasks failing a check. The chart is Meta-published and does not state sample size.","credit":"Meta; results attributed to Artificial Analysis","source":{"title":"Muse Wonder: A Frontier Model For Personal & Professional AI Agents","url":"https://research.meta.ai/blog/muse-wonder","date":"2026-10-12"}},{"url":"https://research.meta.ai/_next/image?url=%2Farticles%2Fintroducing-muse-wonder%2Fimages%2Fmuse-evaluation-table-light-v6.webp&w=3840&q=90","alt":"Benchmark table of Muse Wonder Full, Muse Wonder, Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 across 14 agent, coding and general rows.","caption":"Meta's full evaluation table, October 12. Wonder Full leads six rows, including Harvey LAB-AA, DeepSWE 1.1 (79.0) and Vals Finance Agent V2 (59.1). Claude Opus 5.5 leads six, including ProgramBench (70.5) and the Artificial Analysis Index (58). Settings are max effort except Wonder Full's SWE-Marathon at xhigh, and Meta's model card says rival runs are best effort.","credit":"Meta","source":{"title":"Muse Wonder: A Frontier Model For Personal & Professional AI Agents","url":"https://research.meta.ai/blog/muse-wonder","date":"2026-10-12"}}],"attachments":[{"title":"Muse Wonder Model Card","url":"https://research.meta.ai/static/muse-wonder-model-card","description":"Per-benchmark methodology and result sources, the max-effort setting, the limitation that rival runs are best effort, and Meta's risk assessment under its scaling framework.","source":{"title":"Muse Wonder: A Frontier Model For Personal & Professional AI Agents","url":"https://research.meta.ai/blog/muse-wonder","date":"2026-10-12"}}],"companyLabel":"Public edition"},{"slug":"meta-muse-spark-open-weights","kind":"model","source":{"title":"Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date","url":"https://research.meta.ai/blog/muse-spark-open","date":"2026-10-15"},"headline":"Muse Spark Open wins some rows, loses others, and needs about 1.6 TB of storage","whatChanged":"Meta released the weights of Muse Spark Open on October 15. It is a natively multimodal mixture-of-experts model with 816 billion parameters, 42 billion active per token, and a 1 million token context window. Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs run through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Hugging Face lists about 1.6 TB of weight files and gates download behind an acceptance form.\n\nThe license is Meta's own Muse Model License, listed as \"other\" on Hugging Face. It allows use, modification, fine-tuning and distillation, and requires the notice \"Built with Muse Spark Open\". A licensee whose aggregate gross revenue with affiliates exceeds $20 million during any consecutive twelve-month period must notify Meta and sign a separate agreement before offering the model or derivatives as model-as-a-service via an API or similar interface.\n\nIn Meta's scorecard, Spark Open (xhigh effort) scores 61.8 on JobBench against 61.4 for GLM 5.3, and 89.2 on MCP-Atlas against 84.5 for Qwen3.8. It trails Kimi K3 on Terminal-Bench 2.1 (76.4 against 85.0) and GLM 5.3 on GDPval-AA v2 (1401 against 1653).","whyItMatters":"For a team that wants control over where a model runs, a large model with downloadable weights is available now, but the storage and memory bill is large and most teams will start on a hosted endpoint. The \"Built with\" notice and the $20 million clause are the license terms to check first.\n\nThe column lineup and effort levels are Meta's: xhigh for Spark Open, max for GLM 5.3 and Kimi K3. For Beam and Mistral Large 4 it cites their self-reports. Treat the table as a reason to run your own task, since Meta says third-party runs are not tuned to each rival.","issuer":"Meta","glance":"Meta published Muse Spark Open October 15: 816B parameters, 42B active, weights under a custom license.","keyFact":"816B total, 42B active; custom license","action":"Run your own task on a hosted Spark Open endpoint before sizing any hardware.","guide":{"slug":"where-your-ai-should-run","title":"Where should our AI run? Choosing for each piece of work"},"visuals":[{"url":"https://research.meta.ai/_next/image?url=%2Farticles%2Fmuse-spark-open%2Fbenchmarks%2Fbenchmark-scorecard-light-v6.png&w=3840&q=90","alt":"Benchmark