Weekly Review: Trump Bans Chinese AI and Kimi Steps Up
This week was less about what AI can do and more about who pays for it, who lets it in, and who profits from it. Websites and banks are closing their doors to agents, Anthropic reveals the cost of the frontier in its prospectus, and open models from China are closing the gap to within a single point. Two stories in particular made waves: banks are already finding third-party agents logged into customer accounts, and DeepSeek responds to the US ban with $15 billion in fresh capital.
Monday — Yelp, Delta, JPMorgan, Citi, and Bank of America: Even More Websites Lock Out Agents
Several major US websites are currently blocking autonomous AI agents designed to handle tasks on behalf of users. Yelp stated that it only permits agents that pay for its proprietary data licensing program, while Delta Air Lines is barring agents from booking tickets. JPMorgan Chase and Charles Schwab confirmed that they are excluding agents from accessing customer accounts for security reasons. In recent weeks, users have reported blocks affecting Meta’s Muse, xAI’s Grok, OpenAI’s Dots, and Instinct’s agent; Cars.com and Amtrak also suddenly stopped working for some users, indicating silent blocks by additional providers. Even without an explicit ban, agents frequently run into anti-bot systems.
Amazon set the precedent last year by blocking Perplexity’s shopping agent and subsequently filing a lawsuit. Amazon argued that it had repeatedly asked Perplexity to stop, and that third parties shopping on behalf of customers should act transparently and respect the service provider’s decisions. Perplexity countered that Amazon simply feared fewer human eyes on its ads and compared its agent to an extension of the user. A court initially barred Perplexity from shopping on Amazon, but an appeals court overturned that decision in August, prompting Amazon to file an amended lawsuit.
How Anthropic’s Duet aims to circumvent this exact hurdle on Thursday using permissions and shielded passwords reveals the flip side of these blocks.
Synthszr Update: According to Axios on October 22, major US banks have not only turned away unauthorized agents at the door, but have also detected some already logged into customer accounts. Morgan Stanley is now actively blocking them, while Wells Fargo is still working on countermeasures. Dimon refuses to allow “arbitrary third-party agents,” though read-only access via intermediaries like Plaid remains permitted. The take held up and has become even more urgent: Liability is no longer a theoretical exercise, as agents are already operating in areas where banks fall back on their terms of service. In the future, access via authorized intermediaries will come at a price. Amazon vs. Perplexity remains the legal touchstone, with the hearing scheduled for November 20.
Tuesday — Trump Bans DeepSeek from All Civilian Federal Agencies
The US Department of Homeland Security (DHS) has ordered the removal of all products from Chinese AI company DeepSeek from the IT systems of all civilian federal agencies. The order was published in the Federal Register on October 19, 2026. Affected agencies include the Environmental Protection Agency (EPA) and the communications regulator FCC. They must remove existing DeepSeek products and are barred from deploying new ones. This explicitly applies to third-party services built on DeepSeek technology as well.
In its rationale, the department cites “multiple threat vectors.” The models could grant the Chinese Communist Party “extensive authority” to access US government infrastructure. According to the DHS, users could be tracked or digitally identified through the software. Sensitive data belonging to government personnel and contractors could be leaked to the People’s Republic of China. Additionally, the models could harbor malware or other security vulnerabilities.
The order follows just days after the release of DeepSeek’s latest model, V4.1-Flash-1017. The company promotes it as “smarter, faster, more efficient,” with improvements in reasoning and coding. Initial reports place the model on par with some of the most advanced US systems. DeepSeek had previously slashed the price of its flagship V4 model permanently by 75 percent.
Synthszr Update: On October 20, China’s Ministry of Commerce threatened “all necessary measures.” On October 22, DeepSeek announced a funding round of around $15 billion led by Tencent and CATL, valued at approximately $74 billion, and is pushing ahead with an IPO in Shanghai. The core take held up, even more clearly than anticipated: The federal ban costs DeepSeek virtually nothing. In China, it serves almost as a seal of approval for investors. This flips the logic of the measure on its head. Washington cuts procurement lists, while Beijing funds a challenger to OpenAI and Anthropic. The decisive question remains whether the third-party clause trickles down to federal contractors.
Wednesday — Anthropic Loses $20 Million per Hour Ahead of IPO
Anthropic posted a net loss of $80.7 billion in the first half of 2026. This is revealed in a draft of the S-1 prospectus that the company is preparing for its planned initial public offering. Over the same period, revenue reached $16.3 billion. Preliminary figures had already surfaced in late September, but the draft now provides considerably more detail. According to concurring reports, the IPO is slated for November.
Growth continues unabated. Last week, Anthropic informed prospective investors that revenue in the third quarter alone reached $16.8 billion, representing a 45 percent increase over the previous quarter. However, revenue relies on a narrow base: two customers account for roughly one-fifth of overall sales. The company itself lists this concentration as a risk factor.
