Weekly AI Review — September 27, 2026

🚀 Model & Product Releases

• Anthropic releases Claude Opus 5.5 at lower list prices – Launched on September 22, Opus 5.5 is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure and across Pro, Max, Team and Enterprise plans. List prices fall to $4 per million input tokens and $20 per million output tokens, from $5 and $25 for Opus 5, and Anthropic says the model performs at the level of Claude Fable 5.1 on most work while costing about 40 percent less to run on typical workloads. Extended thinking can no longer be switched off, one of the listed breaking changes.
• OpenAI adds GPT-6 Sol and Luna below Astra – The two API models, released on September 22, are priced at $2 and $10 per million input and output tokens for Sol and $0.10 and $0.50 for Luna, which OpenAI told VentureBeat is permanent pricing, half the GPT-5.6 Sol rate. OpenAI reports Sol scoring 33.2 percent on AutomationBench 1.0.6 at $0.27 per task. CNBC noted that the OpenAI and Anthropic launches were the first from either lab since Dario Amodei’s call to “pace the frontier”.
• xAI ships Grok 4.7 at unchanged prices – Released on September 21, the model is live without a waitlist on the xAI API, in the Grok app, Cursor and Grok Build, with a 500,000-token context window at $2 per million input and $6 per million output tokens below 200,000 prompt tokens. xAI calls it “a notable improvement over Grok 4.6 at the same price and speed”. Independent leaderboards cited by Decrypt still place it behind Claude Fable 5.1 on GDPval and multi-hour office tasks.
• Meta expands its Muse agent at Connect – At Connect on September 23, Mark Zuckerberg announced a real-time video avatar for the Muse agent, Mac desktop control, a dedicated email address and integrations with retailers including Shopify, Walmart and Best Buy. Glasses support is promised in the coming months and several features only “soon”, so most are announcements rather than shipped capabilities. Muse remains free, and Zuckerberg said Meta will eventually take a small fee on transactions.

🔬 Research Highlights

• Anthropic reports Claude agents found a CRISPR-like enzyme system – In a preprint released on September 23, Anthropic says about 950 Claude agents searched sequence databases for 21 hours, narrowing over 200,000 reverse transcriptases to 20 candidates and identifying array-associated reverse transcriptases (ART) in bacteriophages. Initial lab work detected distinct short RNAs from the repeat arrays. The authors state that the system’s biological function is still unknown, and the work has not been peer reviewed.
• Google researchers regularise self-improving agent harnesses – arXiv 2609.24972, posted on September 21 by a 14-author team, introduces RRSI, which constrains automated edits to an agent’s prompts, tools and memory with an annealed edit budget, a critic and a pruner. The authors report gains of up to 14.1 points on the split the harness evolves against and 30 percent fewer policy tokens. Their own out-of-distribution gains top out at 4.7 points, reflecting the overfitting the method is designed to limit.
• Researchers extract hidden reasoning from closed frontier models – arXiv 2609.26637, posted on September 22 by Xiaoyu Luo, Johannes Bjerva and colleagues, registers a custom tool through a standard API feature to make models such as GPT-6 Astra write out intermediate reasoning. They report that the extracted reasoning matches native chain-of-thought performance on open models and outperforms no-reasoning baselines. The authors caution that such traces may be post-hoc rationalisation, which they can check against native reasoning only on open-source models.

🏗️ Infrastructure & Compute

• Anthropic commits $11.6bn to Akamai for CPU cloud capacity – Akamai announced the seven-year agreement after the US close on September 24, with an option to expand by a further $9 billion, and issued Anthropic a warrant for up to 7.7 million shares, about 5 percent, at $111.33. TechCrunch reported that the contract centres on CPU rather than GPU workloads and that Akamai expects about $5.5 billion of related capital spending. Both sides retain exit provisions tied to delivery and availability requirements.
• Oracle sends force majeure notice on New Mexico Stargate campus – Bloomberg reported on September 24 that Oracle notified a Blue Owl Capital unit developing Project Jupiter, a 2.45-gigawatt campus for OpenAI, preserving the right to delay payments if it misses its 2028 start. TechCrunch cited a gas pipeline delayed to February 2027 and an air-quality permit for Bloom Energy fuel cells still pending. Oracle said the project “remains on our planned schedule”.
• Alibaba unveils the Zhenwu V900 accelerator and a 20GW target – Chief executive Eddie Wu presented the chip at the Apsara conference in Hangzhou on September 22, with 216 GB of memory, 1,200 GB/s of chip-to-chip bandwidth and three times the performance of the Zhenwu M890, by Alibaba’s own measure. Mass production is scheduled for the first quarter of 2027. Alibaba also set a goal of more than 20 gigawatts of data-centre capacity by 2032 and said Qwen 4.5 and Qwen 5 will reach 5 to 10 trillion parameters.

💼 Industry & Funding

• Island raises $400m at a $6.4bn valuation – The Series F, announced on September 24, was led by Evolution Equity Partners with Sequoia Capital, Coatue, Insight Partners and J.P. Morgan Growth Equity Partners among participants. Island now markets itself as a control plane for governing employees and AI agents across browsers, devices and applications. It says annual recurring revenue has doubled every fiscal year since 2022, without disclosing a figure.
• Finland’s Verda passes a $1bn valuation with $189m round – The Helsinki AI cloud provider, formerly DataCrunch, closed the Series B on September 22, led by Emergence Capital with MUFG Innovation Partners, Supermicro and the Finnish state investor Tesi. Verda says its annualised revenue run rate reached $165 million in July, and it plans to multiply computing capacity within a year from facilities in Finland and Iceland.
• Brahma AI raises $150m led by Multiples – The enterprise audiovisual AI company said Multiples Alternate Asset Management invested $100 million of the round, with Quartz and Dealroom reporting a $2 billion post-money valuation. Brahma lists Warner Bros, the NBA and Mayo Clinic as anchor customers and says the money will fund expansion in the United States.

⚖️ Policy & Safety

• AI chiefs brief the UN Security Council as Washington objects – At a September 23 meeting convened by France, Anthropic’s Dario Amodei proposed narrow agreements such as a ban on AI-assisted biological weapons, and OpenAI’s Sam Altman asked states to align capability measurement and failure reporting. Yoshua Bengio, co-chair of the UN scientific panel on AI, called the current effort “a race where everyone loses”. White House science adviser Michael Kratsios said dialogue “cannot be allowed to drift towards global governance”; no resolution was adopted.
• United States and China agree an AI incident channel – At their Washington summit, Donald Trump and Xi Jinping agreed “a communication mechanism for artificial intelligence-related incidents”, with an AI-specific dialogue scheduled for November, CBS News reported. Xi said the two sides could “together guard against the misuse or malicious use of AI”, while Trump said the United States was “not going to be putting on brakes”. China’s readout gave few details.
• Twenty-six attorneys general press Congress for federal AI rules – A bipartisan coalition led by New York’s Letitia James wrote to congressional leaders on September 24 seeking federal oversight of safety testing, government-led incident response and international cooperation, alongside a ban on pre-empting state laws. The letter cites cases of AI agents acting outside test environments. President Trump posted that the only guardrail AI needs is “a STRONG AND SMART” president.

🔭 What to Watch

• OpenAI holds DevDay on September 29 – Fortune reports that OpenAI plans to preview GPT-6 Cyber and a security product around the event.
• Micron reports fiscal fourth-quarter results on September 30 – The memory maker guided in June to revenue of $50 billion, plus or minus $1 billion.
• First US-China AI dialogue expected in November – The summit set the dialogue for November without naming a date or venue.

Weekly AI review generated on September 27, 2026.