ChatGPT & Co can now also directly manage your portfolio

Scalable: ChatGPT & Co can now also directly optimize your portfolio

Scalable Capital is opening its platform to AI assistants, calling the feature 'Agentic Investing.' Since August 25, 2026, customers have been able to link their portfolios with ChatGPT, Claude, or Grok; according to the Munich-based provider, this makes it the first bank in Europe to offer this for all common assistants. From the outset, the range of functions goes beyond read permissions: trading, setting up savings plans, managing watchlists, and setting price alerts are possible at launch. There is also a dedicated search for stocks, ETFs, and derivatives. The company provides news, real-time quotes, and historical price data free of charge, as well as parts of the Scalable Insights analysis tool with diversification checks, scenario analyses, and risk assessments.

Technically, two methods are available. One is a server based on the Model Context Protocol, accessible at mcp.scalable.capital, and the other is a CLI application for the command line that can be installed locally, making it suitable for self-hosted models. Access is enabled on the web under Profile > Security > Agentic Investing, after which the activation must also be confirmed in the app. For Claude, the connection is added as a custom connector, followed by a login and two-factor authentication prompt. In the connector's settings, it’s possible to specify for each command whether it is always allowed, requires approval, or is blocked.

An order is not executed immediately upon the first request. The assistant first summarizes the order, Scalable displays the required preliminary information including costs and, if applicable, the key information document, and then the customer confirms the transaction in the chat. According to the provider, logging in via two-factor authentication will also be required at recurring intervals later on. Scalable points out that AI responses and recommendations do not come from the broker and do not constitute investment advice. Consumer advocates advise disconnecting the link when not actively working on the portfolio and to check beforehand what data the AI provider stores and whether it uses inputs for training. A practical limitation: The interface does not provide the entry price; the AI must calculate it from the purchase transactions itself, which can be inaccurate for positions transferred from another portfolio.

Founder and Co-CEO Erik Podzuweit describes Agentic Investing as the 'biggest technological leap in financial technology since internet banking.' The company cites use cases such as a daily 7 a.m. newsletter about one’s own portfolio holdings, the CSV-export of transactions for tax returns, or the calculation of a savings rate with a pre-filled savings plan order.

At the same time, what the assistants on the other side do with such data is also changing. On the same day, Anthropic merged the memory systems of Claude Chat and Claude Cowork: what is created in the chat interface will be available in Cowork in the future, and vice versa. The feature is active by default, and users must actively opt out; the memory is continuously updated during conversation, and according to the company, there are separate controls for sensitive data. Claude Code is excluded from the merger. → aktiengram, zdnet, dasinvestment, scalable, finanztip

Synthszr Take: Scalable is selling Agentic Investing as the biggest leap since internet banking, yet hides the feature under Profile > Security, where the customer has to enable it themselves and countersign via the app. There is a gap between the press release and the product architecture, and the architecture is more telling. Every order must be confirmed individually in the chat, every write command can be hard-blocked in the permissions menu, and the 2FA prompt reappears at intervals. This caution is well-founded: The interface doesn’t even provide the entry price; the AI calculates it from the purchase transactions, and this becomes inaccurate with portfolio transfers. As long as a calculation error can directly result in an executed order, the confirmation click remains the cheapest insurance for investors.

Perplexity now runs its agent locally

On Tuesday, Perplexity introduced Portable Computer, a locally running version of its cloud agent Perplexity Computer, which was launched in February. The software runs on Nvidia’s DGX Spark, a desktop system with a GB10-Grace-Blackwell processor, a 20-core CPU, and 128 gigabytes of unified memory, which has been shipping since October 2025. Unlike previous local model offerings, the orchestrator, planner, tool router, scheduler, persistent task queue, and search index run entirely on the device. The models used are the open-source Qwen 3.8 27B and PPLX 27B, a version retrained by Perplexity and optimized for the DGX Spark. Linux is currently a prerequisite; Windows machines with RTX cards and a 30B model are expected to follow.