The largest cost driver is computing power. In the filings, Anthropic indicates that it could spend up to $750 billion on additional compute capacity in the coming years. This total also includes contracts currently being negotiated. Despite this scale, the company is struggling to secure sufficient cloud capacity from AWS, Google, and Microsoft. According to the prospectus, the three major cloud providers increasingly view the Claude developer as a competitor rather than a customer.
Anthropic is responding by charting its own paths to compute power. The company is leasing its own data centers and partnering with third-party data center operators. However, in the prospectus Anthropic raises concerns that these operators have only a “limited operating history.” Consequently, they could prove less reliable than the established cloud giants.
The list of risk factors is long. In addition to dependence on a few major customers and uncertain access to computing power, the draft warns of liability risks. According to the document, Anthropic could be held liable for damages caused by harmful behavior of its AI models to an extent that exceeds available insurance coverage. This ties into previously disclosed information regarding catastrophic risks. Anthropic, led by CEO Dario Amodei, did not initially respond to inquiries regarding the documents. → Reuters, Wall Street Journal, TechCrunch
Synthszr Update: According to reports from October 22 and 23, Anthropic could publicly file its S-1 on schedule so that the roadshow can kick off in mid-November. Chamath Palihapitiya cited dependence on cloud computing capacity as a key risk. The take held up: investors see the exact same problem. Anthropic purchases computing power from its competitors.
Thursday — Anthropic positions Claude Duet, its own AI assistant, against Muse, Dots & Co.
With Claude Duet, Anthropic has introduced an agent that monitors connected applications, flags priorities, and completes tasks in the background. The product is based on the Opus 5.5 model and launches as a Research Preview. Access is initially restricted to Pro and Max subscribers in the US via a waitlist. Duet runs as a dedicated app for iOS and Android, as well as on claude.ai. The same usage limits apply as for the rest of Claude. According to Anthropic, the target audience is freelancers and small businesses. Team and Enterprise plans are excluded for now. There is neither an API nor a timeline for enterprise customers.
The agent is designed to act proactively. Every morning, it delivers a brief overview of key activity across connected services. Throughout the day, it checks these accounts and reaches out when something requires attention. Users decide for themselves which topics are reported and how often. Checks occur every few hours, meaning there is no monitoring in real time. According to Anthropic, more frequent queries would deplete the usage allowance more quickly.
Anthropic lists examples such as tracking unpaid invoices, drafting customer follow-ups, scheduling meetings, creating financial and sales forecasts, building dashboards across dozens of tools, and spinning up prototypes and web apps. Duet is controlled via text or voice in an ongoing conversation. The agent adopts memory, skills, and existing app integrations directly from the Claude account. Technically, it operates inside a virtual machine with its own browser, where it determines which applications and steps a task requires. The work continues even when the laptop or smartphone is powered off.
Before major steps, such as making a payment or deleting data, Duet requests approval. According to Anthropic, the agent never directly sees passwords or login tokens. Actions are vetted by the same classifier designed to prevent unwanted actions in Claude Code. Defenses against hidden instructions and prompt injection are based on research into browser control. In-house red teams and an external firm tested Duet for vulnerabilities, the company stated. → Reuters, VentureBeat, Claude, The Verge
Duet enters a market where banks and retailers just shut their doors to agents like Muse and Dots on Monday.
Synthszr Update: On October 22, Anthropic confirmed that morning briefings and intermediate checks count against regular quotas. Newly admitted users receive a one-time refreshed quota valid until November 30. The take held up: the reset is a band-aid over the conflict between a flat-rate model and agents. Pricing, access, and enterprise plans remain unchanged.
Friday — Orchestrated frontier models reportedly outperform primary care doctors
Microsoft has published a study in the journal Science showing that its proprietary diagnostic system MAI-DxO correctly identifies complex medical cases more than twice as often as primary care physicians. Across 304 difficult cases from the New England Journal of Medicine, the system achieved an accuracy rate of 83 percent. Family physicians and internists reached 36 percent when using reference materials, and around 22 percent without aids. For more routine emergency room cases taken from the Beth Israel Deaconess Medical Center’s MIMIC-IV-ED database in Boston, the ratio was 80 to 47 percent. The paper, titled “Sequential Diagnosis with Language Models,” had already appeared the previous year as a non-peer-reviewed preprint. Microsoft AI chief Mustafa Suleyman called the result a step toward “medical superintelligence”.
MAI-DxO operates sequentially. Initially, the system receives only the chief complaint and decides independently which questions to ask and which diagnostic tests to order. Multiple AI agents assume roles similar to a medical panel: proposing diagnoses, selecting examinations, challenging assumptions, and weighing costs while refining the differential diagnosis across several rounds. The researchers argue that this approach can mitigate cognitive biases, such as anchoring on an initial suspected diagnosis. For testing, Microsoft developed its own benchmark called SDBench, as licensing exams like the US USMLE provide all information upfront, failing to reflect daily clinical practice. The company plans to release parts of the benchmark, including evaluation and cost estimation code, as open source.
