The AI Upload

Major AI labs are racing to cut costs, and early evidence suggests the strategy is working. OpenAI this week released GPT-6 Sol and GPT-6 Luna at prices 50% lower than previous iterations, while Anthropic's new Opus 5.5 costs around 40% less to run than its predecessor. xAI released Grok 4.7 with a similar price-performance pitch, only to see competitors neutralize its advantage within 24 hours.

The price reductions appear to be driving meaningful demand increases. A Citadel Securities report found that falling per-token costs are fueling additional usage and higher overall AI spending, consistent with the economic principle known as Jevons Paradox. Chinese open-weight providers are adding competitive pressure: DeepSeek's newly released V4.1 Flash, which the company says outperforms its prior flagship model on several benchmarks, surged to the top of OpenRouter's leaderboard with a 172% spike in usage. Morgan Stanley views Chinese competition as an additional bullish signal, predicting it will expand overall demand for computing infrastructure. For legal professionals and institutions evaluating AI adoption, declining costs may lower the practical barriers that previously made enterprise AI tools difficult to justify.

AMD briefly crossed a $1 trillion market capitalization on Monday, placing it among a small group of chip designers to reach that milestone. The company still trails Nvidia, which holds a market cap of nearly $5.5 trillion, but has surpassed long-time rival Intel, currently valued at $640 billion.

Much of AMD's recent valuation growth is tied to its Instinct GPU line, which has gained traction with major AI labs and cloud providers including Microsoft, OpenAI, Oracle, Anthropic, and Meta. After years of criticism over its ROCm software ecosystem, AMD has worked to close that gap while competing on price and memory capacity. Its latest Helios rack platform, launched in July, is claimed to offer 50 percent more HBM4 memory and between 15 and 25 percent higher AI training performance than Nvidia's current Blackwell-based systems.

GPU sales are not the only driver. AMD's Epyc server processors now hold 34.5 percent of the datacenter CPU market, rising to 46.4 percent when measured against Intel Xeon SP alone. CEO Lisa Su told investors the company expects datacenter revenue to more than double year over year in 2027. Some analysts remain cautious, citing concentration risk in AI, though AMD's broader portfolio spanning embedded systems, FPGAs, and consumer CPUs provides exposure well beyond that single market.

A 92-page plaintiffs' brief, unsealed Thursday in the consolidated OpenAI copyright litigation before Judge Sidney H. Stein, quotes Microsoft's Director of Applied Science Dr. Brent Hecht describing AI training on published content as potentially "the largest theft of labor in human history." The brief also cites internal Microsoft documents warning that its AI content strategy has triggered a "doom loop" damaging both its own models and the broader web, with company data showing click-through rates to plaintiff publishers falling as much as 94% when users interact with Copilot instead of traditional Bing search results.

The plaintiffs include The New York Times, the Daily News group, the Center for Investigative Reporting, and Ziff Davis. The filing quotes OpenAI's VP of Research stating the company trains its networks specifically to memorize training data, and internal documents show OpenAI held over 91,692 copies of plaintiffs' works in its mid-training datasets. An unnamed OpenAI engineer wrote plainly: "No matter how prominently we show the links, users won't click."

Microsoft has characterized Hecht's remarks as one employee's personal perspective rather than legal analysis. The defendants have not conceded the core copyright question, and courts have generally favored AI companies in similar disputes. Judge Stein must now decide whether AI training and output constitute fair use or actionable substitution.

British Columbia has filed a lawsuit against OpenAI in California over the February 10, 2026, shooting at Tumbler Ridge Secondary School, which killed eight people, six of them children. The province alleges that OpenAI's internal safety team flagged the shooter's ChatGPT conversations for gun violence-related content as early as June 2025, nearly eight months before the attack, and recommended contacting law enforcement. According to the suit, that recommendation was not followed, reportedly due to concerns about OpenAI's commercial prospects and its anticipated IPO.

The lawsuit was formally filed on September 21, 2026, by Attorney General Niki Sharma, who stated the conversations were never reported to the Royal Canadian Mounted Police. British Columbia's action joins a growing body of litigation: 37 separate civil suits have been filed by victims' families in California federal court, some seeking over $1 billion in damages and demanding structural changes to OpenAI's safety protocols. OpenAI has moved to dismiss many of those cases on jurisdictional grounds, issued a public apology, and pledged to strengthen its safety practices.