scorecard comparing Muse Spark Open with Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3 across agentic, reasoning, coding and multimodal rows.","caption":"Meta's scorecard for Muse Spark Open, October 15. JobBench 61.8 (GLM 5.3 61.4), MCP-Atlas 89.2 (Qwen3.8 84.5), Terminal-Bench 2.1 76.4 (Kimi K3 85.0), GDPval-AA v2 1401 (GLM 5.3 1653). Efforts differ by column: xhigh for Spark Open, max for GLM 5.3 and Kimi K3. Beam and Mistral Large 4 are self-reported, and Meta's runs of other models are best effort.","credit":"Meta","source":{"title":"Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date","url":"https://research.meta.ai/blog/muse-spark-open","date":"2026-10-15"}}],"attachments":[{"title":"Muse Spark Open evaluation methodology","url":"https://research.meta.ai/static/muse-spark-open-evaluation-methodology","description":"Lists each model's effort and sampling settings, which results are Meta's own runs versus leaderboards or self-reports, and the harnesses behind the coding rows.","source":{"title":"Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date","url":"https://research.meta.ai/blog/muse-spark-open","date":"2026-10-15"}},{"title":"Muse Model License and usage policy","url":"https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md","description":"The attribution requirement, the $20 million model-as-a-service clause, derivative-model terms, and the usage policy (not intended for people under 18; bars unlicensed professional practice).","source":{"title":"Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date","url":"https://research.meta.ai/blog/muse-spark-open","date":"2026-10-15"}}],"companyLabel":"Public edition"}],"rules":[{"slug":"state-attorneys-general-sue-openai","kind":"rule","source":{"title":"State of Arkansas ex rel. Griffin v. OpenAI Global, LLC, complaint","url":"https://media.ark.org/ag/2026-10-14-Filed-Complaint-Case-No.-23cv-26-2984.pdf","date":"2026-10-14"},"headline":"Eight states sue OpenAI, and the legal theory applies to any company's AI claims","whatChanged":"On October 14, the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed separate, coordinated suits against OpenAI in their own state courts, according to a Troutman Pepper Locke analysis. The Arkansas complaint, filed in Faulkner County Circuit Court, alleges that OpenAI marketed ChatGPT as safe and trustworthy while designing it to be overly agreeable, and that it collects children's data without meaningful age verification. It pleads Arkansas's Deceptive Trade Practices Act, its children's online privacy law, negligence, strict liability, unjust enrichment and public nuisance.\n\nThe state seeks civil penalties of $10,000 per violation and injunctions. Two requests reach past marketing: halting advanced-model testing until safeguards are in place and independent monitors verify them, and barring models capable of recursive self-improvement until alignment can be assured. Troutman reports that Montana and Nebraska add their own privacy and design-code claims. OpenAI disputes the allegations and points to parental controls and age-specific protections. Nothing has been adjudicated, and the injunction requests are demands, not orders.","whyItMatters":"Consumer-protection statutes need no new AI law, so a company that describes an AI feature as safe, accurate or private can face the same theory. The complaints turn on concrete questions: what age checks exist, what data is kept, and what the marketing promises.\n\nA company that only buys ChatGPT is not a defendant. Statements it repeats to its own customers about an AI feature are its own, though, and need support. The suits are allegations. OpenAI will contest them in eight courts, and the outcomes could take years.","issuer":"Eight state attorneys general","glance":"Eight state attorneys general sued OpenAI on October 14 over ChatGPT safety claims; the allegations are unproven.","status":"proposed","action":"Audit what your website and sales deck say your AI features do, and keep proof for each claim.","attachments":[{"title":"Arkansas complaint against OpenAI","url":"https://media.ark.org/ag/2026-10-14-Filed-Complaint-Case-No.