The system delivered the best results using OpenAI’s GPT-5 and GPT-5.5. Models from Google, Anthropic, Meta, and DeepSeek were also tested. Study data indicates that a major share of the performance stems from the underlying model. GPT-5 alone achieved 78.9 percent on complex cases at an estimated cost of around $8,400 per case. Orchestration reduced these costs by roughly $3,000 and improved diagnostic accuracy by about four percentage points. With GPT-5.5, accuracy on complex cases climbed from 72 to nearly 87 percent, while both variants hit roughly 85 percent on emergency room cases. Weaker models benefited the most: Gemini 2.5 Flash jumped from 46.4 to nearly 69 percent. Microsoft attributes diminishing returns to models approaching the theoretical upper limit. Newer models like GPT-6 Astra or Opus 5.5 were not part of the study.
The limitations are significant. Testing was conducted exclusively on historical cases. Specialists and clinical teams who handle such cases in daily practice were not part of the comparison, and the authors themselves anticipate that the performance gap with them would be narrower. The claim that the system reaches the correct diagnosis at one-third to one-half the cost is based on simulated cost estimates, according to Microsoft. Medical professionals emphasize that such systems are intended as assistants and point to unresolved liability questions. → TechCrunch, The Verge, Gizmodo, Microsoft AI
Synthszr Take: Four percentage points in accuracy and $3,000 less per case—that is all orchestration adds on top of GPT-5. The rest is powered by OpenAI’s model. Microsoft’s real asset here is SDBench: a homegrown yardstick against which every vendor aiming for hospital deployments will have to measure itself. Superintelligence is marketing; the benchmark is the product.
Saturday — Kimi’s open K3.1 Pro just barely behind Claude Opus 5.5
With K3.1 Pro, Kimi has unveiled the first model in its new K3.1 series. According to the provider, it is its most capable open-weight model in the 3T class to date and processes a context window of one million tokens. In parallel, K3.1 Pro Highspeed is being released, which is said to respond two to three times faster. Both variants are available immediately via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The complete model weights are scheduled to follow by November 4.
Technically, K3.1 Pro is a post-trained version of Kimi K3. Into its post-training went significantly more computing power. Kimi describes joint reinforcement learning across hundreds of application domains. To achieve this, the entire stack was reportedly overhauled, from the RL framework and long-context inference to the sandbox infrastructure and rollout planning. The goal was for each additional unit of RL compute to deliver stable and predictable gains.
In benchmarks, the provider sees itself on par with the strongest proprietary models. On Terminal-Bench 4.0, which measures coding agents on realistic terminal tasks under resource constraints, K3.1 Pro reportedly scores 65.7 percent. GPT-6 Astra reaches 58.2 percent there, and Claude Opus 5.5 comes in at 66.4 percent. In the in-house Kimi Code Bench 3.0 long-horizon, the score rises from 55.95 to 61.2 percent compared to K3. On the 3D benchmark BenchCAD, which evaluates object reconstruction from multiple views using CAD code, the model scores 92.7 percent with tools. Claude Opus 5.5 stands at 96.2 percent here, and GPT-6 Astra at 95.9 percent.
In terms of focus, Kimi emphasizes coding, 3D creation, and agentic tasks. The model is designed to work with game engines such as Unity, Unreal Engine, and Godot, combining them with web search and 3D generation. For parametric models like machine parts, architectural models, or furniture, it independently writes and executes scripts for tools like Blender according to the manufacturer, even without MCP connectors.
As evidence of long task chains, Kimi cites an internal test. Based on the open human-atlas body model with 296 skeletal and 402 muscle components, K3.1 Pro built a simulator for human walking complete with a training and evaluation pipeline. Across 167 steps and 420 independent reviews, it improved the gait controller to the point where the simulated figure walked for 11.2 seconds with alternating steps instead of collapsing after seconds. When arm movement failed to improve, the model reportedly investigated the simulator’s dynamics directly. In doing so, it identified and corrected 36 errors in the calculation of shoulder muscle lever arms. → Kimi, BottleCap AI
For Anthropic, which on Wednesday disclosed almost five dollars in losses for every dollar of revenue, open price pressure hits the exact spot on which the IPO narrative hinges.
Synthszr Take: Just one point on Terminal-Bench separates Claude Opus 5.5 from a model whose weights any hoster will be able to load starting November 4. With this, Anthropic loses the basis for a premium on coding agents, while GPT-6 Astra even lags significantly behind. Moving forward, prices will be dictated by the hosting market—and that market only cares about compute costs.