Newly unsealed internal documents in The New York Times' lawsuit against OpenAI and Microsoft have handed publishers some of their most potent legal ammunition to date. The filings reveal that OpenAI's head of ChatGPT, Nick Turley, described the company's products as "largely substitutive, period" to publishers, while Microsoft recorded 83-93% drops in click-through rates to Times and Daily News domains when comparing traditional Bing search with its Copilot chatbot. Microsoft CEO Satya Nadella testified that chatbot conversations substitute for visiting the underlying source.

Legal experts say the documents seriously damage the defendants' fair use defense, particularly on the fourth factor: market harm. Because OpenAI and Microsoft have themselves participated in licensing deals valuing this type of content, experts argue a commercial market clearly exists. Evidence of deliberate paywall circumvention adds potential Digital Millennium Copyright Act liability on top of copyright claims. Some lawyers caution that substitution was already the weakest part of defendants' fair use argument, and recent rulings in Anthropic and Meta cases show courts can still find AI training transformative.

Both sides are seeking summary judgment before Judge Sidney Stein. A ruling favoring the Times could accelerate licensing negotiations industry-wide, with executives and legal experts predicting a wave of additional publisher lawsuits and pressure on AI companies to enter structured, paid licensing arrangements.

OpenAI has expanded its GPT-6 model lineup with updated versions of Sol and Luna, two smaller models positioned below the recently released GPT-6 Astra. Sol is designed for complex tasks such as coding, while Luna targets high-volume clerical work like document summarization and information extraction. Both models are now available in ChatGPT Work, Codex, and the API for most paid accounts, with Luna also accessible to Free and Go users.

The updates come with notable cost reductions: API access to the GPT-6 Sol and Luna models is priced at half the cost of their 5.6 predecessors, a drop OpenAI attributes to improvements in caching and inference. OpenAI also reports that GPT-6 Sol makes roughly half as many factual errors as its predecessor, claiming it reaches the reliability level of Astra at a lower price point.

The release landed just 90 minutes after Anthropic published a new version of its Opus 5.5 model, underscoring the continued competitive pressure between the two AI developers. Legal professionals and researchers relying on AI for document review, summarization, or coding assistance may find the improved accuracy and lower costs relevant to their workflows.

SpaceX has released Grok 4.7, its latest large language model, built on a new base model and an enhanced reinforcement learning workflow that gave the model more difficult training tasks over longer periods. The model supports multi-agent parallel processing through SpaceX's Grok Bot harness, which allows complex tasks to be distributed across several AI agents that can also verify each other's outputs.

Benchmark results are of particular interest to legal professionals: Grok 4.7 outperformed competitor model Fable 5.1 on the Harvey Legal Agent Benchmark, which tests performance on legal tasks, though it trailed OpenAI Group's GPT-6 Astra on a chip design benchmark. SpaceX also reports record scores on safety benchmarks targeting malicious biology research and cybersecurity requests. The model is priced starting at $2 per million input tokens and $6 per million output tokens, with a faster, higher-cost option available for latency-sensitive applications.

AI-powered bug hunting is producing a surge in discovered software vulnerabilities that security teams are struggling to keep pace with. Microsoft has already patched 974 CVEs this month alone, setting a new record. Oracle shipped 1,448 patches in July compared to 309 during the same month the prior year. Google Chrome's two major June releases included 1,072 patches, exceeding the combined total from the prior 23 major releases. As of mid-September, 66,401 CVEs have been recorded for 2026 — nearly double the 33,512 logged by the same point in 2025, and more than double the roughly 25,000 recorded in all of 2022.

Researchers disagree on whether this represents a crisis or a sign that vulnerability detection systems are functioning better. Jerry Gamblin of Empirical Security cautions that more CVEs means more known vulnerabilities, not necessarily more actual risk. The deeper concern, however, is that patching capacity — which depends on human labor — cannot scale as quickly as AI-driven discovery. Britain's National Cyber Security Centre has noted that finding vulnerabilities alone does not improve security. For legal and compliance professionals tracking software risk and cyber liability, the widening gap between discovery and remediation may be the more pressing development to watch.

