Weekly AI Review — October 9, 2026

🚀 Model & Product Releases

• Anthropic ships Claude Haiku 5.5 with tiered pricing – Released on October 7 on the Claude Platform, AWS, Google Cloud and Microsoft Azure, the small model costs $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, rising to $0.50 and $2.50 above that threshold. Anthropic reports 39.2 percent on Terminal-Bench 4.0 against 16.4 for GPT-6 Luna, a figure no third party has reproduced. It follows Sonnet 5.5, released on September 28, which Anthropic says runs 30 percent faster and costs up to 30 percent less.
• Mistral previews Large 4, open weights promised this month – Announced on October 6, the 1-trillion-parameter multimodal model with 52 billion active parameters is in public preview on Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens, with weights due at the end of October. Mistral reports 61.7 percent on DeepSWE v1.1 and 28.3 percent on Terminal-Bench 4, vendor figures not yet independently run.
• OpenAI’s DevDay brings GPT-6.1 Sol and always-on agents – At its September 29 developer conference OpenAI released GPT-6.1 Sol, which it says nearly matches GPT-6 Astra on agentic coding and computer use at one-fifth of Astra’s token prices, available in the API, ChatGPT Work and Codex. It also began rolling out Dots, always-on agents on GPT-6 Astra with their own cloud computers and access to more than 4,000 apps, to Pro and Business Premium subscribers at one Dot per user, and added a $500-a-month Pro 500 plan with an Ultrafast speed tier.
• Google makes Nano Banana 2.1 image model generally available – Google’s Gemini API changelog lists gemini-nano-banana-2.1 as generally available on October 6, with improved prompt adherence, multi-turn character consistency and text rendering, and output at 1K, 2K and 4K. Google says it keeps Flash-level speed and cost and deprecates gemini-3.1-flash-image without a shutdown date.

🔬 Research Highlights

• Kandinsky 6.0 Video generates picture and sound together – Kandinsky Lab’s report, posted to arXiv on October 4, describes 3-billion and 29-billion-parameter models that produce 5-second clips with synchronised 44 kHz audio, including lip-sync. A dual-stream CrossDiT joins a pretrained video stream and a new audio stream through bidirectional cross-attention. The authors note that native output is standard definition and that trajectory compression alone does not fully preserve quality during distillation; code and weights are MIT-licensed.
• AgentGarten trains agents in code-defined worlds – Researchers from MirroS, Tsinghua and Peking universities report on October 8 that agents in their five hand-built environments learned behaviours such as shelter construction in four rounds, against roughly 25 million episodes for a reinforcement-learning baseline. Simulators and game engines share a neural renderer that adapts a pretrained video model to geometry conditions. The authors acknowledge that geometry alone leaves parts of the world state unspecified and that the worlds were built one at a time.
• Zhejiang team reconstructs Meta’s Muse agent as open source – nanoMuse, posted on October 6 under GPL-3.0, is a personal agent that runs across a person’s devices and shares one conversation through a relay. The authors rebuilt the design from Meta’s public Muse documentation and a circulating production prompt, labelling each statement as documented, observed or inferred. They state plainly that they have no success rate or count of human take-overs to report and that the gatekeeping Sentinel is a policy boundary, not a privilege boundary.

🏗️ Infrastructure & Compute

• Micron posts record quarter and guides higher on AI memory – The company’s September 30 release reports fiscal fourth-quarter revenue of $54.23 billion, up from $11.32 billion a year earlier, and GAAP net income of $37.70 billion. Its Core Data Center unit generated $18.0 billion at a 90 percent gross margin, and full-year revenue reached $133.19 billion. Micron guides fiscal first-quarter revenue to $61.5 billion plus or minus $1.5 billion.
• Google signs 20-year nuclear supply deal with Constellation – Announced on October 6, the agreement covers 890 megawatts of new capacity from uprates at Constellation reactors in Illinois, Pennsylvania and New Jersey, which the utility says will cost about $4.3 billion, plus a separate 15-year contract for 2,700 megawatts from existing plants. Constellation says the first uprated reactor should come online by 2028. The Register notes the deal upgrades existing units rather than building new ones.