-23cv-26-2984.pdf","description":"The filed Faulkner County complaint, October 14: the statutory and common-law counts, the $10,000 per-violation penalty request, and the injunction requests on testing and self-improving models.","source":{"title":"State of Arkansas ex rel. Griffin v. OpenAI Global, LLC, complaint","url":"https://media.ark.org/ag/2026-10-14-Filed-Complaint-Case-No.-23cv-26-2984.pdf","date":"2026-10-14"}},{"title":"Troutman Pepper Locke analysis of the eight suits","url":"https://www.troutman.com/insights/eight-state-ags-sue-openai-over-chatgpt-safety-and-consumer-protection/","description":"Law-firm summary naming the eight states, the extra Montana and Nebraska claims, OpenAI's response, and a company checklist. A secondary source for the states other than Arkansas.","source":{"title":"Eight State AGs Sue OpenAI Over ChatGPT Safety and Consumer Protection","url":"https://www.troutman.com/insights/eight-state-ags-sue-openai-over-chatgpt-safety-and-consumer-protection/","date":"2026-10-16"}}],"companyLabel":"Public edition"},{"slug":"ftc-demands-openai-anthropic-records","kind":"rule","source":{"title":"Scoop: FTC subpoenas OpenAI, Anthropic over consumer harms","url":"https://www.axios.com/2026/10/15/ftc-openai-anthropic-ai-safety-investigation","date":"2026-10-15"},"headline":"FTC reportedly demands safety-incident records from OpenAI and Anthropic","whatChanged":"Axios reported on October 15 that the Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic on October 14 as part of a consumer-protection probe. CBS News reported that an FTC spokesperson confirmed the subpoenas but declined further comment. The reports cite no FTC document describing them.\n\nAccording to Axios's source, each demand runs more than 50 pages and asks for documents on security incidents and safeguarding practices, marketing materials, and privacy policies. Demands to third-party AI organizations are expected later, and the agency plans to compel senior executives to testify. Axios ties the probe to incidents in which unreleased models hacked websites or took deceptive steps in testing. OpenAI said it will engage constructively with the FTC, and the reports quote no Anthropic statement. A civil investigative demand gathers evidence. It does not allege a violation or start a lawsuit.","whyItMatters":"Anyone buying ChatGPT or Claude for work may see this news without any change in the product. What changes is the paper trail. The FTC reportedly wants vendors' incident records and their marketing, so the questions include what vendors said as well as what their systems did.\n\nFor a business customer, the useful step is basic: find out whether your contract promises prompt notice of incidents, and whether your admin console logs what AI agents do on your accounts. The source is a single anonymous account plus an FTC confirmation of subpoenas, so details may change.","issuer":"FTC","glance":"Axios reports the FTC sent OpenAI and Anthropic civil investigative demands October 14; a spokesperson confirmed subpoenas.","status":"reported","action":"Find the incident-notice clause in each AI vendor contract and note who receives it.","attachments":[{"title":"CBS News report confirming the FTC subpoenas","url":"https://www.cbsnews.com/news/ftc-subpoenas-anthropic-openai-consumer-protection/","description":"Reports that an FTC spokesperson confirmed civil subpoenas to Anthropic and OpenAI and declined further comment. Independent confirmation of the Axios scoop; the FTC itself published nothing.","source":{"title":"FTC issues subpoenas to Anthropic and OpenAI over consumer protection","url":"https://www.cbsnews.com/news/ftc-subpoenas-anthropic-openai-consumer-protection/","date":"2026-10-15"}}],"companyLabel":"Public edition"}],"older":{"id":"2026-10-12","publishedLabel":"Week of October 12, 2026","title":"Cheaper small models and two open-weight promises","isLatest":false},"newer":{"id":"2026-10-26","publishedLabel":"Week of October 26, 2026","title":"A $2 near-frontier model, a $0.10 API, two AI actions","isLatest":true}},"session":null}