Microsoft, partnering with Health-ISAC, Cloudflare, Coinbase, OpenAI, and several other organizations, took down EvilTokens, a phishing-as-a-service platform that exploited the OAuth 2.0 device authorization flow to hijack authenticated email sessions without capturing user passwords. Acting under authorization from the U.S. District Court for the Eastern District of America's Eastern District of Virginia, the coalition seized 50 websites and disabled more than 150 domains tied to the operation. The Metropolitan Police Service separately arrested two men on September 11, 2026.

EvilTokens, sold on Telegram starting in mid-February 2026, integrated AI tools to analyze compromised inboxes in over twenty languages, map organizational relationships, identify financial decision-makers, and draft impersonation emails targeting business email compromise and invoice fraud. The platform operated on a subscription model starting at $500 per month, generating an estimated $1.1 million in platform revenue traced through cryptocurrency between October 2025 and June 2026. Microsoft linked the service to more than 12,000 compromised inboxes across over 10,000 organizations worldwide, with victims concentrated in financial services, healthcare, higher education, and real estate sectors.

As the Trump-Xi summit gets underway in Washington, one of its most consequential potential outcomes may be neither trade nor diplomacy but an emergency AI hotline modeled on the Cold War-era "red telephone" established after the Cuban Missile Crisis. Treasury Secretary Scott Bessent has already pitched Chinese officials on a "notification mechanism" for AI incidents with national-security implications, and the two leaders are expected to discuss shared AI risks when they meet Thursday.

The proposal reflects a narrow but significant overlap in U.S. and Chinese concerns. While both governments are committed to winning the AI race, each also worries about near-term dangers such as cyberattacks and disruptions to critical infrastructure. The arrangement would not slow either side down, making it more politically palatable than a broader safety regime.

Serious implementation questions remain. No one has publicly defined what would trigger the hotline, and in a frontier-model incident, private companies may have more situational awareness than governments. Past crisis channels between Washington and Beijing have also gone unused at critical moments, including during the 2022 surveillance balloon incident, raising questions about whether a new mechanism would fare any better.

A new report from the Basel Action Network (BAN) warns that AI infrastructure could generate between 395 million and 617 million metric tons of electronic waste between 2025 and 2050, a figure the organization says is 40 to 60 times higher than widely cited academic projections. The disparity stems from prior estimates focusing almost exclusively on servers and GPUs, which represent only 13 percent of a datacenter's equipment by weight. BAN's model incorporates power distribution, networking, cooling, and backup power systems, arriving at roughly 7,000 metric tons of equipment for a reference 100 MW facility.

The report attributes much of the projected waste to short replacement cycles: AI accelerators turn over every two to three years, networking gear every three to four years, and entire power infrastructure must often be replaced when facilities upgrade to higher-density racks. BAN notes that no hyperscaler, government, or international body has published a plan to handle e-waste at this projected scale, and that existing processing infrastructure cannot safely manage even current volumes. A planned follow-up report will examine how reuse and refurbishment might reduce the total, with later installments addressing toxicity and mitigation strategies.

OpenAI has formed a nine-member independent advisory panel of elite mathematicians, called the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), to help the company manage how it communicates and releases AI-generated mathematical results. The group, hosted by Princeton's Institute for Advanced Study, includes researchers from Stanford, Harvard, Oxford, and Cambridge, and will have early access to OpenAI's research. Members are unpaid and have not signed restrictive agreements limiting their public commentary.

The panel's formation follows significant backlash over OpenAI's handling of recent high-profile mathematical announcements, including accusations of failing to credit human researchers and inadequate coordination with the academic community. Its immediate task is advising on how to release what OpenAI says are more than 100 significant open mathematical problems resolved by an unreleased internal model.

Mathematicians outside the group have raised pointed questions about its representativeness, independence in practice, and whether its elite composition reflects the concerns of the broader research community. Group member Martin Hairer acknowledged in a blog post that members are not naive about the risk of being used for PR purposes, but argued that refusing to engage would be counterproductive given OpenAI's apparent willingness to consult the field.