💼 Industry & Funding

• Manus parent raises over $500 million after Meta deal collapse – Butterfly Effect announced on October 8 a round of more than $500 million led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund following on, TechCrunch and TechNode report. It is the first financing since Chinese authorities ordered the $2 billion sale to Meta unwound in April. The company did not disclose a valuation; last month it was reported to be seeking $4 billion.
• Nous Research raises $90 million at $1.5 billion valuation – The Wall Street Journal reported on October 7 that the open-source lab behind the Hermes Agent closed a Series B led by Robot Ventures, with Nvidia, Microsoft’s M12, Samsung and Union Square Ventures participating. The money goes to Hermes for Businesses, an enterprise version of the MIT-licensed agent. Nous says the agent has been cloned more than 24 million times, a company figure.
• AMD buys Fei-Fei Li’s World Labs for $8.2 billion in stock – Announced on September 28, the all-stock deal brings the spatial-intelligence startup founded in 2024 into the chipmaker, CNBC and The Next Platform report. Li joins AMD as executive vice president and chief scientist reporting to Lisa Su. Closing is expected by the end of 2026, subject to customary conditions.
• Schneider Electric agrees $22.6 billion PTC acquisition – On October 5 the French group agreed to buy the US engineering-software firm for an enterprise value of about €22 billion, a premium of more than 40 percent to PTC’s last close, Bloomberg and Energy Connects report. The all-cash deal is expected to close in the third quarter of 2027; chief executive Olivier Blum called the combination a software and AI powerhouse.

⚖️ Policy & Safety

• New York City Council questions four AI labs under oath – On October 5 executives from OpenAI, Anthropic, Google and Meta testified before a Committee of the Whole while xAI did not appear despite a subpoena, amNY and City & State report. Speaker Julie Menin sought a yes-or-no commitment not to release models that fail safety tests; the companies described their review procedures without giving one. Bills under consideration cover third-party validation, kill switches and 24-hour incident reporting.
• Anthropic widens cyber access programme into three tiers – On October 6 the company restructured its Cyber Verification Program, absorbing Project Glasswing, so that vetted security teams get Opus 5.5, Sonnet 5.5 and Mythos 5.1 with reduced blocking classifiers at Defense, Red Team and Specialized levels. Anthropic says Glasswing partners found at least 129,000 verified vulnerabilities between April and July. In its own test, Opus 5.5 was stopped in 46 of 50 runs under Defense Access but completed 34 of 50 under Red Team Access.
• Third Circuit affirms Westlaw copyright win over ROSS – On September 29 the appeals court became the first federal circuit to rule on fair use in AI training, upholding Judge Bibas’s finding that ROSS Intelligence infringed Thomson Reuters’s headnotes. It found the use minimally transformative and saw harm to a developing market for licensing headnotes as training data, per Ballard Spahr’s analysis. The court said the technology at issue could not generate original expression, leaving generative-AI questions open.
• Newsom signs No Robo Bosses Act among 13 AI bills – On September 30 California’s governor signed SB 947, barring employers from relying solely on automated systems to fire or discipline workers from July 1, 2027, reversing his 2025 veto of an earlier version, CNBC and The Next Web report. He vetoed a smart-glasses recording bill as overly broad. The California Federation of Labor Unions backed the package; TechNet, representing Meta, Google and Amazon, opposed it.

🔭 What to Watch

• TSMC reports third-quarter results on October 15 – ASML is scheduled to report the day before, per earnings calendars.
• Mistral Large 4 weights are due by the end of October – Mistral has committed to the date but not to a licence.
• Florida court weighs motion to restrict OpenAI’s model development – The attorney general’s September 28 filing has no reported hearing date.

Weekly AI review generated on October 9, 2026.