Perplexity states the context window for Qwen 3.8 27B is 256,000 tokens but admits that the model encounters practical issues beyond about 100,000 tokens. To counter this, the company uses a compaction feature that summarizes long inputs beforehand, as well as Multi-Token-Prediction for faster processing. According to the provider, the models run in a sandbox that prevents access to unnecessary parts of the operating system, and unauthorized network connections are blocked. Work done locally does not consume token credits; billing only occurs when a task is explicitly moved to the cloud, for example, for current web information, browser usage, or one of over 15 frontier models. Portable Computer works with Google Drive, Gmail, Slack, and GitHub via connectors. → Computerworld, SiliconANGLE, RuntimeWire

Synthszr Take: Nvidia is negotiating an investment that values Perplexity at over $30 billion, and on the same day, a Perplexity product is released that runs exclusively on Nvidia silicon. This simultaneity describes a relationship beyond that of supplier and customer: The chip manufacturer has a seat at the table when it comes to deciding which models compute where. When licenses for Perplexity’s stack and the recruitment of key personnel are also being discussed, the product roadmap is no longer solely up to the startup. Nvidia’s calculation is obvious: A desktop with 128 gigabytes of unified memory only sells with software that utilizes it, and an agent that orchestrates locally and only goes to the cloud when needed is the best justification for buying one. It will be interesting to see the day Perplexity has good reasons to port the same stack to Apple Silicon or AMD. Experience shows that co-owners have a clear opinion on such plans.

According to SemiAnalysis, OpenAI’s own chip Jalapeño beats Nvidia’s Blackwell in efficiency

At the Hot Chips conference, OpenAI unveiled its self-designed inference chip 'Jalapeño', an ASIC developed jointly with Broadcom, built exclusively for the inference of large language models. According to SemiAnalysis, design work began in mid-2024, with about 16 months passing from the first hiring to tape-out. The chip uses HBM4 memory, making it comparable to top models from Nvidia and AMD in terms of memory equipment. SemiAnalysis was able to see the chip in the lab with its own InferenceX suite and reports that Jalapeño beats Blackwell in throughput per megawatt in almost all scenarios, without Multi Token Prediction or speculative decoding. At concurrency 1, the chip achieves over 700 tokens per second per user on DeepSeek R1, and around 1,400 on GPT-OSS, according to these figures. → SemiAnalysis

Synthszr Take: 16 months from the first hire to tape-out is the real news in this text. In a year and a half, OpenAI has acquired an expertise that took Nvidia two decades and an entire supplier ecosystem to build, and Nvidia’s largest customer now sits on both sides of the table. Still, one should handle the efficiency curves with care: The figures come from OpenAI itself, the real competitor is named Rubin, and Rubin is in the racks today, while Jalapeño is still a lab sample.

Deutsche Bank was a design partner for Google’s financial AI and uses it in corporate banking

Deutsche Bank is using Google’s newly introduced tool, Gemini Enterprise for Financial Services, in its corporate client business after participating in an Early Access test. Germany’s largest bank announced that it was a design partner for the product, which launched on Tuesday. It is an agentic research tool for financial professionals that combines market information, company data, and key metrics to generate analyses more quickly. According to Google Cloud, the outputs are consistent, traceable, and auditable. The bank plans to start in the corporate client division and later expand its use to other business areas. → Capital Brief

Synthszr Take: A firm that usually pushes every software decision through three committees has embedded itself in a provider’s product development before the product was even on the market. Traceability and auditability are now being advertised as core features, and that comes from precisely these kinds of partnerships: regulatory requirements can’t be bolted onto a finished model after the fact. Citi Wealth, Lloyds, and Macquarie are already on the same platform, and Deutsche Bank is starting in its corporate business with an expansion in mind.

Chris Malone leaves OpenAI: Head of data centers departs amidst Stargate expansion

Chris Malone, responsible for building data centers at OpenAI, left the company last week. The Wall Street Journal reports, citing people familiar with the matter. Malone had only joined OpenAI in March 2025, shortly after the company announced Stargate, the joint data center project with Oracle and SoftBank. According to the WSJ, his departure is part of a series of executive-level exits preceding the planned IPO. → Wall Street Journal

Synthszr Take: Eleven months in the job for a project with construction cycles measured in years: Malone didn’t even see the Stargate expansion through to its first full stage. Building data centers requires contracts for power, land, and chips with terms that extend far beyond the current model generation, and you’re up against an organization that reshuffles its priorities every six weeks. This friction rarely ends in a compromise; it ends in a departure.

Dr. Dre backs AI music, comparing it to the synthesizer

In a joint interview with Jimmy Iovine, Dr. Dre has spoken out in favor of using artificial intelligence in music production. According to TechRadar, he literally says, “I’m embracing it,” and places the technology alongside synthesizers and drum machines—tools that were themselves considered inauthentic in their time. The same conversation touches on creativity and the lessons learned from selling Beats to Apple, which acquired the headphone brand in 2014 for about three billion dollars. The statement comes at a time when streaming services are removing mass quantities of automatically generated tracks from their catalogs, usually under the label AI Slop. → TechRadar

Synthszr Take: Dre is arguing from experience, as his entire sound was built on equipment the industry considered fraudulent at the time, from the drum machine to the sampler. The Roland TR-808 went out of production in 1983 as a slow seller because it supposedly sounded too artificial, and subsequently became the foundation of two decades of hip-hop. The pattern reliably repeats itself: the people who pick up a despised tool early and bring taste to it define how the next decade sounds.

Meta launches AI agent Hatch, charges up to $200 per month

Meta is reportedly launching a consumer agent named Hatch in the coming weeks, designed to perform tasks on behalf of the user. According to The Neuron, the premium tier could cost up to $200 per month, putting it on par with the most expensive plans from OpenAI and Anthropic. Additionally, the company’s next flagship model, internally codenamed Watermelon, is announced for October. Facebook and Instagram have so far been ad-funded and free for users; a subscription price this high would be new for Meta’s consumer business. → The Neuron

Synthszr Take: The $200 per month for Hatch is copied from OpenAI and Anthropic, cent for cent. With this price tag, Meta is declaring which league the agent is supposed to play in before anyone has even used it. The competition itself is currently showing how viable this signal is: two months after launch, Anthropic’s Fable 5 captured 11 percent of corporate spending on AI because customers opted for the cheaper GPT-5.6 (data based on 70,000 companies).

Anthropic rebuilds Claude’s renderer, ends stuttering on long responses

Anthropic has rewritten Claude’s streaming renderer, which brings the model’s responses to the screen word by word. Previously, with each new incoming token, the entire response was re-rendered from scratch, which caused long responses to increasingly stutter on less powerful machines. The new renderer only updates the parts of the text that are actively changing. According to Anthropic, this makes long responses stutter nine times less often on a slower laptop, the longest freezes are 4.5 times shorter, and on a MacBook with a 120 Hz display, the output consistently maintains 120 fps. → AlphaSignal

Synthszr Take: A renderer rewrite doesn’t show up in any benchmark table, but it determines whether someone keeps a tool open eight hours a day. Friction has been removed here: 4.5x shorter freezes, stable 120 fps, zero configuration effort for the user. In companies where the hardware is three years old on average and IT decides on rollouts, this detail carries more weight than the next point on a reasoning benchmark.

Playwright gives coding agents their own CLI instead of screenshot automation

Playwright, the browser testing framework from the Microsoft ecosystem, now offers its automation in three variants: the classic Test Runner, a command-line version for coding agents, and an MCP server. According to the project, the CLI (installed via npm i -g @playwright/cli) is aimed at agents like Claude Code and GitHub Copilot and is designed to manage with little context overhead through installable skills. The MCP server gives agents full browser control via the Model Context Protocol and can be integrated into VS Code, Cursor, Claude Desktop, and Windsurf. Playwright passes structured accessibility snapshots to the agents—i.e., element roles, names, and references—instead of screenshots; according to the documentation, vision models are not required for this. → The Pragmatic Engineer

Synthszr Take: Token efficiency isn’t listed in any feature comparison, yet it determines whether an agent remains affordable in continuous operation. A screenshot has to go through a vision model, whereas an accessibility tree is text and is read directly; that’s the difference between a test run a team performs daily and one they can only afford twice a month. Of the three entry points, the CLI with its installable skills is the most pragmatic because it only loads context when needed, instead of pre-filling the window with tool descriptions.

OpenClaw Hype: New Macs with 2-Nanometer Chips for Local Models

Apple has introduced the M6, the company’s first processor built on TSMC’s 2-nanometer node, and is using it in the new Mac mini. The chip has a 12-core processor and a 12-core graphics unit—two more cores each than the M4 and M5—plus a dual 16-core Neural Engine and 170 GB/s of memory bandwidth (M5: 153 GB/s, M4: 120 GB/s). For the first time, Apple is combining three CPU core types in one chip: two super-cores for single-thread loads, four performance cores, and six efficiency cores. According to TSMC, the new manufacturing process embeds tiny SHPMIM capacitors into the chip, which are designed to compensate for voltage fluctuations under highly variable loads. The starting price is $899 with 16 GB of unified memory and 256 GB of flash storage, configurable up to 32 GB.

All performance figures come directly from Apple: up to 40 percent more multithread performance than the M4, up to 4.8 times faster processing of prompts for locally running large language models in LM Studio compared to the M4, and 13.5 times faster compared to the M1. For spreadsheet calculations in Excel, Apple cites a factor of 1.5 compared to the M4, and a factor of 2 for ray-tracing gaming performance.

The second Mac mini variant uses the M5 Pro, which previously debuted in the MacBook Pro and continues to be manufactured on the 3-nanometer process. It offers up to 18 CPU cores, up to 20 GPU cores, a maximum of 64 GB of memory, and 307 GB/s of bandwidth, with prices starting at $1,699. In parallel, Apple is updating the Mac Studio: the entry-level version with the M5 Max starts at $2,499, and the top version with the M5 Ultra starts at $5,499. The M5 Ultra, using UltraFusion technology, connects four dies in an M-series chip for the first time, achieving up to 36 CPU cores, up to 80 GPU cores, and 1.2 TB/s of memory bandwidth—50 percent more than the M3 Ultra. For wireless connectivity, the Mac Studio uses Apple’s own N1 chip with Wi-Fi 7 and Bluetooth 6.

All four models are available for pre-order now and will ship starting September 22, equipped with MacOS 27 “Golden Gate”. Apple declined to comment on the operating system’s release date in a background briefing.

This is the first Mac mini update in nearly two years. The device was previously reported as a surprise interim release before the September event, based on information from people with knowledge of the plans. In February, Mac minis were sold out in many places as buyers used them as workhorses for agentic AI systems like OpenClaw and for locally run models with open weights; Tim Cook said in April that it would take several months to balance supply and demand. Mac revenue rose by 29 percent to $10.4 billion in the third quarter of 2026. Gartner analyst Ranjit Atwal told CIO Dive that the new chips are designed to run agentic workloads locally within enterprises. Cook will hand over leadership to hardware chief John Ternus on September 1. → SiliconANGLE, Daring Fireball, 9to5Mac, CIO Dive, Gizmodo

Synthszr Take: Apple is putting the first 2-nanometer chip in the $899 Mac mini, while the $5,499 machine makes do with the 3-nanometer M5 Ultra. Two extra CPU cores and 170 instead of 153 GB/s is a solid update for buyers, but nothing more. For contract manufacturing, its placement in the entry-level device is the real statement: N2 is in mass production, with yields that allow for volume at the lowest Mac price point, and Apple has secured the initial capacity for it. The fact that this specific device processes prompts 4.8 times faster than the M4, according to Apple’s own figures, makes it the default machine for local agents—a market that caught Cook by surprise with its supply demands in February. Whether Intel with 18A or Samsung can attract a customer of this magnitude will be decided in the next twelve months. Until then, Apple has bought its place at the front of the line.